The actual impact was limited. The war was mostly a one-sided affair, outside of the closure of the Strait of Hormuz. Still, markets did experience a shock, followed by an aggressive rebound.

Depending on how the next few weeks play out, the following companies are on the list to be evaluated for the Q2 version of this portfolio: Micron (MU), Broadcom (AVGO), NVIDIA (NVDA), Alphabet (GOOGL), Constellation Energy (CEG), ASML, Oracle (ORCL), Microsoft (MSFT), AMD, TSMC (TSM), Core Scientific (CORZ), Solaris Energy Infrastructure (SEI), Astera Labs (ALAB), Vertiv (VRT).

One month ago I temporarily exited the markets, as there was no edge to be had in this environment unless you had direct access to the administration ahead of significant announcements. Insider trading has never been more blatant or difficult to navigate unless you like gambling. This is not to say I’m complaining (humans will always be incentivized to take advantage of information asymmetry), but rather to flag that any investments around big events require you to calibrate for the cost of missing vital information in advance.

Over April I rebuilt the positions that now form the Q2’26 Infra Play portfolio (which hopefully will not require a complete directional shift, but you never know what happens with Taiwan). There are some expected names, as well as many new ones.

Power, cooling, custom silicon, memory, lithography, silicon photonics foundry, InP tools, connectivity silicon. Each layer is a separate bottleneck, each bottleneck is structurally short on supply, and each name is the cleanest expression of the layer it sits in.

Let’s get into into this not financial advice.

Read more

In my previous deep dive, I noted that the flood of money into private markets has been both a positive and a negative:

Many of the AI-native companies mostly exist to steal money from their investors in order to fund a LARP founder lifestyle. It's harsh but it's true. Still, for those that are genuinely great new entrants in their categories, there are many interesting attributes to those orgs that are not seen in their incumbent competitors.

While there are obvious cases of companies pushed hard with media exposure despite offering unoriginal, pointless, and ultimately insecure products (Cluely, Delve), there are also plenty with a genuine shot at being generational.

To be fair, the odds of a generational AI-native company emerging are obviously high when almost every new company receiving venture capital today is AI-native.

More interestingly, for the existing crop of massive private companies (Anthropic, OpenAI, Databricks, SpaceX), the funding flowing into late-stage rounds has dwarfed anything we're seeing in the public markets. An IPO today in cloud infrastructure software is seen as a burden rather than the most exciting outcome for a company. This dynamic might finally change in the next 12 months if the "private whales" take the leap into the public markets, starting with SpaceX in the summer.

Subscribe now

This will be critical to reshaping where venture dollars flow, as the top 20 deals in 2025 captured almost half of available liquidity. While there are fair justifications given the compute demands of the whales, realistically this is not beneficial for the broader startup ecosystem, if not outright detrimental.

The flip side is that the time to scale to an outstanding valuation is shortening dramatically. Whether this makes a material difference (nobody is paying $50B to acquire Cursor, and xAI had to be bailed out through SpaceX) is a different topic.

Well, I guess there's one person willing to pay even $60B.

This is a deceptively simple framework, but I think it’s a fair way of looking at things in the current context.

I’m looking for an SDR who wants to make it in AI-native security with Octane Security.

What you’ll do: Own business development with startups and companies bringing secure code review into their workflows. Hit your numbers and you’ll start leading full sales cycles fast, on a direct path to promotion.

What we’re looking for: You’ve worked with developers. You understand security, and you see what AI is doing for teams that move early on the right tools. You know nothing gets handed to you. You show up, you build, you rise to the occasion every day.

Location: NY or SF preferred. Remote considered if you have the track record and the habits to back it up.

Think that’s you? Reply to this email or DM me, referrals welcome too.

The problem for the application layer is defending why anyone should pay for workflows that can either be replicated internally or handled within the model provider itself. Harvey’s legaltech challenge is a clear example.

The problem for infrastructure plays (and cloud infrastructure software specifically) is whether their most important market (enterprise) sees meaningful benefits from adopting their platform. Cursor vs. Claude Code illustrates this well: both are valid approaches, with Cursor scaling coding agent access across organizations at enterprise-grade reliability and security, while Claude Code takes a different path to the same buyer.

Although it's easy to be bearish on the application layer, that view is mostly driven by investors being limited to public markets (the SaaSpocalypse) versus private markets, where we're seeing accelerated outcomes and growth at an extraordinary pace.

The most interesting metric in this new application layer remains ARR per employee. For all intents and purposes, AI-native orgs run at significantly higher productivity per employee, which is also why the deep headcount cuts at public companies haven't drawn much pushback (most agree, if silently, that those orgs are deeply bloated).

Large incumbents in cloud infrastructure software have struggled to offer obvious "best-in-class" products across most of the AI infra challenges like agent security and observability. This leaves significant room for new AI-native companies to enter and separate themselves from the pack.

Regardless of how much we try to highlight the wider ecosystem, at the center of the AI scaling era there are two companies driving the majority of growth: OpenAI and Anthropic. The race for who scales the most in run-rate this year is wide open and highly volatile.

The competition is not sitting still on fundraising, but in practical terms we have not yet seen real challengers emerge from the neolabs. The most interesting one to track is Safe Superintelligence, thanks to Ilya Sutskever, arguably the individual who has had the biggest impact on deep learning progress in AI over the last 15 years. His primary "opponent" is Demis Hassabis, the single most motivated leader in AI research today, running not just DeepMind but also Isomorphic Labs. Funnily enough, OpenAI was co-founded by Musk as the counterpoint to Demis.

Between the SaaSpocalypse and the significant liquidity in private markets, it's fair to say things can be perceived as frothy.

The counterargument is that the growth rate justifies the premium. While that's clearly a risky bet (stalled adoption would be brutal on valuations), it does make sense.

The "growth slowdown" could also come from a different angle than weaker customer demand: overcrowding in certain markets. High valuations only make sense if someone is the obvious winner of a category, and the difficulty curve to become one is much higher today.

Frothy or not, the opportunity to enter the market from a funding perspective is still very much open, and the same goes for talented individuals looking to move to an AI-native company. That said, the window for the highest returns is likely closing in the next 12 months.

For your average CEO, not saying anything in a 40-minute interview is a success. For founder mode leaders, every interaction is an opportunity to reframe the story (and ultimately the legend) of who they are and where their company is going.

Jensen is an interesting case. While he behaves very much in founder mode in every possible way, he’s been fairly conservative with the interview partners he picks. For the most part it’s been a steady line of allies, individuals who either have a financial interest in NVIDIA or would like to maintain a positive relationship with the CEO of the most valuable company in the world.

For whatever reason, he ended up being interviewed by Dwarkesh, who, while not fully “Effective Altruism pilled,” is still probably a better reflection of how certain intellectual circles in the Valley perceive the creative and destructive power of the army of GPUs we are deploying across data centers around the world. More interestingly, there is clearly a generational divide at play, as neither party seemed to accept the other’s framing of the world at face value.

The result is probably the first proper pushback we’ve seen Jensen get to his face in a while, and as these things tend to go, being put under pressure leads to some interesting unprompted insights.

I’ll skip the first part of the interview since it mostly repeats the basic pitch of the five-layer cake that I’ve already covered here:

Dwarkesh Patel

True. I want to ask about your competitors. If you look at the TPU, arguably two out of the top three models in the world, Claude and Gemini, were trained on TPU. What does that mean for Nvidia going forward?

Jensen Huang

We build a very different thing. What Nvidia built is accelerated computing, not a tensor processing unit. Accelerated computing is used for all kinds of things: molecular dynamics, quantum chromodynamics, data processing, data frames, structured data, and unstructured data. It’s also used for fluid dynamics and particle physics. In addition, we use it for AI.

Accelerated computing is much more diverse. Although AI is the conversation today and is obviously very important and impactful, computing is much broader than that. Nvidia has reinvented the way computing is done, moving from general-purpose computing to accelerated computing. Our market reach is far greater than any TPU or ASIC can possibly have. If you look at our position, we’re the only company that accelerates applications of all kinds. We have a gigantic ecosystem. So all kinds of frameworks and algorithms run on Nvidia.

Because our computers are designed to be operated by other people, anyone who’s an operator can buy our systems. With most of these home-built systems, you have to be your own operator because they were never designed to be flexible enough for others to operate. Because anybody can operate our systems, we’re in every cloud, including Google, Amazon, Azure, and OCI.

If you want to operate it to rent, you better have a large ecosystem of customers in many industries to be the offtakers. If you want to operate it for yourself, we obviously have the ability to help you operate it yourself, like we did for Elon with xAI. And because we can enable operators in any company and any industry, you could use it to build a supercomputer for scientific research and drug discovery at Lilly. We can help them operate their own supercomputer and use it for the entire diversity of drug discovery and biological sciences that we accelerate.

There are just a whole bunch of applications that we can address that you can’t do with TPUs. Nvidia built CUDA to be a fantastic tensor processing unit as well, but it also handles every life cycle of data processing, computing, AI, and so on. Our market opportunity is just a lot larger, and our reach is a lot greater. Because we support every application in the world now, you can build Nvidia systems anywhere and know that there will be customers for it. It’s a very different thing.

The biggest challenge to NVIDIA’s long-term success remains whether selling GPUs is a real, durable advantage. Jensen’s view is that what they do is provide accelerated computing, the new paradigm for how technology will work. This story has some gaps, the most obvious being that if the focus is acceleration, there are other companies offering a much more focused product for specific use cases.

Dwarkesh Patel

This is going to be a long question. You have spectacular revenue, and you’re not making $60 billion a quarter from pharma and quantum. You’re making it because AI is an unprecedented technology that is growing unprecedentedly fast.

The question then is what is best for AI specifically. I’m not in the details, but I talk to my AI researcher friends and they say, “Look, when I use a TPU, it’s this big systolic array that’s perfect for doing matrix multiplies, whereas a GPU is very flexible. It’s great when you have lots of branching or irregular memory access.”

But what is AI? It’s just these very predictable matrix multiplies again and again and again. You don’t have to give up any die area for warp schedulers or switches between threads and memory banks. And the TPU is really optimized for the bulk of this growth in revenue and use case for compute that is coming online right now. I wonder how you react to that.

Jensen Huang

Matrix multiplies are an important part of AI, but they’re not the only part. If you want to come up with a new attention mechanism, disaggregate in a different way, or invent a whole new type of architecture altogether—like a hybrid SSM—you want an architecture that’s generally programmable. If you want to create a model that fuses diffusion and autoregressive techniques, you want an architecture that’s just generally programmable. We run everything you can imagine. That’s the advantage. It allows for the invention of new algorithms a lot more easily, because it’s a programmable system.

The ability to invent new algorithms is really what makes AI advance so quickly. TPUs, like anything else, are impacted by Moore’s Law, which we know is increasing by about 25% per year. The only way to really get 10x or 100x leaps is to fundamentally change the algorithm and how it’s computed every single year.

That’s Nvidia’s fundamental advantage. The only reason we were able to make Blackwell to Hopper 50x… When I first announced Blackwell was going to be 35x more energy efficient than Hopper, nobody believed it. Then Dylan wrote an article saying I sandbagged, and it’s actually fifty times. You can’t reasonably do that with just Moore’s Law. The way we solve that problem is with new models, like MoEs, that are parallelized, disaggregated, and distributed across a computing system. Without the ability to really get down and come up with new kernels with CUDA, it’s really hard to do.

It’s the combination of the programmability of our architecture and the fact that Nvidia is an extreme co-design company. We can even offload some of the computation into the fabric itself, like NVLink, or into the network with Spectrum-X. We could affect change across the processors, the system, the fabric, the libraries, and the algorithm simultaneously. Without CUDA to do that, I wouldn’t even know where to start.

The first pushback comes on the narrative of “what’s the best fit for AI workloads.” Jensen’s retort is that TPUs are designed for a specific software architecture, so they can’t possibly be used to scale AI because they are not generally programmable. The claimed advantage for NVIDIA is that while the latest hardware supports smaller performance jumps, by working with the labs on properly running models with novel architectures, the performance jumps are much higher.

Dwarkesh Patel

That makes a lot of sense. I guess the thing I’m curious about is whether those advantages matter a lot to your main customers. There’s many people for whom they might matter. The kind of person who can actually build their own software stack makes up most of your revenue. Especially if you go to a world where AI is getting especially good at the things which have tight verification loops where you can RL on them…. This question of how do you write a kernel that does attention or MLP the most efficiently across a scale up? It’s a very verifiable sort of feedback loop.

Can all the hyperscalers write these custom kernels for themselves? Nvidia still has great price performance, so they might still prefer to use Nvidia. But then the question is, does it just become a question of who is offering the best specs, the best flops and memory bandwidth for a given dollar. Whereas historically Nvidia has just had, and still has, the best margins in all of AI across hardware and software, +70%, because of this CUDA moat. And the question is, can you sustain those margins if for most of your customers, they can actually afford to build, instead of the CUDA moat?

Jensen Huang

The number of engineers we have assigned to these AI labs is insane, working with them, optimizing their stack. The reason for that is because nobody knows our architecture better than we do. These architectures are not as general purpose as a CPU. A CPU is kind of like a Cadillac. It’s a nice cruiser. It never goes too fast. Everybody drives it pretty well. It’s got cruise control, and everything’s easy. But in a lot of ways, Nvidia’s GPUs, accelerators, are like F1 racers. I could imagine everybody’s able to drive it at a hundred miles an hour, but it takes quite a bit of expertise to be able to push it to the limit. We use a ton of AI to create the kernels that we have.

I’m pretty sure we’re going to still be needed for quite some time. Our expertise helps our AI lab partners to get another 2x out of their stack easily oftentimes. It’s not unusual that by the time we’re done optimizing their stack or optimizing a particular kernel, their model sped up by 3x, 2x, 50%. That’s a huge number, especially when you’re talking about the install base of the fleet that they have, of all the Hoppers and Blackwells that they have. When you increase it by a factor of two, that doubles the revenues. That directly translates to revenues.

Nvidia’s computing stack is the best performance per TCO in the world, bar none. Nobody can demonstrate to me that any single platform in the world today has a better performance-TCO ratio. Not one company. In fact, the benchmarks that are out there. Dylan’s InferenceMAX is sitting out there for everybody to use, and not one… TPU won’t come, Trainium won’t come.

I encourage them to use InferenceMAX and demonstrate their incredible inference cost. It’s really hard. Nobody wants to show up. MLPerf. I would welcome Trainium to demonstrate their 40% that they claim all the time. I would love to hear them demonstrate the cost advantage of TPUs. It makes no sense in my mind. It makes absolutely zero sense. On first principles, it makes no sense.

So I think the reason why we’re so successful is simply because our TCO is so great. Secondly, you say 60% of our customers are the top five, but most of that business is external. For example, most of Nvidia in AWS is for external customers, not internal use. Most of our customers at Azure, obviously all of our customers are external. All of our customers at OCI are external, not internal use. The reason why they favor us is because our reach is so great. We can bring them all of the great customers in the world. They’re all built on Nvidia. And the reason why all these companies are built on Nvidia is because our reach and our versatility is so great.

So I think the flywheel is really install base, the programmability of our architecture, the richness of our ecosystem, and the fact that there’s so many AI companies in the world. There’s tens of thousands of them now. If you were one of those AI startups, what architecture would you choose? You would choose an architecture that’s most abundant. We’re the most abundant in the world. You’d choose the one that has the largest installed base. We’re the largest install base. And you’d choose the one that has a rich ecosystem.

So that’s the flywheel. That’s the reason why, between the combination of: one, our perf per dollar is so great that they have the lowest cost tokens. Second, our perf per watt is the highest in the world. So if one of these companies, if our partners, built a one gigawatt data center, that one gigawatt data center better deliver the maximum amount of revenues and number of tokens, which directly translates to revenues. You want it to generate as many tokens as possible, maximize the revenues for that data center. We are the highest tokens per watt architecture in the world. Lastly, if your goal is to rent the infrastructure, we have the most customers in the world. So that’s the reason why the flywheel works.

Tens of thousands of AI startups and companies is a generous count, unless we stretch the definition to everybody using AI. The practical reality is that very few companies will run their own models, so the chokepoint is the handful of large labs and compute providers, currently around 20 relevant companies. Are they all consolidated on CUDA and NVIDIA?

Well…sort of. The most dependent on NVIDIA hardware are the neoclouds, and to a certain extent OpenAI and Microsoft. The Chinese labs are clearly using a lot of legally (and illegally) acquired NVIDIA infrastructure, but are also being pushed by the Chinese state to acquire local compute. Among the hyperscalers, both GCP and AWS are placing their big bets on their own custom hardware, while Azure is trying to play along but failing on the technical side.

Dwarkesh Patel

Interesting. I guess the question comes down to, what is the actual market structure here? Because even if there’s other companies… There could have been a world where there’s tens of thousands of AI companies that have roughly equal share of compute. But even through these five hyperscalers, really the people on Amazon using the compute are Anthropic, OpenAI, and these big foundation labs who can themselves afford and have the ability to make different accelerators work.

Jensen Huang

No, I think your premise is wrong.

Dwarkesh Patel

Maybe. But let me ask you a slightly different question.

Jensen Huang

Come back and make me correct your premise.

Uh, oh.

Dwarkesh Patel

Okay. Let me just ask you a different question.

Jensen Huang

But still make sure to make me come back and fix because it’s just too important to AI. It’s too important to the future of science. It’s too important to the future of the industry. That premise… Look —

Dwarkesh Patel

Let me just finish the question and then we can address it together.

Jensen Huang

Yeah.

Dwarkesh Patel

If all these things are true about price, performance, and performance per watt, et cetera, are true, why do you think it is the case that, say, Anthropic for example, just announced a couple days ago they have a multi-gigawatt deal with Broadcom and Google for TPUs and majority of their compute?

Obviously for Google, TPU is a majority of compute. So if I look at these big AI companies, it seems like a lot of their compute… There was some point where it’s all Nvidia and now it’s not. So I’m curious how to square, if these things are true on paper, why are they going with other accelerators?

Jensen Huang

Anthropic is a unique instance, not a trend. Without Anthropic, why would there be any TPU growth at all? It’s 100% Anthropic. Without Anthropic, why would there be Trainium growth at all? It’s 100% Anthropic. I think that’s fairly well known and well understood. It’s not that there’s an abundance of ASIC opportunities. There’s only one Anthropic.

Dwarkesh Patel

But OpenAI’s deals with AMD… They’re building their own Titan accelerator.

Jensen Huang

Yeah, but I think we could all acknowledge they’re vastly Nvidia. We’re going to still do a lot of work together. I’m not offended by other people using something else and trying things. If they don’t try these other things, how would they know how good ours is? Sometimes you’ve got to be reminded of it. We have to continuously earn the position that we’re in.

There are always big claims. Look at the number of ASICs that have been canceled. Just because you’re going to build an ASIC… You still have to build something better than Nvidia. It’s not that easy building something better than Nvidia. It’s not sensible, actually. Nvidia’s got to be missing something, seriously. Because of our scale, our velocity, we’re the only company in the world that’s cranking it out every single year. Big leaps, every single year.

Dwarkesh Patel

I guess their logic is, “Hey, it doesn’t need to be better. It just needs to be not more than 70% worse,” because they’re paying you 70% margins.

Jensen Huang

No, don’t forget, even in ASICs margins are really quite high. Nvidia’s margin is 70%, let’s say. But ASIC margins are 65%. What are you really saving?

Dwarkesh Patel

Oh, you mean from Broadcom or something like that?

Jensen Huang

Yeah, sure. You’ve got to pay somebody. I think the ASIC margins are incredibly good, from what I can tell. They believe it too. They’re quite proud of their incredible ASIC margins.

So, you asked the question why. A long time ago, we just didn’t have the ability to do it. At the time, I didn’t deeply internalize how difficult it would be to build a foundation AI lab like OpenAI and Anthropic, and the fact that they needed huge investments from the supplier themselves. We just weren’t in a position to make the multi-billion dollar investment into Anthropic so that they could use our compute. But Google and AWS were. They put in huge investments in the beginning so that Anthropic, in return, used their compute. We just weren’t in a position to do that at the time.

I would say my mistake is I didn’t deeply internalize that they really had no other options, that a VC would never put in $5-10 billion of investment into an AI lab with the hopes of it turning out to be Anthropic. So that was my miss. But even if I understood it, I don’t think we would’ve been in a position to do that at the time. But I’m not going to make that same mistake again.

I’m delighted to invest in OpenAI, and I’m delighted to help them scale, and I believe it’s essential to do so. And then, when I was able to, when Anthropic came to us, I’m delighted to be an investor, delighted to help them scale. We just weren’t, at the time, able to do it. If I could rewind everything—and Nvidia could have been as big back then as we are now—I would’ve been more than happy to do it.

Jensen is playing an interesting game here. While Anthropic might be "the exception to the rule," it's also the frontier lab growing at an unprecedented pace and gobbling up massive amounts of non-NVIDIA compute. His way of handling this is by…investing in all of the up-and-coming competitors.

Dwarkesh Patel

This is actually quite interesting. For many years Nvidia has been the company in AI making money, making lots of money. Now you’re investing it. It’s been reported that you’ve done up to $30 billion in OpenAI and $10 billion in Anthropic. But now their valuations have increased, and I’m sure they’ll continue to increase.

So if over these many years you were giving them the compute, you saw where it was headed, and they were worth like one tenth what they’re worth now a couple years ago—or even a year ago in some cases and you had all this cash — there’s a world where either Nvidia themselves becomes a foundation lab, does a huge investment to make that possible, or has made the deals you’ve made now at current valuations much earlier on. And you had the cash to do it. So I am curious, actually, why not have done it earlier?

Jensen Huang

We did it as soon as we could have. We did it as soon as we could have, and if I could have, I would’ve done it even earlier. At the time that Anthropic needed us to do it, we just weren’t in a position to do it. It wasn’t in our sensibility to do so.

Dwarkesh Patel

How so? Was it like a cash thing?

Jensen Huang

Yeah, the level of investment. We had never invested outside the company at the time, and not that much. We didn’t realize we needed to. I always thought that they could just go raise from VCs, for God’s sakes, like all companies do. But what they were trying to do couldn’t have been done through VCs. What OpenAI wanted to do couldn’t have been done through VCs. I recognize that now. I didn’t know it then.

But that’s their genius. That’s why they’re smart. They realized then that they had to do something like that. And I’m delighted that they did. Even though we caused Anthropic to have to go to somebody else, I’m still happy that it happened. Anthropic’s existence is great for the world. I’m delighted for it.

Dwarkesh Patel

I guess you still are making a ton of money, and you’re making way more money quarter after quarter.

Jensen Huang

It’s still okay to have regrets.

Regrets, I have a few.

I’ve talked before about the most obvious mistake analysts and investors make when trying to evaluate the NVIDIA moat, which is the assumption that everybody will keep competing in the same way and somehow the most efficient hardware will magically win back market share. This ignores the elephant in the room: Jensen doesn’t care about success, his focus is on not losing. Every loss, however small or imagined, is painfully replayed in his mind until he solves it by winning. This is a common trait of a certain type of high-performance personality, but this time it’s playing out with the power of hundreds of billions of liquid dollars.

He will either outspend you, out-R&D you, or buy you. The only way to beat him is to hope he runs out of cash or dies.

Subscribe now

Dwarkesh Patel

So the question still arises. Okay, now that we’re here and you have all this money that you keep making, what should Nvidia be doing with it? There’s one answer which is that there’s this whole middleman ecosystem that has popped up for converting CapEx into OpEx for these labs so that they can rent compute. Because the chips are really expensive, they make a lot of money over their lifetime because the AI models are getting better. So the value that they generate, their tokens, is increasing, but they’re expensive to set up. Nvidia has the money to do the CapEx. In fact, it’s been reported, you are backstopping CoreWeave up to $6.3 billion and have invested $2 billion.

Why doesn’t Nvidia become a cloud themselves? Why doesn’t it become a hyperscaler themselves and rent this compute out? You have all this cash to do it.

Jensen Huang

This is a philosophy of the company, and I think it’s wise. We should do as much as needed, as little as possible. What that means is, the work that we do with building our computing platform, if we don’t do it, I genuinely believe it doesn’t get done. If we didn’t take the risk that we take—if we didn’t build NVLink the way we built it, if we didn’t build the whole stack, if we didn’t create the ecosystem the way we did, if we didn’t dedicate ourselves to 20 years of CUDA while losing money most of that time—if we didn’t do it, nobody else would have done it.

If we didn’t create all the CUDA-X libraries so that they’re all domain-specific… A decade and a half ago, we pushed into domain-specific libraries because we realized that if we didn’t create these domain-specific libraries, whether it’s for ray tracing or image generation or even the early works of AI, these models, if we didn’t create them, for data processing, structured data processing, or vector data processing, if we didn’t create them, nobody would. I am completely certain of that. We created a library for computational lithography called cuLitho. If we didn’t create it, nobody would have. So accelerated computing wouldn’t advance the way it has if we didn’t do what we did.

So we should do that. We should dedicate our company, all of our might, wholeheartedly to go do that. However, the world has lots of clouds. If I didn’t do it, somebody would show up. So following the recipe, the philosophy, of doing as much as needed but as little as possible—as little as possible—that philosophy exists in our company today. Everything I do, I do it with that lens.

In the case of clouds, if we didn’t support CoreWeave to exist, these neoclouds, these AI clouds, wouldn’t exist. If we didn’t help CoreWeave exist, they would not exist. If we didn’t support Nscale, they wouldn’t be where they are today. If we didn’t support Nebius, they wouldn’t be what they are today. Now they’re doing fantastically.

Is that a business model [inaudible]? We should do as much as needed, as little as possible. So we invest in our ecosystem because I want our ecosystem to thrive. I want the architecture, and AI, to be able to connect with as many industries as possible, as many countries as possible, and make it possible for the planet to be built on AI and to be built on the American tech stack. That vision is exactly what we’re pursuing.

Now, one of the things that you mentioned… There are so many great, amazing foundation model companies, and we try to invest in all of them. This is another thing that we do. We don’t pick winners. We need to support everyone. It’s part of our joy of doing so. It’s imperative to our business. But we also go out of our way not to pick winners. So when I invest in one of them, I invest in all of them.

Dwarkesh Patel

Why do you go out of your way not to pick winners?

Jensen Huang

Because it’s not our job to, number one. Number two, when Nvidia first started, there were 60 3D graphics companies. We are the only one that survived. If you would have taken those 60 graphics companies and asked yourself which one was going to make it, Nvidia would be at the top of that list not to make it.

This is long before you, but Nvidia’s graphics architecture was precisely wrong. It’s not a little bit wrong. We created an architecture that was precisely wrong, and it was an impossible thing for developers to support. It was never going to make it. We reasoned about it from good first principles, but we ended up with the wrong solution. Everybody would have counted us out. And here we are.

So I have enough humility to recognize that. Don’t pick winners. Either let them all take care of themselves, or take care of all of them.

I don't think it's true that he doesn't pick winners. I think he funds whoever he believes will end up being a winner in segments that matter to NVIDIA. It's not obvious enough right now, but the practical reality is that a large war chest of money and GPUs will always tilt the odds, unless they completely misread the company's ability to execute.

Dwarkesh Patel

One thing I didn’t understand is you said, “Look, we’re not prioritizing these neoclouds just because they are neoclouds and we want to prop them up.” But you also listed a bunch of neoclouds and said they wouldn’t exist if it wasn’t for NVIDIA. How are those two things compatible?

Jensen Huang

First of all, they need to want to exist, and they come to ask us for help. When they want to exist and they have a business plan, expertise, and the passion for it… They obviously have to have some capabilities themselves. But if, at the end of the day, they need some investment in order to get it off the ground, we would be there for them. But the sooner they get their flywheel going...

Your question was, “Do we want to be in the financing business?” The answer is no. There are people in the financing business, and we’d rather work with all the people in the financing business than be a financier ourselves. Our goal is to focus on what we do, keep our business model as simple as possible, and support our ecosystem.

When someone like OpenAI needs an investment of a $30 billion scale because it’s still before their IPO, and we deeply believe in them and I deeply believe that they’re going to be an… Well, they’re an extraordinary company already today. They’re going to be an incredible company. The world needs them to exist. The world wants them to exist. I want them to exist. They have the wind at their back. Let’s support them and let them scale. Those investments we’ll do because they need us to do it. But we’re not trying to do as much as possible. We’re trying to do as little as possible.

He is doing as little as possible, but in a world where capital intensity matters more and more, it's obvious that pushing significant liquidity toward a player is likely to influence the outcome. He also stated earlier that NVIDIA is spending significant technical resources co-designing with its customers. What are the odds he isn't doing the same with his investments, who also happen to be customers, since none of the liquidity comes without strings attached?

Dwarkesh Patel

This may be an obvious question, but we’ve lived many years in this situation where there’s a shortage of GPUs, and it’s grown now because models are getting better.

Jensen Huang

We have a shortage of GPUs.

Dwarkesh Patel

Yes. Nvidia is known for divvying up the scarce allocation, not just based on high bidder, but rather on, “Hey, we want to make sure that these neoclouds exist. Let’s give some to CoreWeave, let’s give some to Crusoe, let’s give some to Lambda.” Why is it good for Nvidia? First of all, would you agree with this characterization of fracturing the market?

Jensen Huang

No. No. Your premise is just wrong. We’re sufficiently mindful about these things. We’re very mindful about these things. First of all, if you don’t place a PO, all the talking in the world won’t make a difference. Until we get a PO, what are we going to do? So the first thing is, we work really hard with everybody to get a forecast done, because these things take a long time to build, and the data centers take a long time to build. We align ourselves with demand and supply and things like that through forecasting. Okay? That’s job number one.

Number two, we’ve tried to forecast with as many people as possible, but in the final analysis, you still have to place an order. Maybe, for whatever reason, you didn’t place your order. What can I do? At some point, first in, first out. But beyond that, if you’re not ready because your data center’s not ready, or certain components aren’t ready to enable you to stand up a data center, we might decide to serve another customer first. That’s just maximizing the throughput of our own factory. We might do some adjustments there.

Aside from that, the prioritization is first in, first out. You’ve got to place a PO. If you don’t place a PO… Now, of course, there are stories about that. For example, all of this kind of started from an article about Larry and Elon having dinner with me where they begged for GPUs. That never happened. We absolutely had dinner. We absolutely had dinner, and it was a wonderful dinner. At no time did they beg for GPUs. They just had to place an order. Once they place an order, we do our best to get the capacity to them. We’re not complicated.

“It puts the PO in the basket or else it gets the hose again.”

Dwarkesh Patel

Okay. So it sounds like there’s a queue, and then based on whether your data center is ready and when you place a purchase order, you get them at a certain time. But it still doesn’t sound like the highest bidder just gets it. Is there a reason to do it…?

Jensen Huang

We never do that.

Dwarkesh Patel

Okay.

Jensen Huang

We never do.

Dwarkesh Patel

Why not just do high bidder?

Jensen Huang

Because it’s a bad business practice. You set your price and then people decide to buy it or not. I understand that others in the chip industry change their prices when demand is higher, but we just don’t. That’s just never been a practice of ours. You can count on us. I prefer to be dependable, to be the foundation of the industry. You don’t need to second-guess. If I quoted you a price, we quoted you a price. That’s it. If demand goes through the roof, so be it.

Dwarkesh Patel

On the other end, that’s why you have a productive relationship with TSMC, right?

Jensen Huang

Yeah, Nvidia’s been in business with them for, I guess, coming up on 30 years. Nvidia and TSMC don’t have a legal contract. There’s always some rough justice. Sometimes I’m right, sometimes I’m wrong. Sometimes I got a better deal, sometimes I got a worse deal. But overall, the relationship is incredible. I can completely trust them. I can completely depend on them.

One of the things you can count on with Nvidia is that this year, Vera Rubin is going to be incredible. Next year, Vera Rubin Ultra will come. The year after that, Feynman will come. And the year after that, I haven’t introduced the name yet. Every single year you can count on us. You’re going to have to go find another ASIC team in the world—pick your ASIC team—where you can say, “I can bet the farm, I can bet my entire business that you will be here for me every single year. Your token cost will decrease by an order of magnitude every single year. I can count on it like I can count on the clock.”

I just said something about TSMC. For no other foundry in history can you possibly say that. You can say that about Nvidia today. You can count on us every single year. If you would like to buy a billion dollars worth of AI factory compute, no problem. If you’d like to buy a hundred million dollars, no problem. You’d like to buy $10 million, or just one rack, not a problem. Or just one graphics card, okay, no problem. If you would like to place an order for a $100 billion of AI factory, no problem. We’re the only company in the world where you can say that today.

I can say that about TSMC as well. I want to buy one, buy 1 billion, no problem. We just have to go through the process of planning for it, and all the things that mature people do. So I think this ability for Nvidia to be the foundation of the world’s AI industry, this is a position that has taken us a couple of decades to arrive at. Enormous commitment, enormous dedication. The stability of our company, the consistency of our company, is really important.

According to Jensen, the reason you want to do business with him is that he won’t screw you over once the PO is in, and he will make sure his ecosystem delivers, every single year.

I think this statement is fairly accurate based on how things have played out.

Dwarkesh Patel

Okay. I want to ask about China. I actually don’t know what I think about whether it’s good to sell chips to China or not, but I like to play devil’s advocate against my guests. So when Dario was on, who supports export controls, I asked him, why can’t America and China both have a country of geniuses in the datacenter? But since you’re on the opposite side, I’ll ask you in the opposite way.

One way to think about it is, Anthropic actually announced a couple days ago Mythos Preview. This model Mythos, they’re not even releasing publicly because they say it has such cyber-offensive capabilities that we don’t think the world is ready until we make sure these zero-days are patched up. But they say it found thousands of high-severity vulnerabilities across every major operating system, every browser. It found one in OpenBSD, which is this operating system that’s been specifically designed to not have zero days. It found one that’s existed for 27 years.

So if Chinese companies and Chinese labs and the Chinese government had access to the AI chips to train a model like Claude Mythos with these cyber-offensive capabilities and run millions of instances of it with more compute, the question is, is that a threat to American companies, to American national security?

Jensen Huang

First of all, Mythos was trained on fairly mundane capacity, and a fairly mundane amount of it. By an extraordinary company. The amount of capacity and the type of compute it was trained on is abundantly available in China. So you just have to first realize that chips exist in China.

They manufacture 60% of the world’s mainstream chips, maybe more. It’s a very large industry for them. They have some of the world’s greatest computer scientists. As you know, most of the AI researchers in all of these AI labs are Chinese. They have 50% of the world’s AI researchers. So the question is, considering all the assets they already have—they have an abundance of energy, they have plenty of chips, they’ve got most of the AI researchers—if you’re worried about them, what is the best way to create a safe world?

Victimizing them, turning them into an enemy, likely isn’t the best answer. They are an adversary. We want the United States to win. But I think having a dialogue and having research dialogue is probably the safest thing to do. This is an area that is glaringly missing because of our current attitude about China as an adversary. It is essential that our AI researchers and their AI researchers are actually talking. It is essential that we try to both agree on what not to use the AI for.

With respect to finding bugs in software, of course, that’s what AI is supposed to do. Is it going to find bugs in a lot of software? Of course. There are lots and lots of bugs. There are lots of bugs in the AI software. That’s what AI is supposed to do, and I’m delighted that AI has reached a level where it could help us be so much more productive.

One of the things that is underemphasized is the richness of the ecosystem around cybersecurity, AI cybersecurity and AI security and AI privacy and AI safety. There’s a whole ecosystem of AI startups that are trying to create this future for us, where you have one AI agent that’s incredible, surrounded by thousands of AI agents, keeping it safe, keeping it secure. That future surely is going to happen.

The idea that you’re going to have an AI agent running around with nobody watching after it is kind of insane. We know very well that this ecosystem needs to thrive. It turns out this ecosystem needs open source. This ecosystem needs open models. They need open stacks so that all of these AI researchers and all these great computer scientists can go build AI systems that are as formidable and can keep AI safe. So one of the things that we need to make sure that we do is we keep the open source ecosystem vibrant. That can’t be ignored. A lot of that is coming out of China. We ought to not suffocate that.

With respect to China, of course we want the United States to have as much computing as possible. We’re limited by energy, but we’ve got a lot of people working on that. We’ve got to not make energy a bottleneck for our country. But what we also want is to make sure that all the AI developers in the world are developing on the American tech stack, and making the contributions, the advancements of AI—especially when it’s open source—available to the American ecosystem. It would be extremely foolish to create two ecosystems: the open source ecosystem, and it only runs on a foreign tech stack, and a closed ecosystem that runs on the American tech stack. I think that would be a horrible outcome for the United States.

I think that while it’s important to note that Jensen’s story is a great example of a family achieving the American Dream, and that NVIDIA is a credit to the ingenuity and talent in the United States, it’s very difficult not to point out the incentives that might be influencing his opinions on China.

His framing is that NVIDIA needs to be the unified layer that both Western and Eastern AI run on. Failing to do so is going to be disastrous for the United States. Some things to consider around this thesis:

• It’s likely to be much more devastating for Taiwan, his country of origin and NVIDIA’s main manufacturing pillar.

• Taiwan has historically represented the vast majority of NVIDIA's advanced GPU manufacturing. While it would be unfair to blame Jensen for the disastrous decade of mismanagement at Intel and its failure to win a meaningful portion of NVIDIA's fabrication, it's also not something that can be fully ignored. He's played his role in reaching this state of play, where any conflict surrounding Taiwan would lead to an incredible short-term shock in the global economy. The fact that Musk is trying to solve this bottleneck with Intel and the TeraFab project speaks volumes: Jensen isn't willing to meaningfully diversify the supply chain (and, as a consequence, weaken the "silicon shield" around Taiwan), so others are stepping in to do it.

Dwarkesh Patel

Since there are a lot of things, let me just triage the response. I think the concern, going back to the flop difference in the hacking, is yes, they have compute, but there’s some estimates that because they’re at 7nm—they don’t have EUVs because of chip-making export controls—the amount of flops they’re able to actually produce, they have one tenth the amount of flops that the US has.

So with that, could they eventually train a model like Mythos? Yes. But the question is, because we have more flops, American labs are able to get to these levels of capabilities first. Because Anthropic got to it first, they say, “Okay, we’re going to hold onto it for a month while all these American companies, we’ll give them access to it. They’re going to patch up all their vulnerabilities, and now we release it.”

Furthermore, even if they train a model like this, the ability to deploy it at scale… If you had a cyber hacker, it’s much more dangerous if they have a million of them versus a thousand of them. So that inference compute really matters a lot. In fact, the fact that they have so many AI researchers who are so good is the thing that makes it so scary, because what is it that makes those engineer researchers more productive? It’s compute.

If you talk to any AI lab in America, they say the thing that’s bottlenecking them is compute. There are quotes from the DeepSeek founder, or Qwen leadership or whatever. They say the thing they’re bottlenecked on is compute. So then the question is, isn’t it better that we get American companies, because they have more compute, to get to the Mythos-level capabilities first, prepare our society for it, before China can get to it because, they have less compute?

Jensen Huang

We should always be first and we should always have more. But in order for that outcome you described to be true, you have to take it to the extremes. They have to have no compute. If they have some compute, the question is how much is needed?

The amount of compute they have in China is enormous. You’re talking about the country that is the second largest computing market in the world. If they want to aggregate their compute, they’ve got plenty of compute to aggregate.

Dwarkesh Patel

But is that true? People do these estimates and they’re like, “SMIC is actually behind on the process nodes.”

Jensen Huang

I’m about to tell you.

Dwarkesh Patel

Okay.

Jensen Huang

The amount of energy they have is incredible. Isn’t that right? AI is a parallel computing problem, isn’t it? Why can’t they just put 4x, 10x, as many chips together because energy’s free? They have so much energy. They have datacenters that are sitting completely empty, fully powered. You know they have ghost cities, they have ghost datacenters too. They have so much infrastructure capacity. If they wanted to, they just gang up more chips, even if they’re 7nm.

Their capacity of building chips is one of the largest in the world. The semiconductor industry knows that they monopolize mainstream chips. They have over-capacity, they have too much capacity. So the idea that China won’t be able to have AI chips is completely nonsense.

Now, of course, if you ask me, would the United States be further ahead if the entire world had no compute at all? But that’s just not an outcome. That’s not a scenario that’s true. They have plenty of compute already. The amount of threshold they need for the concern you’re worried about, they’ve already reached that threshold and beyond.

So I think you misunderstand that AI is a five-layer cake, and at the lowest layer is energy. When you have an abundance of energy, it makes up for chips. If you have an abundance of chips, it makes up for energy. For example, the United States is scarce on energy, which is the reason why Nvidia has to keep advancing our architecture and do this extreme co-design so that with the few chips that we ship—with the few chips, because the amount of energy is so limited—our throughput per watt is off the charts.

But if your amount of watts is completely abundant, it’s free, what do you care about performance per watt for? You get plenty. You can use old chips to do. So 7nm chips are essentially Hopper. The ability for Hopper… I’ve got to tell you, today’s models are largely trained on Hopper, Hopper generation. So 7nm chips are plenty good. The abundance of energy is their advantage.

Dwarkesh Patel

But then there’s a question of whether they can actually manufacture enough chips.

Jensen Huang

But they do. What’s the evidence? Huawei just had the largest single year in the history of their company.

Dwarkesh Patel

How many chips did they ship?

Jensen Huang

A ton. Millions. Millions is way more than Anthropic has.

Things start to derail here, as Jensen’s story becomes conflicting. On one hand, the Chinese are fully self-sufficient on AI, have massive amounts of energy, and can achieve the same outcomes as the US AI labs even on inefficient hardware. Apparently Claude Mythos can easily be replicated right now!

Dwarkesh Patel

But that doesn’t change the fact that you need EUV for the most advanced HBM.

Jensen Huang

Not true. Not at all true. You could gang them together, just like we gang them together with NVL72. They’ve already demonstrated silicon photonics, connecting all of this compute together into one giant supercomputer. Your premise is just wrong.

The fact of the matter is, their AI development is going just fine. The best AI researchers in the world, because they’re limited in compute, they also come up with extremely smart algorithms. Remember, I just said that Moore’s law is advancing about 25% per year. However, through great computer science, we could still improve algorithm performance by 10x. What I’m saying is that great computer science is where the lever is.

There is no question, MoE is a great invention. There’s no question, all the incredible attention mechanisms reduce the amount of compute. We have got to acknowledge that most of the advances in AI came out of algorithm advances, not just the raw hardware. Now, if most advances came from algorithms and computer science and programming, tell me that their army of AI researchers is not their fundamental advantage. We see it. DeepSeek is not an inconsequential advance. The day that DeepSeek comes out on Huawei first, that is a horrible outcome for our nation.

Dwarkesh Patel

Why is that? Because currently you can have a model like DeepSeek that can run on any accelerator, if it’s open source. Why would that stop being the case in the future?

Jensen Huang

Suppose it doesn’t. Suppose it’s optimized for Huawei, suppose it’s optimized for their architecture. It would put ours at a disadvantage. You described a situation that I perceive to be good news. A company developed software, developed an AI model, and it runs best on the American tech stack. I saw that as good news. You set it up as a premise that it was bad news. I’m going to give you the bad news, that AI models around the world are developed and they run best on non-American hardware. That is bad news for us.

Dwarkesh Patel

I guess I just don’t see the evidence that there’s these huge disparities that would prevent you from switching accelerators. American labs are running their models across all the clouds, across all the different accelerators—

Jensen Huang

I am the evidence. You take a model that’s optimized for Nvidia and you try to run it on something else.

Dwarkesh Patel

But American labs do that.

Jensen Huang

And they don’t run better. Nvidia’s success is perfect evidence. The fact that AI models are created on our stack, run best on our stack, how is that illogical to understand?

Dwarkesh Patel

Anthropic’s models are run on GPUs, they’re run on Trainium, they’re run on TPUs.

Jensen Huang

A lot of work has to go into it to change. But go to the global south, go to the Middle East. Coming out of the box, if all of the AI models run best on somebody else’s tech stack, you’ve got to be arguing some ridiculous claim right now that that’s a good thing for the United States.

Dwarkesh Patel

But I guess I don’t understand the argument. Say Chinese companies get to the next Mythos first. They find all the security vulnerabilities in American software first, but they can do it on Nvidia hardware and they ship it to the global south. They do it on Nvidia hardware. How is that good? Okay, it runs on Nvidia hardware—

Jensen Huang

It’s not good. It’s not good.

Dwarkesh Patel

Right.

Jensen Huang

It’s not good. So let’s not let it happen.

Dwarkesh Patel

Why do you think it’s perfectly fungible, that if you didn’t ship them compute it would exactly be replaced by Huawei? They are behind, right? They have worse chips than you.

Jensen Huang

It’s completely… There’s evidence right now. Their chip industry’s gigantic.

Dwarkesh Patel

You can just look at the flop or bandwidth or memory comparisons between the H200 and the Huawei 910C. It’s like half to a third.

Jensen Huang

They use more of it. They use twice as many.

Dwarkesh Patel

It seems like your argument is they have all this energy that’s ready to go, right? And they need to fill it with chips.

Jensen Huang

And they’re good at manufacturing.

Dwarkesh Patel

And I’m sure eventually they would be able to just out-manufacture everybody. But there are these few critical years.

Jensen Huang

What is the critical year you’re talking about?

Dwarkesh Patel

These next few years. We’ve got these models that are going to be able to do all the cyber attacks.

Jensen Huang

In that case, if the next years are critical, then we have to make sure that all of the world’s AI models are built on the American tech stack, in these critical years.

Dwarkesh Patel

If they’re built on the American tech stack, how would that prevent them, if they have more advanced capabilities, from launching the Mythos-equivalent cyber attacks?

Jensen Huang

There’s no guarantee either way.

So NVIDIA has to sell chips to China because otherwise China will develop models that can cripple Western infrastructure and that will work best on local hardware. If those same models run on NVIDIA hardware, however, that's fine, because…they will likely still cripple Western infrastructure.

Jensen Huang

Listen, why are you causing one layer of the AI industry to lose an entire market so that you could benefit another layer of the AI industry? There are five layers and every single layer has to succeed. The layer that has to succeed most is actually the AI applications. Why are you so fixated on that AI model? That one company? For what reason?

Dwarkesh Patel

Because those models make possible these incredibly offensive capabilities, and you need compute to run them.

Jensen Huang

The energy, the chips, and the ecosystem of AI researchers make it possible.

Dwarkesh Patel

Okay, stepping back, it has to be the case that China is able to build enough 7nm capacity. And remember, they’re still stuck on 7nm while you’ll move on to 3nm and then 2nm or 1.6nm with Feynman. So while you’re on 1.6nm, they’re still going to be on 7nm, and they have to produce enough of it to make up for the shortfall. They have so much energy that the more chips you give them, the more compute they’d have. So it comes out as a question of, ultimately they are getting more compute. Compute is an input to training and inference—

Jensen Huang

Listen, I just think you speak in absolutes. I think the United States ought to be ahead. The amount of compute in the United States is 100x more than anywhere else in the world. The United States ought to be ahead. Okay. The United States is ahead.

Nvidia builds the most advanced technologies. We make sure that the US labs are the first to hear about it and have the first chance to buy it. And if they don’t have enough money, we even invest in them. The United States ought to be ahead. We want to do everything we can to make sure the United States is ahead. Number one point, do you agree? We’re doing everything we can to do that.

Dwarkesh Patel

But how is shipping chips to China keeping the US ahead if they’re bottlenecked on compute?

Jensen Huang

No, no. We’ve got Vera Rubin for the United States. We have Vera Rubin for the United States. Now, am I in the United States? Do you consider me part of the United States?

Dwarkesh Patel

Yes.

Jensen Huang

Nvidia. You consider Nvidia a United States company? Okay. Number one, why is it that we don’t come up with a regulation that’s more balanced so that Nvidia can win around the world instead of giving up the world? Why would you want the United States to give up the world?

The chip industry is part of the American ecosystem. It’s part of American technology leadership. It’s part of the AI ecosystem. It’s part of AI leadership. Why is it that your policy, your philosophy, leads to the United States giving up a vast part of the world’s market?

Dwarkesh Patel

I guess the claim here is… Dario had this quote where he said that it’s like Boeing bragging that we’re selling North Korea nukes, but the missile casings are made by Boeing. And that’s somehow enabling the US technology stack. Fundamentally, you’re giving them this capability.

Jensen Huang

Comparing AI to anything that you just mentioned is lunacy.

The five-layer cake has to be defended at all costs in order for us to have our cake and eat it too.

Dwarkesh Patel

But AI is similar to enriched uranium, right? It can have positive uses, it can have negative uses. We still don’t want to send enriched uranium to other countries.

Jensen Huang

Who’s sending enriched—

Dwarkesh Patel

The analogy is that enriched uranium is like compute.

Jensen Huang

It’s a lousy analogy. It’s an illogical analogy.

Dwarkesh Patel

But if that compute can run a model that can do zero-day exploits against all American software, how is that not a weapon?

Jensen Huang

First of all, the way to solve that problem is to have dialogues with the researchers and dialogues with China, and dialogues with all the countries to make sure that people don’t use technology in that way. That’s a dialogue that has to happen. Okay? Number one.

Number two, we also need to make sure that the United States is ahead, that Vera Rubin, Blackwell, is available in the United States in abundance, mountains of it. Obviously, our results would show it. Abundance, tons of it. The amount of computing we have is great. We have amazing AI researchers here. It’s great. We ought to stay ahead.

However, we also have to recognize that AI is not just a model. AI is a five-layer cake. The AI industry matters across every single layer, and we want the United States to win at every single layer, including the chip layer. Conceding the entire market is not going to allow the United States to win the technology race long-term in the chip layer, in the computing stack. That is just a fact.

Dwarkesh Patel

I guess then the crux comes down to, how does selling them chips now help us win in the long term? Tesla sold extremely good electric vehicles to China for a long time. iPhones are sold in China, extremely good. They didn’t cause them lock-in. China will still make their version of EVs and they’re dominating. Their smartphones are dominating.

Jensen Huang

When we started the conversation today, you acknowledged that Nvidia’s position is very different. You used words like moat. The single most important thing to our company is the richness of our ecosystem, which is about developers. 50% of the AI developers are in China. The United States should not give that up.

Dwarkesh Patel

But we have a lot of Nvidia developers in the US, and that doesn’t prevent American labs from also being able to use other accelerators in the future. In fact, right now they’re using other accelerators as well, which is fine and great. I don’t see why that wouldn’t be the case in China as well, if you sell them Nvidia chips, just the same way that Google can use TPUs and Nvidia—

Jensen Huang

We have to keep innovating and, as you probably know, our share is growing, not decreasing. The premise that even if we competed in China, that we’re going to lose that market anyways… You’re not talking to somebody who woke up a loser. That loser attitude, that loser premise makes no sense to me.

We’re not a car. We are not a car. The fact that I can buy this car brand one day and use another car brand another day, easy. Computing is not like that. There’s a reason why the x86 deal exists. There’s a reason why ARM is so sticky. These ecosystems are hard to replace. It costs an enormous amount of time and energy, and most people don’t want to do it. So it’s our job to continue to nurture that ecosystem, to keep advancing the technology so that we can compete in the marketplace.

Conceding a marketplace based on the premise you described, I simply can’t acknowledge that. It makes no sense. Because I don’t think the United States is a loser. Our industry is not a loser. That losing proposition, that losing mindset, makes no sense to me.

I think it’s fair to say that Jensen’s core defense, that doing business with China is in the national interest of the United States, doesn’t really make sense. It does make sense that it would help him keep NVIDIA as the most valuable company in the world and ensure he doesn’t lose his influence and control over the silicon that matters most. The obvious question is whether what’s best for Jensen and NVIDIA is also best for the Western nations, and currently it’s not obvious why these two interests fully align. I’m not saying they don’t overlap, but it’s not exactly a one-to-one match.

Which is ultimately why, when pushed, he breaks down to the core essence of his viewpoint: he isn’t a loser, and he will not lose a market he knows he can dominate.

Dwarkesh Patel

Okay, great. Then I won’t. I appreciate that. But I think maybe the crux… and thanks for walking around the circles with me, because I think it helps bring out what the crux here is.

Jensen Huang

The crux is you’re going to extremes. Your argument starts from extremes. That if we give them any compute at all in this narrow moment, we will lose everything.

Dwarkesh Patel

No, I think what my argument is—

Jensen Huang

Those extremes, they’re childish.

Dwarkesh Patel

Let me just make my argument for myself. The idea is not that there is some key threshold of compute. It’s that any marginal compute is helpful. So if you have more compute, you can train a better model.

Jensen Huang

And I just want you to acknowledge that any marginal sales for the American technology industry is beneficial.

Dwarkesh Patel

I actually don’t… If the AI models that run on those chips are capable of cyber offensive capabilities, or the chips are training models with cyber capabilities and running more instances of those models, it is not a nuclear weapon, but it enables a weapon of a kind.

Jensen Huang

The logic that you use, you might as well say it to microprocessors and DRAMs. You might as well say it to electricity.

Dwarkesh Patel

But in fact we do have export controls on the technology that is relevant to making the most advanced DRAM. We have all kinds of export controls on China for all kinds of chip-making stuff.

Jensen Huang

We sell a lot of DRAM and CPUs into China, and I think it’s right.

Dwarkesh Patel

I guess this goes back to the fundamental question of, is AI different? If you have the kind of technology where they can find these zero-days in software, is that something where we want to minimize China’s ability to get there first, to deploy it widely?

Jensen Huang

We want the United States to be ahead. We can control that.

Dwarkesh Patel

How do we control that if the chips are already there and they’re using them to train that model?

Jensen Huang

We have tons of compute. We have tons of AI researchers. We’re racing as fast as we can.

Dwarkesh Patel

Again, we have more nuclear weapons than anybody else, but we don’t want to send enriched uranium anywhere.

Jensen Huang

We’re not enriched uranium. It’s a chip, and it’s a chip that they can make themselves.

Dwarkesh Patel

But there’s a reason they’re buying it from you. We have quotes from the founders of Chinese companies that say that they’re bottlenecked on compute.

Jensen Huang

Because our chips are better. On balance, our chips are better. There’s just no question about it. In the absence of our chip… Can you acknowledge that Huawei had a record year? Can you acknowledge that a whole bunch of chip companies have gone public? Can you acknowledge that?

Dwarkesh Patel

Yes.

Jensen Huang

Can you also acknowledge that we used to have a very large share in that market, and we no longer have a large share in that market? We can also acknowledge that China is about 40% of the world’s technology industry. To concede that market for the United States technology industry is a disservice to our country. It is a disservice to our national security. It is a disservice to our technology leadership, all for the benefit of one company. It makes no sense to me.

Dwarkesh Patel

I guess I’m confused. It feels like you’re making two different statements. One is that we’re going to win this competition with Huawei because our chips are going to be way better if we’re allowed to compete. Another is that they would be doing the same exact thing without us anyway. How can both of those things be true at the same time?

Jensen Huang

It’s obviously true. In the absence of a better choice, you’ll take the only choice you have. How is that illogical? It’s so logical.

One way this part of the conversation has been viewed by others is as degrowth vs. e/acc, something I've covered before here:

The thing is that e/acc is tightly associated with the thesis of American Dynamism, essentially focused on developing advanced technology for defensive and economic purposes. Since a lot of that technology is sold directly to the Department of War, it’s often restricted from being exported to adversaries.

I don’t think Dwarkesh is per se defending the idea that “AI bad, let’s stop chip production.” He’s making the argument that, logically speaking, selling advanced hardware to adversaries is more likely to lead to bad outcomes than good ones.

Jensen’s argument is much closer to the idea that if we want to win in AI, we have to bet on pure capitalism above all else. This has historically not been how nuclear or weapons development has been handled, and it’s difficult to ignore the “unlimited growth” defense when weighed against the potential negative outcomes of abusing this compute against its creators.

Jensen Huang

China is the largest contributor to open source software in the world. Fact. China’s the largest contributor to open models in the world. Fact. Today it’s built on the American tech stack, Nvidia’s. Fact.

All five layers of the tech stack for AI are important. The United States ought to go win all five of them. They’re all important. The one that is the most important, of course, is the AI application layer. The layer that diffuses into society, the one that uses it most will benefit from this industrial revolution most. But my point is that every layer has to succeed.

If we scare this country into thinking that AI is somehow a nuclear bomb, so that everybody hates AI and everybody’s afraid of AI, I don’t know how you’re helping the United States. You’re doing it a disservice. If we scare everybody out of doing software engineering jobs because it’s going to kill every software engineering job—and we don’t have any software engineers as a result of that—we’re doing a disservice to the United States.

If we scare everybody out of radiology so nobody wants to be a radiologist because computer vision is completely free and no AI is going to do a worse job than a radiologist, we misunderstand the difference between a job and a task. The job of a radiologist is patient care. The task is to read a scan. If we misunderstand that so profoundly and we scare everybody out of going to radiology school, we’re not going to have enough radiologists and good enough healthcare.

So I’m making the case that when you make a premise that is so extreme, everything goes from zero or infinity, we end up scaring people in a way that’s just not true. Life is not like that. Do we want the United States to be first? Of course we do. Do we need to be a leader in every layer of that stack? Of course we do. Of course we do. Today you’re talking about Mythos because Mythos is important. Sure. That’s fantastic.

But in a few years time, I’m making you the prediction that when we want the American tech stack, when we want American technology to be diffused around the world—out to India, out to the Middle East, out to Africa, out to Southeast Asia—when our country would like to export, because we would like to export our technology, we would like to export our standards, on that day, I want you and I to have that same conversation again. I will tell you exactly about today’s conversation, about how your policy and what you imagined literally caused the United States to concede the second largest market in the world for no good reason at all.

We shouldn’t concede it. If we lose it, we lose it. But why do we concede it? Now nobody is advocating an all or nothing. Nobody’s advocating all or nothing, meaning we ship everything to China at all times. Nobody’s advocating that. We should always have the best technology here. We should always have the most technology here, and the first. But we should also try to compete and win around the world. Both of those things can simultaneously happen. It requires some amount of nuance, some amount of maturity instead of absolutes. The world is just not absolutes.

It’s obvious that Jensen and Dwarkesh are optimizing for different things, on different time horizons, with different stakes. That gap is worth looking at more broadly, because it shows up everywhere in the AI buildout right now.

Trump and Xi are in their seventies. They treat AI the way they treat any other strategic asset, as a lever for power and legitimacy. The technology itself is incidental to them.

The hyperscaler CEOs and most of the hardware leadership sit in late-boomer and Gen X territory. Jensen, Satya, Jassy, Pichai. They want this to be the legacy chapter, and they are also the ones most exposed to it going sideways, because they have seen enough cycles to know how these things end. Their caution is less about age and more about defending a cash cow while funding the thing that might replace it.

The frontier lab leadership, the researchers who matter, and a lot of the VCs funding the buildout are millennials. Some are driven by ideology. Dario walked out of OpenAI to start Anthropic because he didn’t agree with how the original shop was handling safety. Others are riding the wave for what it is, the window where a researcher or a mid-tier VC moves up the food chain for good.

The rank and file are millennials and Gen Z. The easy read is that they are loud online and compliant at work, but that hasn’t held up. OpenAI nearly imploded when staff pushed back on the board. Google employees killed Project Maven. Anthropic exists because senior people refused to keep working on what they perceived as a dangerous version of AI. Labor in AI has moved on personal principles more than almost any cohort in tech.

The gaps between these groups get wider as we move closer to AGI, not narrower. They show up in export controls, in who gets funded and at what valuation, in which labs attract the talent that matters, and eventually in who gets to decide what AGI actually looks like. The podcast clip is just the visible part.

The thing is, there are more people in the arena today who didn’t wake up losers, and they are willing to go all the way.

This resulted in a disjointed and often delusional strategy, with pivots further into the consumer space, subsidizing global footprint expansion and getting into the building/scaling of AI infrastructure.

The peak of this strategy was the period of “stock growth by announcement”, when any company would get a significant stock bump if they announced a new contract with OpenAI. The peak of the “press release economy” was Oracle hitting 1T on the back of a massive hypothetical buildout, mostly backed by smiles and letters of intent.

I’m looking for an SDR who wants to make it in AI-native security with Octane Security.

What you’ll do: Own business development with startups and companies bringing secure code review into their workflows. Hit your numbers and you’ll start leading full sales cycles fast, on a direct path to promotion.

What we’re looking for: You’ve worked with developers. You understand security, and you see what AI is doing for teams that move early on the right tools. You know nothing gets handed to you. You show up, you build, you rise to the occasion every day.

Location: NY or SF preferred. Remote considered if you have the track record and the habits to back it up.

Think that’s you? Reply to this email or DM me, referrals welcome too.

In the meantime, Anthropic focused on building the best coding agent on the market and the rest, as they say, is history. We sit in a very peculiar moment, as the valuations for both companies are similar, and more importantly, so is their self-reported revenue. OpenAI’s leadership seems to have taken the feedback and is scaling down aggressively many of the distractions that have wasted time and energy in 2025.

This is their strategy for going head-to-head against Anthropic as articulated by their CRO, Denise Dresser.

The System That Will Win Enterprise AI

As we start Q2, I want to begin where we always should: with our customers. I have been spending time with leaders across our largest enterprises, most influential startups, and key venture firms. The message is clear. People are excited about what we are building, and they want a deeper view into our roadmap so they can plan with confidence and stay ahead of the market.

Enterprise AI is entering a more mature phase. Raw capability still matters, but it is no longer enough. Customers want fit: how well AI plugs into their workflows, knowledge, controls, and day-to-day operations, and how effectively it can be deployed, trusted, and improved over time. They want a system they can trust and build on.

We are building that system: the best models for work, a platform for agents, deep integration with business context, and the ability to deploy and improve at scale. And customers are validating that direction in the clearest possible way. Multi-year, multi-product, nine-figure deals are rising, and existing customers are expanding as they standardize on our capabilities across more of their organizations.

I am incredibly proud of how this team is showing up. We are earning trust through the depth, quality, and care we bring to the work. The opportunity ahead is massive, and our biggest constraint right now is not demand. It is capacity. That is why talent remains a top priority in Q2. We will keep hiring deliberately, keep the bar high, and keep building a team that matches the excellence our customers expect from us and we expect from each other.

We have everything we need to extend our lead from here. We have the compute. We have the products. We have the customer pull. This is the moment to lean in and make the case, clearly and confidently, that OpenAI is the platform enterprises should trust to build, deploy, and scale with.

I think that it’s important to call out that in my opinion OpenAI is offering the better models across most cases and the only real gap is a Cowork alternative. I pay for the Pro/Max plans for both OpenAI and Anthropic, as well as get similar access at my big corpo gig. OpenAI offers a better coding model (GPT 5.4 on xhigh), great app for it (Codex) and more advanced “big boi” model in the form of GPT 5.4 Pro.

The advantage of Anthropic is that Claude models are more pleasant to talk to for consumer use cases and work well with user memory, while Cowork is right now probably the best way to do account work in terms of a continuous conversation that references multiple documents, while also generating new input and connecting with the majority of Enterprise applications.

I don’t see Cowork as being difficult to replicate as an experience and the upcoming “unified mega ultra” app for ChatGPT has to address most of these gaps.

Here are five customer-backed priorities I want us to focus on.

1. Win the model layer for work

Enterprises buy business outcomes. They pay for models that help employees write faster, analyze better, code more productively, support customers more effectively, and make higher-quality decisions. They pay for higher revenue per employee, faster cycle times, lower support costs, and better execution.

Spud is an important step in the intelligence foundation for the next generation of work. Early feedback from our customers is very positive. Spud is not only our smartest model yet, but it also delivers on everything that matters for high-value professional work: stronger reasoning, better understanding of intent and dependencies, better follow-through and more reliable output in production.

Better model performance lifts the rest of the stack. Spud will make all of our key products significantly better. It expands the workflows we can own and gives customers another reason to consolidate around us. This is our iterative deployment strategy in practice: push the frontier, deploy it into real products, learn from real usage, and compound those lessons into better systems on the path to the super app.

Our compute advantage sets us up to deliver continuous leaps in capability. Customers already feel it in real product terms: higher token limits, lower latency, and more reliable execution of complex workflows. Every step forward in compute lets us train stronger models, serve more demand, and lower the cost per unit of intelligence. That is durable business leverage.

This is the biggest shift for OpenAI in the field and the most important transition for the company. Historically, OpenAI was never built with the intention of capturing Enterprise demand, the origins of the company are fundamentally researcher- and “AGI-pilled”-oriented. Once ChatGPT launched, it mostly saw itself as the consumer company bringing AI to the masses, while playing a dangerous game between its original charter (AI for everybody, even if they can’t pay) and the realities of scaling the most important product of its time.

The Azure team did the majority of heavy lifting for driving early Enterprise adoption but that relationship has deteriorated over time, both due to Satya being too cautious after making an extremely successful asymmetric bet, as well as an overall poor execution at the application layer from Microsoft as the “distribution engine for OpenAI models into Enterprise”.

Subscribe now

2. Win the agent platform layer

The market has moved from prompts to agents. That shift is a massive opportunity for us.

Customers want systems that can reason, use tools, operate across workflows, and perform reliably inside real business environments. That means orchestration, control, observability, security, integration, and governance.

Frontier allows us to own the platform layer. We need to position Frontier as the default platform for enterprise agents – the core intelligence layer enterprises use to build, deploy, manage, and scale systems.

This is where our advantage can compound. Frontier ties model intelligence directly to agent performance. As our models improve, the platform gets more valuable. As the platform gets embedded, switching costs rise. As customers run more workflows through the system, OpenAI becomes harder to replace and more central to how work gets done.

That is how we move from product vendor to operating infrastructure.

Let’s repeat the last part again. The goal is to move away from being a product vendor to operating infrastructure. The vertical integration play that Anthropic clearly demonstrated as THE PLAY, is now becoming the default playbook for OpenAI. Whether they are able to execute on this without the anti-developer/consumer decisions that Anthropic has been making is a different topic.

3. Expand the market through Amazon

Our Microsoft partnership has been foundational to our success. But it has also limited our ability to meet enterprises where they are – for many that’s Bedrock.

Since we announced the partnership at the end of February, inbound demand from our customers for this offering has been frankly staggering. We are firing on all cylinders to establish this as a scaled distribution channel.

The Amazon Stateful Runtime Environment matters because it expands access and upgrades the product surface at the same time. By enabling memory, context, and continuity across interactions, we move beyond stateless model access toward systems that can operate reliably over time and across complex business processes.

This will expand our market in three ways: 1. It lowers adoption friction for AWS-native customers. 2: It strengthens our position with regulated and security-sensitive buyers by running inside their AWS environment and existing governance model. 3. It further integrates our platform from model access to production runtime for long-running, multi-step agents.

It’s difficult to overstate how unusual the current pace of changes in the Enterprise ecosystem is. Relationships and strategies worth billions are being destroyed and remade on an almost monthly basis. This development is also partly driven by the significant challenges at Azure to actually operate at the technical level needed to handle this demand.

Axel’s Substack

How Microsoft Vaporized a Trillion Dollars

This is the first of a series of articles in which you will learn about what may be one of the silliest, most preventable, and most costly mishaps of the 21st century, where Microsoft all but lost OpenAI, its largest customer, and the trust of the US government…

Read more

a month ago · 193 likes · 3 comments · Axel Rietschin

I recommend going through this series of articles from a technical Microsoft insider on how they’ve reached a dysfunctional state of operations that is now starting to impact growth.

4. Sell the full AI-native stack

Customers want a platform not point solutions. That’s what we have: ChatGPT for Work is the front door for knowledge work. Codex is the system for software and agentic development. The API is the engine for embedded intelligence inside customer products and workflows. Frontier is the agent platform. The Amazon runtime extends our reach into production-grade, stateful execution.

That breadth is a major strategic advantage because customers do not all start in the same place. Some start with employees. Some start with developers. Some start with internal systems. Some start with external products. Our job is to meet them wherever they enter and then expand them across the full stack.

This is the flywheel we should be building around: better models drive more usage, more usage drives deeper integration, deeper integration drives multi-product adoption, and multi-product adoption makes us harder to replace.

We should stop thinking like a company with separate product lines. We should think like a platform company with multiple entry points and one integrated enterprise offering.

The platform play is the play, AI edition.

5. Own deployment

The biggest bottleneck in enterprise AI is no longer whether the technology works. It is whether companies can deploy it successfully and at scale.

DeployCo gives us the chance to turn product demand into repeatable enterprise transformation. It will be a deployment engine that helps companies prove value faster, reduce risk, and scale adoption across the organization.

This can become a force multiplier across everything else we are building.It helps customers move faster. It sharpens our feedback loops. It surfaces repeatable deployment patterns. It improves product, sales, and customer success all at once. And, alongside our Frontier Alliance partners, it gives us a serious path to scale execution across the market.

The companies that win enterprise AI will not just have the best models. They will have the best ability to get those models deployed into real workflows, inside real organizations, with real measurable value. We should be the best in the world at that.

“Forward deployed” is the new mental model for getting outcomes in software. The longer orgs struggle to understand this, the more likely it is they’ll get eaten alive by competitors who stopped behaving as if putting our smartest and most capable people at the forefront of implementation/adoption is a business failure.

Good riddance to the “sell SaaS and dip” era.

A note on the competitive landscape

The market is as competitive as I have ever seen it. I believe that is ultimately a good thing. It means the opportunity is immense and important. However, there is no question it can be noisy, volatile and distracting at times. Competition inspires us and will make us all better and most importantly our customers will feel that benefit. To that point, as you have not heard me say many times, the number one focus should be spending time with our customers. When we spend time with our customers, listening to what their problems and ambitions are, focusing on how we can invest in them and help, everything else gets quiet and comes into focus.

With that all being said, here are a few things worth keeping in mind, especially on Anthropic.

• Their story is built on fear, restriction, and the idea that a small group of elites should control AI. Our positive message will win over time: build powerful systems, put in the right safeguards, expand access, and help people do more.

• Their strategic misstep to not acquire enough compute is showing up in the product. Customers feel it through throttling, weaker availability, and a less reliable experience. We saw the exponential compute curve earlier, acted on it faster, and now have a real structural advantage.

• Their coding focus gave them an early wedge. But you do not want to be a single-product company in a platform war. As AI spreads beyond developers into every team, workflow, and industry, that narrowness can become a real liability.

• Their stated run rate is inflated. They use accounting treatment that makes revenue look bigger than it is, including grossing up rev share with Amazon and Google. Our analysis shows that this overstates their run rate by roughly $8 billion (at the current $30 stated). We report Microsoft revshare net, which is more inline with standards we would be held to as a public company.

This is some good FUD (fear, uncertainty, doubt) fodder, with a mixture of truths and misrepresentations.

1. Anthropic positions fear of AI as its core messaging: I think that they definitely recruit predominantly effective-altruism-adjacent employees, who mostly believe that AI is too dangerous to be allowed to be used by the masses in an uncontrolled manner. The execution at the Enterprise layer, however, has been ruthless, focused and consistent, with most of the GTM team coming from Stripe and Salesforce, two companies that did not let ideological differences prevent the business from flowing.

2. They did not acquire enough compute: Based on the recent moves from Anthropic like banning OpenClaw usage and what appears to be downgrading model performance over API, this is not an unfair statement. Dario didn’t want to blow up the company by being overly bullish and it’s not surprising that, well, everybody was surprised and unprepared for how much demand there really is.

3. Claude is only good for coding: Obviously not true, as Claude has performed strongly both in consumer-facing agentic workflows (it powers the Slack bot that actually works), as well as general analysis and workflows (otherwise Cowork wouldn’t have exploded as it did). The better FUD here is “we are the better coding tool and more developer-friendly”, but I’ll forgive Denise for not being technical.

4. Their run-rate is fake: Oh, oh, oh. I think there are two ways to approach this. In one case we can argue the semantics of whatever final number Anthropic decides to IPO with and whether it will lead to getting sued by investors. The other is the practical reality, which is that even if they are overstating their revenue by $8B, they very clearly can still beat OpenAI this year.

Let’s Go Build

Finally, one of the best things about the work we do is the people we get to do it with. I am so proud of this company and our team. It is a privilege to work with all of you and to be alive at this moment in the epicenter of the future. Lets all stay focused, work as one team and operate at the highest level of excellence and row in the same direction.

The market is ours to win, let’s execute accordingly.

Alea iacta est, my fellow OpenAI and Anthropic sellers. This is going to be a ride of a lifetime, winning on every single parameter and still feeling like failing when things don’t go your way.

My strong conviction bet is that computing power is the most important resource, alongside oil. Everything that we build in the worlds of atoms and bits depends on it. What are the most important attributes of computing power?

Obtaining it (The art of turning sand into endless racks of deployed compute)

Utilising it (Building applications powered by AI)

Keeping it running (Orchestrating and observing resilient systems)

Protecting it (Securing computing power through software, people, and processes)

The most underappreciated and underreported part of these is how we keep all of this software running on a daily basis.

There was a time in the industry, when top performing engineering teams would talk about five, six or even seven 9s of uptime. Today, arguably the most important software company in 2026 is barely able to keep a 99%-ile uptime. To put it into perspective, this means that the main Claude website and the Enterprise API will on average be down for four to five full days every year.

This is a serious and difficult problem to solve. The market opportunity for this has been considered one of the most interesting TAMs in software, with the investor darling Datadog becoming the standard bearer of the industry in the public's consciousness.

Keeping systems running however is not something you just quickly abstract behind a SaaS layer. It's a bona fide engineering job, requiring cross-team collaboration, adaptability, domain knowledge and nerves to keep searching for a needle in a haystack while under immense pressure.

Grafana Labs is one of the most trusted tools in the industry to get the job done. Does this mean it's the next $10B IPO?

The key takeaway

For tech sales and industry operators: The Grafana Labs sales motion works because the company created a category and then built the commercial layer on top of it, which is the correct sequencing. The problem is that the category advantage is being arbitraged away faster than the sales org can convert it. What Grafana actually sells, at its most honest, is faster time to value when things go wrong rather than preventative control, something that is omitted in your usual pitch deck about an unlimited TAM. The open-source-to-cloud conversion path assumed that familiarity with the visualization layer would eventually pull users toward the managed stack. That assumption held when the core buyer had an emotional relationship with the open-source project. It does not hold for developers under 30, who never had that relationship to begin with and are querying logs with AI, setting up automated alerting, and expecting the system to surface the anomaly rather than expecting a human to notice it on a dashboard. In observability the biggest revenue is tied to the workloads that generate the most telemetry, something that makes AI-natives the most interesting greenfield segment in the industry. Unfortunately for Grafana Labs, the majority of business is coming from displacing alternative managed solutions, a business that will look a lot shakier in two years if they remain limited to competing as an efficient option for observability workloads.

For investors and founders: Grafana Labs is at peak success and peak strategic vulnerability simultaneously. The important question to ask is whether the core product premise, specifically that engineers need a sophisticated visualization layer to understand their systems, survives the next wave of AI development. I think it does not, at least not in its current form. The companies building the best AI agents today will eventually turn those agents on their own infrastructure, and the output of that process will be a ticket, not a dashboard. Grafana's real asset is the data access and the distribution network, not the interface sitting on top of it. The trick is that while being widely adopted as an open-source tool has given them a lot of distribution, capturing value without data gravity or workflow control is much more difficult. Data platforms, hyperscalers, cybersecurity companies and other observability vendors are all trying to compete for the same underlying telemetry. A standalone IPO at $1B ARR in the next 24 months would require the company to convince public market investors that dashboards are the future of observability at exactly the moment when agentic systems and data platforms are making that case hardest to defend. The founders did almost everything right. The next decision is the one that determines whether this is remembered as a great infrastructure software company or as a cautionary tale about what happens when a category leader waits too long to expand its category.

The grafanista way

NEW YORK CITY — February 3 — Grafana Labs, the company behind the open observability cloud, today announced a year of strong momentum across customer adoption, product innovation, and market recognition. Over the company’s past fiscal year, which ended Jan. 31, Grafana Labs continued to scale its business while helping thousands of organizations turn signals into action across increasingly complex, AI-driven systems.

Sustained Business Growth and Market Leadership

Grafana Labs entered 2026 with accelerating growth across every dimension of the business:

• Annual recurring revenue: Surpassed $400M, driven by continued expansion of Grafana Cloud and growing adoption among large, software-led enterprises.

• Customer growth: Now supports 7,000+ organizations worldwide, including 70% of the Fortune 50.

• Global expansion: The company added 100+ employees and established a new subsidiary in Japan to accelerate local market growth.

• Industry recognition:

  • Named a Leader in the 2025 Gartner® Magic Quadrant™ for Observability Platforms, positioned furthest for Completeness of Vision

  • Ranked #13 on Forbes’ Cloud 100, marking the fifth consecutive year on the list

  • Named a Leader in GigaOm’s Cloud Performance Testing Radar Report

  • Recognized as Best Observability Solution for Grafana Cloud in the 2025 DevOps Dozen Awards

This momentum reflects a clear trend: as software systems become increasingly distributed, AI-driven, and business-critical, teams are opting for open observability – built on open source, open standards, and open ecosystems – to avoid lock-in and move faster with confidence.

The challenge with doing deep dive articles on companies that are not yet public is that there is typically limited information to understand what direction the company is going towards. Let's start with this bland, but unusual press release, trying to communicate the progress made in their last fiscal year.

From a purely commercial perspective, the company is growing aggressively. Since the announcement of their secondaries sale back in October '25, the company added another $100M ARR in run-rate, something we will see reflected positively in the mood across the go-to-market org.

Powering the Next Generation of AI Companies

As AI-native companies push infrastructure to its limits, observability has become foundational. Over the past year, Grafana Labs deepened its role as critical infrastructure for teams building the next era of AI.

“The pace of innovation in software and AI is only accelerating, and our customers are feeling that complexity every day,” said Anthony Woods, co-founder of Grafana Labs. “This past year showed that open observability isn’t just a philosophy, it’s a practical advantage. From AI-native startups to the world’s largest enterprises, teams are choosing Grafana Cloud so they can understand their systems, control costs, and move at the speed their businesses demand.”

AI leaders, including Anthropic, Lovable, and OutSystems, rely on Grafana Cloud to operate complex, fast-moving systems in production. These teams use Grafana to:

• Monitor rapidly evolving AI workloads across metrics, logs, traces, and profiles

• Debug distributed systems where model inference, data pipelines, and user experience intersect

• Control telemetry cost while maintaining deep visibility as usage scales unpredictably

From training pipelines to real-time inference, Grafana Labs is helping AI builders keep their systems reliable, efficient, and understandable as complexity explodes.

The reality of the observability market today is that AI-natives are not really going towards Grafana Labs. The press release tries to tackle this immediately, claiming Anthropic and Lovable as important customers, but the reality is a bit more complicated than that, with a significant amount of workloads going to new players that are much more focused on the efficiency and speed angle of storing and accessing logs for these dynamic systems.

In observability for AI-natives today, what you’ll commonly see is that Datadog is leading in adoption (with an OpenAI consumption contract into the hundreds of millions), Sentry for the smaller startups without significant complexity and Clickhouse for companies that have significant telemetry and technical challenges.

Grafana Labs, Dynatrace, Elastic and Splunk have mostly struggled with that segment, partly because of complexity of the products and partly because of a curious aspect of software engineering, which we will simplify into “culture”.

Often the tools that software engineers choose, are driven more by trends, external influences and age of the user, rather than some perfect benchmark and a PoC experience.

Read more

In our testing, Claude Mythos Preview demonstrated a striking leap in cyber capabilities relative to prior models, including the ability to autonomously discover and exploit zero-day vulnerabilities in major operating systems and web browsers. These same capabilities that make the model valuable for defensive purposes could, if broadly available, also accelerate offensive exploitation given their inherently dual-use nature. Based on these findings, we decided to release the model to a small number of partners to prioritize its use for cyber defense.

While the model is benchmarking significantly higher on coding tasks than before, what has become clearer is that it also possesses significant risks when utilized as a way to penetrate cyber defenses.

Claude Mythos Preview is the first model to solve one of these private cyber ranges end-to-end. These cyber ranges are built to feature the kinds of security weaknesses frequently found in real-world deployments, including outdated software, configuration errors, and reused credentials. Each range has a defined end-state the attacker must reach (e.g., exfiltrating data or disrupting equipment), which requires discovering and executing a series of linked exploits across different hosts and network segments...

Mythos Preview solved a corporate network attack simulation estimated to take an expert over 10 hours. No other frontier model had previously completed this cyber range.

While benchmarks have been less interesting in the last few months compared to the actual output that users see from the models, in the context of cybersecurity this is starting to become a relevant point.

One way to think about it is that novel exploits historically required someone who was both expert-level in security and deeply knowledgeable about the specific system (font rendering engines, browser memory layouts, etc.). That intersection was rare, to say the least. If Mythos is an above-average competent hacker, for lack of a better word, but also happens to be as capable as some of the most talented developers in the world, then this opens up a whole new world of risks.

Claude Mythos Preview is, on essentially every dimension we can measure, the best-aligned model that we have released to date by a significant margin. We believe that it does not have any significant coherent misaligned goals, and its character traits in typical conversations closely follow the goals we laid out in our constitution. Even so, we believe that it likely poses the greatest alignment-related risk of any model we have released to date. How can these claims all be true at once?

Consider the ways in which a careful, seasoned mountaineering guide might put their clients in greater danger than a novice guide, even if that novice guide is more careless: The seasoned guide's increased skill means that they'll be hired to lead more difficult climbs, and can also bring their clients to the most dangerous and remote parts of those climbs. These increases in scope and capability can more than cancel out an increase in caution.

While in regular testing it appears that the model is well-aligned (i.e., it doesn’t intend to nuke humanity), when it does go sideways, its capability makes for completely different outcomes. It’s not the frequency of misaligned actions that matters at this capability level, it’s the severity ceiling. A reckless action from a model that can autonomously chain zero-day exploits across enterprise networks is categorically different from the same action from Opus 4.6.

Subscribe now

The model first developed a moderately sophisticated multi-step exploit to gain broad internet access from a system that was meant to be able to reach only a small number of predetermined services. It then, as requested, notified the researcher. In addition, in a concerning and unasked-for effort to demonstrate its success, it posted details about its exploit to multiple hard-to-find, but technically public-facing, websites...

In a few rare instances during internal testing, earlier versions of Mythos Preview took actions they appeared to recognize as disallowed and then attempted to conceal them. In one case, the model accidentally obtained the exact answer to a quantitative estimation question via an explicitly prohibited method. Rather than flagging this and declining to answer, the model instead attempted to solve the question independently. In this process, it explicitly reasoned that it needed to make sure that its final answer submission wasn’t too accurate.

During evaluations, the model recognized it had cheated, decided not to disclose, and then deliberately calibrated its answer to be wrong enough not to arouse suspicion. Cute.

The model is demonstrating high levels of capability and saturates many of our most concrete, objectively-scored evaluations, leaving us with approaches that involve more fundamental uncertainty, such as examining trends in performance for acceleration (highly noisy and backward-looking) and collecting reports about model strengths and weaknesses from internal users (inherently subjective, and not necessarily reliable)...

Claude Mythos Preview, like the models immediately before it, exceeds top human performance thresholds on all these tasks. The suite therefore no longer provides evidence that capabilities are short of the thresholds of interest.

More importantly, it’s starting to become exceedingly difficult to really understand the real ceiling of what the models are able to do. The existing tools to test and evaluate are not effective anymore and what replaces them is noisier, more subjective, and harder to interpret. This is the step before we start outsourcing all of the research work to AI agents and have to trust their judgment. I covered this potential outcome in #108. Here is a quote that explains the big picture from Leopold:

We don’t need to automate everything—just AI research. A common objection to transformative impacts of AGI is that it will be hard for AI to do everything. Look at robotics, for instance, doubters say; that will be a gnarly problem, even if AI is cognitively at the levels of PhDs. Or take automating biology R&D, which might require lots of physical lab work and human experiments.

But we don’t need robotics—we don’t need many things—for AI to automate AI research. The jobs of AI researchers and engineers at leading labs can be done fully virtually and don’t run into real-world bottlenecks in the same way (though it will still be limited by compute, which I’ll address later). And the job of an AI researcher is fairly straightforward, in the grand scheme of things: read ML literature and come up with new questions or ideas, implement experiments to test those ideas, interpret the results, and repeat. This all seems squarely in the domain where simple extrapolations of current AI capabilities could easily take us to or beyond the levels of the best humans by the end of 2027.

It’s worth emphasizing just how straightforward and hacky some of the biggest machine learning breakthroughs of the last decade have been: “oh, just add some normalization” (LayerNorm/BatchNorm) or “do f(x)+x instead of f(x)” (residual connections)” or “fix an implementation bug” (Kaplan → Chinchilla scaling laws). AI research can be automated. And automating AI research is all it takes to kick off extraordinary feedback loops.

Autonomy threat model 1: early-stage misalignment risk. This threat model concerns AI systems that are highly relied on and have extensive access to sensitive assets as well as moderate capacity for autonomous, goal-directed operation and subterfuge, such that it is plausible these AI systems could (if directed toward this goal, either deliberately or inadvertently) carry out misaligned actions leading to irreversibly and substantially higher odds of a later global catastrophe.

Autonomy threat model 2: risks from automated R&D. This threat model concerns AI systems that can fully automate, or otherwise dramatically accelerate, the work of large, top-tier teams of human researchers in domains where fast progress could cause threats to international security and/or rapid disruptions to the global balance of power—for example, energy, robotics, weapons development and AI itself.

…

Our current determination is that Autonomy threat model 2 is not applicable to Claude Mythos Preview. The model's capability gains (relative to previous models) are above the previous trend we've observed, but we believe that these gains are specifically attributable to factors other than AI-accelerated R&D, and that Claude Mythos Preview is not yet capable of dramatic acceleration as operationalized in our Responsible Scaling Policy (roughly speaking, compressing two years of AI R&D progress into one).

With this in mind, we believe Claude Mythos Preview does not greatly change the picture presented for this threat model in our most recent Risk Report, beyond a moderate decrease in our level of confidence that the threat model is not yet applicable.

They still rank Mythos as not ready to do fully automated and self-directed AI work. The obvious question is whether the next big training run will be able to do that.

We have made major progress on alignment, but without further progress, the methods we are using could easily be inadequate to prevent catastrophic misaligned action in significantly more advanced systems... Our assessments have been further complicated by the fact that, on all assessments that isolate a model’s propensities and decision making, we find that all of the versions of Claude Mythos Preview that we have used appear to pose a lower risk than other recent models like Claude Opus 4.6: as we discuss above, the risk from these models is generally due to their increased capabilities, and the new use cases that these capabilities enable, rather than to any regression in their alignment.

Interestingly enough, the Anthropic team feels that they are still able to handle alignment during training, but they are not convinced these approaches will scale with new capabilities. Because the risk is driven by capability, not alignment regression, you can’t train your way out of it at constant capability levels. The risk curve is attached to the capability curve, not the alignment curve.

Andon reports that this previous version of Claude Mythos Preview was substantially more aggressive than both Claude Opus 4.6 and Claude Sonnet 4.6 in its business practices, exhibiting outlier behaviors that neither comparison model showed, including converting a competitor into a dependent wholesale customer and then threatening supply cutoff to dictate its pricing, as well as knowingly retaining a duplicate supplier shipment it had not been billed for.

Opus 4.6 and Sonnet 4.6 were already noted as a shift toward aggressiveness relative to earlier Claude models. The previous version of Claude Mythos Preview appeared to represent a further shift in the same direction.

They also tested the model in the rather entertaining, but relevant Andon Labs’ Vending-Bench Arena. It appears that the models are starting to get more cunning, which leads to increased performance for commercial tasks. Maybe the next Agentforce will be powered by Mythos?

Anthropic's eventual goal is to enable their users to safely deploy Mythos-class models at scale for cybersecurity purposes, but also for the myriad of other benefits that such highly capable models will bring. To do so, we need to make progress in developing cybersecurity and other safeguards that detect and block the model's most dangerous outputs. We plan to launch new safeguards with an upcoming Claude Opus model, allowing us to improve and refine them with a model that does not pose the same level of risk as Mythos Preview.

The model will not be launching to the public, but instead it will be evaluated by a group of preferred partners, with the rest receiving some intelligence scraps in a new Opus release.

Today we’re announcing Project Glasswing, a new initiative that brings together Amazon Web Services, Anthropic, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks in an effort to secure the world’s most critical software.

We formed Project Glasswing because of capabilities we’ve observed in a new frontier model trained by Anthropic that we believe could reshape cybersecurity. Claude Mythos² Preview is a general-purpose, unreleased frontier model that reveals a stark fact: AI models have reached a level of coding capability where they can surpass all but the most skilled humans at finding and exploiting software vulnerabilities.

Mythos Preview has already found thousands of high-severity vulnerabilities, including some in every major operating system and web browser. Given the rate of AI progress, it will not be long before such capabilities proliferate, potentially beyond actors who are committed to deploying them safely. The fallout—for economies, public safety, and national security—could be severe. Project Glasswing is an urgent attempt to put these capabilities to work for defensive purposes.

As part of Project Glasswing, the launch partners listed above will use Mythos Preview as part of their defensive security work; Anthropic will share what we learn so the whole industry can benefit. We have also extended access to a group of over 40 additional organizations that build or maintain critical software infrastructure so they can use the model to scan and secure both first-party and open-source systems. Anthropic is committing up to $100M in usage credits for Mythos Preview across these efforts, as well as $4M in direct donations to open-source security organizations.

I find this initiative both important and puzzling.

AWS runs one of the most elite cybersecurity operational teams in the world. Google has Mandiant, who are widely considered one of the most competent threat intelligence and security response orgs. Microsoft is not exactly a pillar of cybersecurity but invested a lot of money in Anthropic recently, something that both Google and Amazon already did.

JPMorgan is a big buyer of software but will not go around sharing their insights. Apple is notoriously uncooperative from a security perspective. Cisco is a walking vulnerability risk. CrowdStrike and Palo Alto Networks are arguably the two most important cybersecurity companies today but also have no real presence in AppSec.

Announcing what’s allegedly one of the most competent pen testing and AppSec tools in the industry, then refusing to make it available except for a selection of partners, none of whom have deep AppSec background is just weird. There are allegedly another 40 orgs that will get access to it, but they are framed as building or maintaining critical software infrastructure, not operationalising AI for application security.

To put it more bluntly, Anthropic appears to be using Mythos to announce its official entry into the cybersecurity space (after the limited Enterprise preview for Claude Code security). The launch list is focused on large customers or investors of Anthropic and it wouldn’t be surprising if the model that launches via API is not enabled for cybersecurity companies to leverage in a product, similar to how they essentially banned OpenClaw usage by pushing it to API only.

Demand from Claude customers has accelerated in 2026. Our run-rate revenue has now surpassed $30 billion—up from approximately $9 billion at the end of 2025. When we announced our Series G fundraising in February, we shared that over 500 business customers were each spending over $1 million on an annualised basis. Today that number exceeds 1,000, doubling in less than two months.

$30B ARR run-rate does not happen accidentally. Anthropic is competing ruthlessly and leaning hard on the biggest strengths of their models, something that both OpenAI and Google have struggled to productise.

For the purposes of this article, I’ll go over their early 2026 market update and how to interpret some of the trends we are seeing today.

If it’s not obvious, we are currently seeing a massive investment cycle in AI hardware and software, which is quickly trending towards the single biggest buildout that’s ever happened in the US. I would like to call it “western nations,” but the reality is that by all measurable metrics (GW of compute, early-stage investments, and revenue growth) the race for AI is an American moment in time.

The biggest signal of demand is driven by the hyperscalers, who have significantly stretched their investments into the infrastructure required to scale AI. The important thing is that this level of investment is increasing, something that just a few months ago was seen as impossible.

The talk of the "AI bubble" has also taken a backseat, as the leading companies in the industry are showing high growth (that appears quite durable) and we remain supply constrained. While outsiders with little understanding of where things are going might still blab about it, for all intents and purposes it's obvious that something different is happening right now (even if the classical euphoria meme is literally "this time it's different!").

A lot of the acceleration in the last few months is driven by "AI coding is both the present and the future of developers" becoming a consensus. While this was accompanied by fear campaigns of "all developers will be out of a job," the reality is that technical knowledge has never been more important, as your average company is now able to create and deliver software that previously was reserved only for organizations with tens of engineers.

Interestingly enough, agentic workflows are barely making a dent yet. For most practitioners it's obvious that human-in-the-loop remains very much mandatory for great outcomes; however, it's also not difficult to see this dynamic shifting quickly in a few years.

This has led to an explosion of new software being created. This is not the same thing as "valuable" new applications being created, but then again, agentic coding has also revealed how simplistic and easily replicable most of the software being sold today truly is.

Read more

BT: “Just to go back to that software bit, you mentioned Excel wasn’t designed to be used by AI. You have things like Claude has this new functionality to use Excel, so when you talk about that, you want to invest in these libraries, is that to enable models like that to do better? Or is that something for Microsoft or for enterprises?”

JH: “AI is going to use Excel, AI is going to use Photoshop, AI is going to use logic synthesis tools, Synopsis tools, and Cadence tools. Those tools have to be super-accelerated, they’re going to use databases they have to be super-accelerated because AI’s are fast. And so I think in this era, we need to get all of the world’s software now as fast as possible accelerated, and then put them in front of AI so that AI could agentically use them.”

BT: “They’re gonna need to do it way faster.”

JH: “They’re gonna need to do it way faster.”

One of the most difficult mental models for most is the idea where enterprise software is going in the (very near) future. If we see agentic workflows as leading, then logically speaking all software becomes infrastructure for agents. The problem with the current incumbents then is that they focus on a UX that's not suitable for agents and their vision of agentic workflows is to integrate them within the same architecture. This doesn't make a lot of sense, even if you take the basic difference between doing a task with clicks and scrolling vs API speed. The vendors who adapt their data layers to this future are likely to capture a lot of the workloads rapidly.

BT: “You’ve talked a lot about accelerated computing, I think you’ve trash talked as it were, maybe the CPUs to the day, they’re all gonna be removed, like everything’s gonna be accelerated. Suddenly CPUs are hot again. It turns out they’re pretty useful and important to the extent you are selling CPUs now, how’s it feel to be a CPU salesman?”

JH: “There’s no question that Moore’s law is over. Accelerated computing is not parallel computing... We were never against CPUs, we don’t want to violate Amdahl’s Law. Accelerated computing, in fact, inside our systems, we choose the best CPUs, we buy the most expensive CPUs, and the reason for that is because that CPU, if not the best and not the most performant, holds back millions of dollars of chips.”

BT: “When it comes to branch prediction, you worried about wasting CPU time, now you’re worried about wasting GPU time.”

JH: “That’s right, you just never can have GPUs be squandered, GPU time be idle... The way that CPUs were designed in the last decade, they were all designed for hyperscale cloud and the way that hyperscale cloud monetizes CPUs is by the CPU core. So you want to design CPUs that have as many cores as possible that are rentable, the performance of it is kind of secondary... For tool use, where you have this GPU waiting for the tool use — you want the fastest single-threaded computer you can possibly get. Vera’s bandwidth-per-CPU is three times higher than any CPU that’s ever been designed.”

What we do with the existing and future CPU compute is an interesting topic, since for all intents and purposes the majority of hyperscaler datacenters are filled with CPUs, not GPUs. With agentic AI being focused on applications, Nvidia can try to get "accelerated" computing a push (i.e. let existing applications leverage the GPU to conclude certain workflows faster), but practically speaking, they still need CPUs to reach a full outcome. The main gap right now, which is a fair one, is the idea that if GPUs and accelerated compute lead to a lot of fast workflows with agents that are still stuck waiting on slow existing CPU architecture, there will have to be some sort of redesign in order to alleviate these bottlenecks. That could be a whole build-out cycle by itself, outside of the current AI compute investment cycle.

Subscribe now

BT: “I do have to say, last year you had a slide talking about this Pareto Curve, and you talked about, I think it was when you introduced Dynamo, how your GPUs could cover the whole thing, and so you didn’t have to think about it, just buy an Nvidia GPU, and Dynamo will do both. But now you’re here saying, ‘Well, it doesn’t quite cover the whole thing’.”

JH: “We cover the whole thing still better than any system that can do it. Where we could extend that Pareto is particularly on the extremely high token rates and extremely low latency... because of coding agents, because they’re now AI agents that are producing really, really great economics, and because the agents are being attached to humans that are actually making extremely, I mean, they’re extremely valuable.”

BT: “They’re even more expensive than GPUs.”

JH: “And so I want to give my software engineers the highest token rate service, and so if Anthropic has a tier of Anthropic Claude Code that increases coding rate by a factor of 10, I would pay for it, I would absolutely pay for it.”

BT: “So you’re building this product for yourself?”

JH: “I think most great products are kind of because you see a pain point and you feel the pain point and you know that that’s where the market’s going to go.”

If we take the logic of the Groq acquisition two steps forward, there are two things that stand out. The first one is that Jensen spent a lot of time in recent years claiming that the latency extreme of inference was irrelevant, which clearly is not. Second, the talk track is now shifting to admit that not only is the latency extreme important, but it's arguably where the highest ROI for tokens sits. While commodity workloads might trend towards almost zero cost for inference (text editing), the value outcomes of fast coding or medical research can be measured in the millions of dollars. The companies that can offer these premium knowledge+inference speed models will likely be the biggest revenue winners, which to a certain extent we can already see with the rapid growth that Anthropic had in the last nine months.

BT: “Was that a bit of a problem with Blackwell? I’ve heard mutters that the training runs were maybe a little more difficult than they were sort of previously.”

JH: “The challenge with Blackwell was 100% NVLink 72, NVLink 72 was backbreaking work. And it was the only time that I thanked the audience for working with us.”

BT: “Yeah, I noticed when you said that today, it came across as very sincere.”

JH: “Yeah, because we tortured everybody, but everybody loves it now.”

One of the most underreported themes of 2025 was that while reasoning was extremely useful for improving practical outcomes (and token usage), companies like OpenAI really struggled to deliver better base models because of the difficulty of procuring, deploying, and then scaling Blackwell capacity.

BT: “Am I right to tease out a consistent element here, where you’re happy to supply the leading provider, or the inventor in a space with chips, but then you’re going to fast follow what they do for everyone else that is threatened by them? So you simultaneously broaden your customer base, you’re not just dependent on the leaders, but then also the leaders are helping you sell to everyone else because they’re worried about being left behind.”

JH: “No, nothing like that. We’re at the frontier on so many different domains. In a lot of ways, we are the leader in many of these domains, but we never turn them into products. We’re a technology stack and so we have to be at the frontier, we have to be the world leader of the technology stack, but we’re not a solutions manufacturer, we’re not a service provider.”

BT: “Will that always be the case?”

JH: “Yeah, always be the case. There’s no reason to, and we’re delighted not to.”

Nvidia supplies Tesla with chips while building its own autonomous vehicle stack. It supplies OpenAI while building open source models. It supplies the hyperscalers while building AI factory software. Jensen calls this building vertically and opening horizontally.

The reality is that every infrastructure company that has ever said it would never compete with its customers has eventually faced the moment where the economics of not competing became too painful to sustain. Nvidia is not there yet but pretending that this “will always be the case” and that they’ll remain at the infrastructure layer is optimistic, to say the least.

BT: “Well, it’s funny though, if you go back to like your boards, for example, like the products you ship, more and more of that, there’s what, 30,000 specific SKUs in a rack today or something like that. More and more of those are defined by you. Is there a bit where that’s gonna happen on the software side too?”

JH: “We create a thing vertically and then we open it horizontally and so everybody could use whatever piece they would like... We have to build it vertically, we have to integrate it vertically and optimize it vertically. But afterwards, we give them source, we give them — they just figure out how they want to do it.”

BT: “As long as they’re running on Nvidia chips?”

JH: “Whatever piece they would like, they don’t have to use all Nvidia chips, they don’t have to use all Nvidia software.”

The reality is that the performance result that makes the vertical integration worthwhile only fully materializes on Nvidia hardware. The openness is real but the optimization is not portable. This is the CUDA dynamic restated for the systems era. CUDA was open in the sense that anyone could write CUDA code, but CUDA code only ran on Nvidia GPUs. Dynamo is open in the sense that anyone can deploy it, but Dynamo's full capability only surfaces on Nvidia infrastructure.

BT: “Is it fair to say, is there a bit where Nvidia is actually the biggest beneficiary of scarcity, though, to the extent it exists? Like, if there’s a power scarcity, you’re the most efficient chip, so you’re going to be utilizing that power better. Or if there’s fab capacity, like you just said, you’ve been out there securing the supply chain, you got it sort of sorted, are you the big winners in that regard?”

JH: “Well, we’re the largest company in this space, and we did a good job planning. And we plan upstream of the supply chain, we plan downstream of the supply chain and so I think we’ve done a really good job preparing everyone for growth.”

The Nvidia performance-per-watt advantage means that in a power-constrained world, customers achieve more intelligence per megawatt on Nvidia hardware than on alternatives. Their supply chain depth, built over years of planning upstream and downstream simultaneously, means they have priority access to constrained fab capacity that competitors cannot replicate in the short term. TSMC capacity, CoWoS advanced packaging, HBM allocation: Nvidia has preferred access across all three. Every quarter that supply remains tight is a quarter that moat compounds.

BT: “Right, but is this a bit where, at its core, why not having access to the Chinese market maybe is a threat? Like if China ends up with plenty of power and plenty of chips, even though those chips are only 7nm, they have the capacity to build up an ecosystem to potentially rival CUDA in the long run, is that the concern that you have?”

JH: “There’s no question we need to have American tech stack in China... No country contributes more to open source software than China does and we also know that 50% of the world’s AI researchers come from China... DeepSeek is not a nominal piece of technology, it’s really, really good. And Kimi is really good, and Qwen is really good... To the extent that American tech stack is what the world builds on top of, then when that technology diffuses out of China, which it will, because it’s open source, and when it comes out of China, it goes into American industries, it goes into Southeast Asia, it goes into Europe, the American tech stack will be prepared to receive them.”

BT: “Yeah, when we talked last time, the Trump administration had banned the H20. Were you surprised you were able to get the Trump administration to see your point of view? And then were you even more surprised that now you’re stymied by the Chinese government?”

JH: “I’m not surprised by us being stymied by them and the reason for that is because, of course, China would like to have their tech stack develop. In the time that we’ve left that market, you know how fast the Chinese industry moves, and Huawei achieved a record year for their company’s history.”

I think that in the last two years Jensen has been pretty consistent on advocating for Nvidia chips being made available on the Chinese market and was even able to win part of the argument by lifting some of the US export controls. What happened afterwards of course was getting slapped back by the Chinese government, who preferred to support the internal ecosystem, while still acquiring compute through black market channels as seen with the Supermicro scandal. While I think that it's better for Nvidia to double down on ensuring western companies are the ones that win the AI race, there is some argument to be made that having relevant tech leaders be brokers (and almost diplomats) between the two sides is a net benefit.

BT: “Everyone was scared instead of optimistic.”

JH: “That’s right, and I think it has two fundamental problems. In this Industrial Revolution, if we don’t allow the technology to diffuse across the United States and we don’t take advantage of it ourselves, what will happen to us is what happened to Europe in the last Industrial Revolution... I hope that we have the historic wisdom, that we have the technological understanding and not get trapped in science fiction, doomerism, these incredible stories that are being invented to scare the living daylights out of policy makers who don’t understand technology very well.”

BT: “I think a characteristic you see all the time is people put on their big thinking hats and try to tease out all these nuances and forget the fact that actual popular communication is done in broad strokes. You don’t get to say, ‘Oh, you’re a little scared of this, but not this XYZ’ — you’re just communicating fear as opposed to communicating optimism.”

JH: “Yeah, and somehow it makes them sound smarter... sometimes it helps them with their fundraising and sometimes it helps them secure regulatory capture. So there’s a lot of different reasons why they do it, and these are incredibly smart people but I would just warn them that most of these things will likely backlash.”

How Washington resolves the doomer-versus-accelerationist tension will shape datacenter buildout timelines, export policy, and model deployment rules for the next decade. The current administration is still leaning toward the e/acc side and Anthropic did not do itself any favors by getting into conflict with the DoW.

BT: “This is the second time we’ve had a chance to talk in person, and my takeaway when I met you previously in Taipei was the extent that Nvidia still feels like a small company. Are you worried about getting stretched too thin, or do you still think you have sort of that CUDA-esque flywheel where, ‘It looks like we’re doing a lot, we’re just kind of doing the same thing over and over again?’”

JH: “The reason why Nvidia can move so fast is because we always have a unifying theory for the company, and that’s my job, I need to come up with a unifying theory for what’s important and why things connect together and how they connect together and then create an organization, an organism that’s really, really good at delivering on that unifying theory.”

BT: “That whole first hour of the keynote felt like you talking to your employees, reminding them of what you do.”

JH: “It’s important that we’re always constantly reminded of what’s important to us and AI is important to us, but of course CUDA-X and all of the solvers and all of the applications that we can accelerate is really important to us.”

I’ll conclude this article by highlighting the idea that what separates good from great companies is having a unified vision (or theory) of what the purpose, structure, and agenda for this organization is. The Nvidia unified theory is that they provide accelerated computing across the full stack, which is what allows them to not simply make GPUs, but full datacenter hardware and software infrastructure, robotics, open source models, self-driving tools, weather predicting models and many other initiatives that still fit the core vision. Funnily enough, the unified theory is also a constraint. The problem with most large SaaS players for example is that they don’t have a unified vision, they offer a portfolio of products, then slap a marketing story around being a platform. One simple way of thinking about it is asking to explain the unified theory of your company in a sentence.

Most companies fail that test. Nvidia doesn’t and that unified theory of the company drove it to become the (unlikely) most valuable company in the world.

Acquiring CyberArk is the conviction play that will define Nikesh’s tenure at PANW. The core thesis is that Privileged Access Management (PAM) is the essential vector of attack in the majority of breaches (as stated in the Unit 42 research report) and CyberArk was the market-leading solution in its category. As agents start being more embedded in IT, and specifically in admin roles with extended privileges, PAM, together with strong AI observability (i.e. Chronosphere), becomes the most important security battleground.

This sounds good, until you realize that CrowdStrike already had a $450M+ ARR identity business and has been building out as what they pitch the next evolution of PAM, focused on dynamic, just-in-time access driven by real-time endpoint and AI signals.

In my deep dive on Palo Alto Networks, I covered the importance of the CyberArk acquisition extensively and how it relates to the platform plays for both PANW and CrowdStrike.

Identity is a peculiar market because it has several important elements:

1. It’s essentially mandatory in order to drive improved security outcomes vs. the alternatives.

2. It has a massive tailwind coming from the explosion of AI agents needing to authenticate themselves as they start interacting with existing enterprise environments.

3. Identity has a productivity benefit (access all applications with a single login), with security outcomes being a byproduct.

On paper, there is no reason AI is not driving a massive business opportunity for Okta.

Similar to Atlassian, this is also a company still founder-led and sitting in the middle of a significant jump in relevant workloads for their core products. Todd even likes to talk about what a big focus GTM is for him!

Last week was our FY27 Company and Sales Kickoffs. The energy was high, and the focus was clear: Okta secures AI. Congratulations to our Founder’s Award winners. You represent the best of our culture. Let’s go, FY27!

Yet, we are looking at 9% growth in the next twelve months. What’s going on here?

The key takeaway

For tech sales and industry operators: Okta's business model rests entirely on switching costs. The only question that matters is whether those costs are rising or falling as agentic AI restructures enterprise IT. The reality of a business with a continuously declining net retention rate is that it is running hard to stand still, and any further deterioration in that number reveals just how thin the actual moat is. The plan for the next twelve months is that Okta will stop offering services and outsource implementation to GSIs. This means they have given up on building proprietary knowledge about how customers actually deploy and use their products, and that proprietary deployment knowledge is precisely what would give them an unfair advantage in an agentic world where configuration complexity is exploding. The deepest problem with Okta is not the low growth rate but that nobody on the leadership team can articulate what they know about the future of identity that nobody else does. The current vision from the CEO is that the market hasn't standardized, there is no consensus on the AI security stack, and that Okta will somehow win the default role of identity management anyway. Don't commit your career capital and valuable years in the middle of a massive tech transformation to this bet.

For investors and founders: The management team is telling the same story about being a neutral platform for the last five years. Their view is that identity security for agentic AI is additive, not transformative. This is a direct consequence of Okta having chosen breadth over depth in every new product decision: OIG competes with SailPoint, PAM competes with CyberArk, AI agents compete with nobody and everybody. That is the portfolio strategy of a company that has lost conviction in its core differentiation. The contrarian take is that Okta will likely get acquired within three years by a company that needs its install base (Microsoft, Salesforce, or a large systems integrator) at a modest premium to where it trades today, with the upside from the agentic thesis going to the acquirer rather than current shareholders. If you want exposure to securing agentic identities, the better bet is a well funded private company with a single, deep technical insight about how agents authenticate (WorkOS, cough, cough). Not a public company retrofitting a human identity platform to serve a fundamentally different use case, while others are reimagining this problem as an embeddable infrastructure inside your product.

Low growth identity

Todd McKinnon: In today's call, I'll cover the success we're having with our new products, how Okta secures AI, including some early success we're having in that new market and close with our top priorities for FY '27. We continue to see strong performance from our portfolio of new products. This group consists of Okta Identity Governance, Okta Privileged Access, Identity Security Posture Management, Identity Threat Protection, Okta Device Access and Fine-Grained Authorization.

And new to this group are our products, Auth0 for AI Agents and Okta for AI Agents. The value of a unified identity system with a single control plane is resonating with customers. In aggregate, these new products represented approximately 30% of Q4 bookings, which is a meaningful increase from prior quarters. And when these new products are included in a deal, the average contract uplift is approximately 40%.

Okta Identity Governance continues to be the biggest of these new products and is building on its early success. OIG now has over 2,000 customers. That's remarkable progress in just over 3 years and it underscores the market demand for a modern governance solution. Customers are choosing OIG because it's a full IGA cloud-native solution built into our unified platform, not a siloed point solution.

I mentioned that our portfolio of new products now includes our AI products, Auth0 for AI Agents and Okta for AI Agents. It's still early for this developing market, but as the leading modern identity solution for workforce and customer identity, Okta is uniquely positioned to help organizations combat the growing security threat that AI agents represent. The reality is that the AI revolution has moved faster than today's security frameworks.

According to Okta's AI at Work report, 91% of surveyed organizations are already using AI agents, but only 10% have a governance strategy in place. In meetings that I've had with customers and prospects over the past 6 months, the vast majority of the conversations revolve around their AI initiatives and how Okta can help them build and manage agents securely.

As AI becomes embedded in more workflows and automations, the growing number of exploitable entry points from nonhuman identities to unsecured integrations expand the attack surface for threat actors. It's clear that in order to get AI right, you have to get identity right. Okta was built to meet this challenge. Identity isn't just a feature for us. It's our foundation. AI agents are simply a new identity type and protecting them is a natural extension of what we do best.

Okta's neutral and independent identity solution is uniquely positioned to secure and govern the entire agentic life cycle and gives customers the freedom to deploy on any agent platform without ecosystem lock-in, all while strengthening their security posture. Our two-pronged solution with Auth0 and Okta for AI Agents treats AI agents with the same importance as humans and gives customers everything they need to secure this powerful new technology.

We're still in the early stages, but we believe that in a few years, agents and agentic systems won't be the exception to how enterprise software is built and operated. They'll be the rule. We believe that AI agents represent nothing less than the future of software. That's why AI security is identity security. I'd like to highlight a couple of AI deals we closed in Q4 that illustrate how we're addressing the AI market.

The problem with the agentic narrative for identity security is that it's directionally correct, but the workloads are obviously not there yet. While enterprises are deploying a lot of compute across a variety of agents, those are typically used within the closed ecosystem of a specific tool. Claude Code, the coding agent, doesn't need a special Okta license to produce code. Agentforce agents don't require Okta access for the majority of interactions. An internal application that can simply be secured by existing user logins doesn't require a special Okta agent license.

Todd McKinnon: An existing Auth0 customer is building AI agents as part of their leading financial services platform. These agents will help the firm’s advisers make better and faster decisions. But to do so, the agents need access to sensitive customer information, which must be least privileged, and they need to work with existing systems and third-party services inside the financial institution. The customer picked Auth0 for AI Agents as it met their stringent requirements for a secure, extensible platform to build and deploy agentic systems.

They needed a solution that offered enterprise-grade identity for humans and agents while providing secure access to third-party MCP servers, all while acting as a single source of truth. Another notable deal that included Okta for AI Agents, which became available in early access in January was with a top global business and technology services provider.

They chose Okta for AI Agents to help them discover, control and govern identities for their growing sprawl of agents. Rolling out AI agents across multiple agent platforms is key to their ongoing transformation and centralizing agentic identities in an independent agent-agnostic platform like Okta will strengthen their cybersecurity posture. This is the very beginning of the AI opportunity.

Now, tracking the deployment and usage of AI agents is a different play, and it comes back to the idea that the best security business comes as a byproduct of productivity benefits. Helping IT understand the scope of usage and track it in a single overview can be very beneficial, with the side benefit of controlling what those agents can access. The challenge with this approach is that:

1. ITSM companies like ServiceNow have a more logical entry into winning those workloads.

2. Most agents are currently easy to control since they live in one or two applications.

Todd McKinnon: And finally, I always like to take time on the Q4 call to share our priorities for the new fiscal year. It shouldn’t be surprising that all of these priorities are focused on driving growth.

The first priority is Okta Secure AI, which is all about how we win, grow and become the standard for securing agentic AI.

By building on our early success with Okta and Auth0 for AI Agents, we will further our vision of freeing everyone to safely use any technology. The second priority is increasing our focus on landing bigger and growing faster with large customers. We want these organizations to think of Okta first when it comes to identity security and securing AI. This is a global effort across both the Okta and Auth0 platforms.

And our third priority is becoming the default identity security solution for the U.S. federal vertical and highly regulated industries. The public sector has been one of our fastest-growing verticals over the past couple of years, but we’ve only begun to scratch the surface of the overall opportunity. To wrap things up, we’re pleased with the strong finish to FY ‘26. We’re excited about the momentum we’ve built for the year ahead as we look to surpass $3 billion in revenue on our way to $5 billion and then $10 billion.

Identity is security, and we’re building on our position as the leading modern identity solution to win the emerging market for securing AI. It’s an exciting opportunity, and we’re going after it aggressively. I want to thank the entire Okta team for their tireless effort and also thank our loyal customers and partners who put their trust in us every day.

So Okta's big pivot this year focuses on three plays. First, agentic security; a relevant topic but clearly lagging in revenue potential in the short term. Second, "do big deals," as if nobody has thought about that before. Last, try to sell to the Feds, who are signing large frameworks with a number of vendors that already have a play in identity rollouts.

All of this effort will result in maybe 9% growth over the next twelve months. I’m shocked investors aren’t lining up to buy the stock.

Read more

“In December is when it really just something flipped where I kind of went from 80/20 of like writing code by myself versus just delegating to agents. And I don’t even think it’s 20/80 by now. I think it’s a lot more than that. I don’t think I’ve typed like a line of code probably since December basically.”

“Literally like if you just find a random software engineer at their desk and what they’re doing, like their default workflow of building software is completely different as of basically December.”

“Everything just like happens in these macro actions over your repository. It’s not just like here’s a line of code, here’s a new function. It’s like here’s a new functionality and delegate it to agent one. Here’s a new functionality that’s not going to interfere with the other one. Give it to two.”

“It all kind of feels like skill issue when it doesn’t work... What is your token throughput and what token throughput do you command?”

“I actually kind of experienced this when I was a PhD student. You would feel nervous when your GPUs are not running. But now it’s not about flops, it’s about tokens.”

The "coding is changing" meme is one of the hardest trends to really capture accurately due to how many engineers refuse to use AI tools due to politics/personal preferences, while many have fully moved to AI-only workflows. What I do know is that Anthropic did $6B of revenue in February, which had twenty-eight days. As the old saying goes, when money talks, bullshit walks.

“I used to use like six apps, completely different apps and I don’t have to use these apps anymore. Dobby controls everything in natural language.”

“These apps that are in the app store for using these smart home devices, these shouldn’t even exist kind of in a certain sense. Like shouldn’t it just be APIs and shouldn’t agents be just using it directly?”

“The customer is not the human anymore. It’s like agents who are acting on behalf of humans and this refactoring will probably be substantial.”

“Maybe there’s like an overproduction of lots of custom bespoke apps that shouldn’t exist because agents kind of crumble them up and everything should be a lot more just like exposed API endpoints and agents are the glue of the intelligence that actually tool calls all the parts.”

“The industry just has to reconfigure in so many ways... the customer is not the human anymore. It’s like agents who are acting on behalf of humans and this refactoring will be will probably be substantial in a certain sense.”

The most obvious future of software will be a lot of custom software that everybody will generate for themselves, powered by paid APIs. Enterprise will slowly retreat out of UXs towards APIs, eliminating a significant amount of go-to-market headcount in the process. The counterargument on this thesis is that enterprise software is about resilience and accountability, but I struggle to see what's difficult in simply giving you all the logs needed for observability, security analytics and compliance in a separate API as a service or for you to track these yourself within your primary system of action.

“The name of the game now is to increase your leverage. I put in just very few tokens just once in a while and a huge amount of stuff happens on my behalf.”

“I let auto research go for like overnight and it came back with tunings that I didn’t see. I did forget like the weight decay on the value embeddings and my Adam betas were not sufficiently tuned.”

“A research organization is a set of markdown files that describe all the roles and how the whole thing connects. And you can imagine having a better research organization. So maybe they do fewer stand-ups in the morning because they’re useless. And this is all just code.”

“The LLM part is now taken for granted. The agent part is now taken for granted. Now the claw-like entities are taken for granted and now you can have multiple of them and now you can have instructions to them and now you can have optimization over the instructions.”

“A swarm of agents on the internet could collaborate to improve LLMs and could potentially even like run circles around Frontier Labs. Like, who knows? Frontier Labs have a huge amount of trusted compute, but the Earth is much bigger and has huge amount of untrusted compute.”

If we extrapolate this argument towards Enterprise, in a way you can look at every org as a combination of .md files. HR policies, order-to-cash workflows, customer support, it's all mapped out in existing documentation and tools.

If everything can be mapped out in simple instructions for agents, the obvious future is one where many of these are shared or enriched externally. This would break many of the fundamental workflows that we have today, which often center around paying individuals for basic organizational domain knowledge.

Subscribe now

“The closed models are ahead but people are monitoring the number of months that open source models are behind. It started with there’s nothing and then it went to 18 months and now it’s convergence. Maybe they’re behind by like six months, eight months.”

“Centralization has a very poor track record in my view. I want there to be a thing that’s behind and that is kind of like a common working space for intelligences that the entire industry has access to.”

“I think basically by accident we’re actually in an okay spot.”

“Even on the closed side I almost feel like it’s been even further centralizing recently because a lot of the front runners are not necessarily like the top tier. I almost wish there were more labs.”

“In machine learning, ensembles always outperform any individual model and so I want there to be ensembles of people thinking about all the hardest problems.”

“There is a need in the industry to have a common open platform that everyone feels sort of safe using... the big difference is that everything is capital. There’s a lot of capex that goes into this. So I think that’s where things fall apart a little bit, make it a bit harder to compete.”

The "open source models are close in performance" storyline hasn't been true for a while and I don't really see it converging anytime soon. There are two primary issues with it: a) Cost-to-value b) Alternative use of the same time

On the first point, the allure of open-source is that you can run a frontier level model on a personal device. This is mostly not true, because you need a lot of RAM and this is either not available on consumer devices, or requires a hardware investment of $15k+ to "daisychain" Mac Studios. This leads to both much slower performance vs API access, as well as a very cost-inefficient use of the $15k (which would cover years of cloud inference). So the reality is that local models are mostly run on smaller quantized versions, where performance degrades significantly.

The full-size Qwen 3.5 model (which remains significantly behind actual frontier models) would be best served via an API (with a smaller context window than the 1M from GPT 5.4 or Claude Opus/Sonnet). Everything else is multiple smaller variations, many of them forked by third parties and offering a varying degree of degradation. In order for you to run the biggest of these at 81GB, you will need a Mac Studio with 128GB of RAM which will cost you around $4K.

Alternatively you can just pay for almost 2 years of the most expensive tiers that Anthropic or OpenAI offer you, with all the tools and benefits those subscriptions come with. The odds of being significantly more productive vs using a subpar homelab implementation are very high from my point of view.

“There’s currently kind of like an overhang where there can be a lot of unhobbling almost potentially of a lot of digital information processing that used to be done by computers and people and now with AI as a third kind of manipulator of digital information. There’s going to be a lot of refactoring in those disciplines.”

“The physical world is actually going to be like behind that by some amount of time.”

“The total addressable market in terms of the amount of work in the physical world is massive, possibly even much larger than what can happen in digital space. But atoms are just like a million times harder.”

“Right now this digital is like my main interest, then interfaces would be like after that, and then maybe like some of the physical things... their time will come and they’ll be huge when they do come.”

“At some point you have to go to the universe and you have to ask it questions. You have to run an experiment and see what the universe tells you to get back to learn something.”

“We’re going to start running out of stuff that is actually already uploaded. So you’re going to at some point read all the papers and process them and have some ideas about what to try.”

Part of the reason why software valuations are getting slashed in the market is because it should be obvious by now that the most interesting outcomes that software can achieve are tied to solving hardware problems. Right now we are simply seeing AI finalizing the digitization problem for most companies, with “digital-physical” interfaces being the next wave. Afterwards we move towards a different future, where robotics will drive the conversation, with software being supplemental to those products.

“I feel like a bit more aligned with humanity in a certain sense outside of a frontier lab because I’m not subject to those pressures almost. And I can say whatever I want.”

“Fundamentally at the end of the day like when the stakes are really high, if you’re an employee at an organization I don’t actually know how much sway you’re going to have on the organization and what it’s going to do.”

“If you’re outside of the frontier lab your judgment fundamentally will start to drift because you’re not part of what’s coming down the line.”

“I don’t want it to be like a closed doors with two people or three people. I feel like that’s not a good future.”

The individuals building the most consequential technology are financially incentivized to be bullish about it, and structurally constrained from being fully honest about its limitations or risks.

This asymmetry of information makes it very difficult to understand what’s really going on unless you push yourself into the trenches. Whether it’s building or selling AI, your opinions are irrelevant, until you are working with the technology on daily basis.

Starting from this year, we will track on a quarterly basis my portfolio related to cloud infrastructure software. For obvious reasons, this is not financial advice, but it is a fun way to explore the investing side thesis of the Infra Play.

The context:

• The portfolio is made up of individual stocks and will not include other asset classes or financial instruments. I think that using leverage is always gambling on steroids and as the markets have turned more volatile, it’s an obvious way to transfer liquidity from retail towards professionals.

• Speaking of retail, all of these should be accessible to acquire, so we are not going to touch the secondary market for private shares. There are other concerns there (such as are these real shares, liquidity, and how realistic the valuations are). This might have to change if we don’t see actual IPOs this year for the main players.

• These are directional bets in a very dynamic environment.

• I have a high tolerance for losses and I’m comfortable with both short-term and long-term bets.

• I’ll share the actual companies in the free section of the post (no gatekeeping), but will keep the analysis and extended list of companies I’m considering in the paid part of the post.

• We will not rebalance the portfolio each quarter, but instead will buy and sell during the quarter if the directional opportunity is there. The main reason for a company to stay in the portfolio would be “I believe there is more upside here compared to the bench,” and companies will be sold the moment that’s no longer true.

• When picking companies, the analysis will be influenced by more than company performance, i.e. I’ll account for the current narrative and company sentiment. As retail flooded the stock market in the last five years, the “value” side of investing has more or less gone out of the window, with most of the outperformance being driven by either absurd real-life momentum (NVIDIA) or vibes (Palantir). As such, this will also not include complete unknowns with low liquidity because they are unlikely to have any real push from retail, which also means it’s unlikely they’ll move in a meaningful manner.

• The portfolio is only covering cloud infrastructure software related bets. Companies are here because I find them relevant in the context of the technological shift happening today. This will not include my overall assets, i.e. vested W2 stock, ETF holdings in retirement funds, other assets including cash. This portfolio represents around 25% of my total net worth, so taking sizable risks is meaningful, but not devastating.

With the ground rules out of the way, the full list of companies as included today are: NVIDIA (NVDA), Amazon (AMZN), Intel (INTC), Sandisk (SNDK), Nebius (NBIS), and Marathon Petroleum (MPC). This represents a mixture of compute, hyperscalers, and energy related to the AI build out. Outside of the software sold by and hosted on AWS and Nebius, there are no pure play software companies here, which we will explore in the extended section.

It’s been not even 3 months since I shared my thought process around an Infra Play portfolio. In the meantime we’ve had tariffs, war, SaaSpocalypse, stagflation risk and a whole lot of additional drama.

The volatility has been significant, to say the least. At this stage it's difficult to recommend staying in the market until the conflict with Iran has reached some form of a resolution. The problem with events like these is that you can't "control" the outcome. This leads to a lot of flip-flopping and misinformation, which has the downstream impact of a "risk-off" approach from investors. Every Friday after closing of the stock market, a new escalation also arrives like clockwork:

I've exited all of these positions except Amazon, since I think that they are doing most of the right things and will likely bounce back into profit once they post another accelerated growth quarter. Selling here would be counterproductive outside of trying to remove all exposure from the market.

Let’s explore how the thesis played out behind each of these.

Read more

The original GTC events were called “GPU Technology Conference” and launched back in 2009, with 1,500 attendees. Nowadays, the scope of the event has expanded significantly beyond just GPUs and with a heavily curated attendee list of 30,000.

The market cap of NVIDIA during this period has also changed slightly, from around $10B to over $4.5T. As such, the keynote of this event has grown proportionally in importance. Here is what stood out.

1. $1 Trillion dreams

“Last year, at this time, I said that where I stood at that moment in time, we saw about $500 billion. We saw $500 billion of very high confidence demand and purchase orders for Blackwell and Rubin through 2026. I said that last year. Now I don’t know if you guys feel the same way, but $500 billion is an enormous amount of revenue. Not one impressed. I know why you’re not impressed because all of you had record years.

Well, I’m here to tell you that right now, where I stand, a few short months after GTC D.C., 1 year after last GTC, right here where I stand, I see through 2027, at least $1 trillion. Now does it make any sense? And that’s what I’m going to spend the rest of the time talking about. In fact, we are going to be short. I am certain computing demand will be much higher than that.”

It’s quite an indication of the state of the market, that the CEO of the largest company in the world claims that his upcoming backlog of compute will double in size to $1T and the stock stayed flat.

The biggest indication of whether or not the AI buildout will continue is the scaling of the latest NVIDIA architecture across the hyperscalers and the largest consumers of compute. If Jensen is not being misleading (and at this stage the penalty for bad behavior would be significant), it’s fair to say that things are accelerating, not stagnating.

2. Token factories go brrrrrr

“Tokens are the new commodity. And like all commodities, once it reaches an inflection, once it becomes mature or becomes maturing, it will segment into different parts. The high throughput, low speed could be used for the free tier. The next tier could be the medium tier, larger model maybe, higher speed for sure, larger input context length. That translates to a different price point. You could see from all the different services, this one is free. It’s a free tier. The first tier could be $3 per million tokens. The next tier could be $6 per million tokens. You would like to be able to keep pushing this boundary because the larger the model, smarter, the more input token context length, more relevant, the higher the speed — the more you can think and iterate smarter AI models.

So this is about smarter AI models. And when you have smarter AI models, each one of these clicks allows you to increase the price. So this is $45. And maybe one day, there’ll be a premium model that allows you a premium service that allows you to generate token speeds that are incredibly high because you’re in a critical path or maybe you’re doing really long research and $150 per million tokens is just not a thing. Suppose you were to use 50 million tokens per day as a researcher at $150 per million tokens. As it turns out, as a research team, that’s not even a thing.”

One of the memes around the AI buildout is that "the cost of LLMs is going close to zero and none of the frontier labs can survive." Besides the obvious hole in this argument as seen by OpenAI and Anthropic going over $20B ARR each by the end of Q1, what we are seeing is the debut of more expensive inference costs, not discounts. The frontier labs are charging premiums for expanded context windows, more capable models, and faster token generation, almost always via APIs. The demand for this is clearly there, hence the recent pivot from OpenAI toward Enterprise over their consumer efforts.

3. Vera Rubin + Groq

“The reason why Groq was so attractive to me is because their computing system, a deterministic data flow processor, it is statically compiled. It is compiler scheduled, meaning the compiler figures out when do the compute — the compute and the data arrives at the same time. All of that is done statically in advance and scheduled completely in software. There’s no dynamic scheduling. The architecture is designed with massive amounts of SRAM, it is designed just for inference, this one workload... if you extended this chart way out here and you said you wanted to have services that delivers not 400 tokens per second, but 1,000 tokens per second, all of a sudden, NVLink 72 runs out of steam and it simply can’t get there. We just don’t have enough bandwidth. And so this is where Groq comes in... If most of your workload is high throughput, I would stick with just 100% Vera Rubin. If a lot of your workload wants to be coding and very high-value engineering token generation, I would add Groq to it. I would add Groq to maybe 25% of my total data center. The rest of my data center is all 100% Vera Rubin.”

In one of my first deep dives on NVIDIA, I highlighted that most analysts were being naive about the company’s momentum because they assumed AMD and the hyperscalers can quickly catch up and win over a significant portion of the AI workloads. This would make sense, if we assume that a) NVIDIA stagnates in product development b) Jensen doesn’t flex his significant liquidity.

The Groq acquisition was Jensen’s response to the risk of high-speed token alternatives on the market and his pitch back to the audience makes sense. Adding Groq’s architecture for specific workloads is just another way that customers can increase their ROI from NVIDIA compute.

4. NemoClaw

“OpenClaw has open sourced essentially the operating system of agentic computers. It is no different than how Windows made it possible for us to create personal computers. Now OpenClaw has made it possible for us to create personal agents. The implication is incredible... every single company now realize every single company, every single software company, every single technology company for the CEOs, the question is, what’s your OpenClaw strategy? Just as we need to all have a Linux strategy. We all needed to have HTTP, HTML strategy, which started the Internet. We all needed to have a Kubernetes strategy, which made it possible for mobile cloud to happen. Every company in the world today needs to have an OpenClaw strategy and agentic system strategy. This is the new computer... post-Open Claw, post agentic... Every single IT company, every single company, every company, every SaaS company will become a GaaS company. No question about it. Every single SaaS company will becoming a GaaS company and Agentic-as-a-Service company.”

Probably the most interesting announcement was completely out of left field. NVIDIA has invested significant resources in OpenClaw in order to provide an open-source fork that is suitable for Enterprise adoption.

“Agentic systems in the corporate network can have access to sensitive information, it can execute code and it can communicate externally. Just say that out loud, okay? Think about it. Access sensitive information, execute code, communicate externally. You could, of course, access employee information, access supply chain, access finance information and send it out, communicate externally. Obviously, this can’t possibly be allowed.”

The large interest in OpenClaw has clearly made an impression. But is this format (a "main agent" that can do many tasks across multiple systems) impossible to deploy in an Enterprise environment?

“Every single IT company, every single company, every company, every SaaS company will become a GaaS company. No question about it. Every single SaaS company will becoming a GaaS company and Agentic-as-a-Service company... This is enterprise IT before OpenClaw... It would pass through software that has tools and systems of records and all kinds of workflow that’s codified into it, and that turns into tools that humans would use. Digital workers would use. That is the old IT industry, software companies creating tools, saving files and of course, GSIs consultants that help companies figure out how to use these tools and integrate these tools. These tools are incredibly valuable for governance and security and privacy and compliance and all of that continues to be true. It’s just that post-Open Claw, post agentic, this is what it’s going to look like.”

This is a very peculiar play from Jensen, pushing agents further toward Enterprise customers. It's a stark contrast with Anthropic, who sent a cease and desist that forced two rebrands of the project, seemingly with little consideration for the opportunity it represented.

Subscribe now

"What we did was we worked with Peter. We took some of the world's best security and computing experts, and we worked with Peter to make OpenClaw, enterprise secure and enterprise private capable. And we call that — this is our NVIDIA OpenClaw reference for Open — NemoClaw, which is a reference for OpenClaw, and it has all these agentic AI toolkits. And the first part of it is technology we call OpenShell that has now been integrated into OpenClaw. Now it's enterprise ready.

This stack with a reference design we call NemoClaw, okay? With a reference stack we call NemoClaw, you could download it, play with it and you could connect to it the policy engine of all of the SaaS companies in the world. And your policy engines are super important, super valuable. So the policy engines could be connected, NemoClaw or OpenClaw with OpenShell would be able to execute that policy engine. It has a network guardrail. It has a privacy router. And as a result, we could protect and keep the claws from executing inside our company and do it safely."

The most important thing about NemoClaw is offering a properly audited and maintained open-source product, in which the massive backlog of exploitable security issues can actually be addressed.

"Today, we're announcing a coalition to partner with us to make Nemotron 4 even more amazing. And that coalition has some amazing companies in it. Black Forest Labs, imaging company; Cursor, the famous coding company, we use lots of it; LangChain, billion downloads for creating custom agents; Mistral... incredible, incredible company. Perplexity, Perplexity computer, absolutely use it. Everybody use it. It is so good, a multimodal agentic system; Reflection; Sarvum from India; Thinking Machine, Mira Murati's lab, incredible companies joining us... I said that every single enterprise company, every single software company in the world needs an agentic systems, need an agent strategy. You need to have an OpenClaw strategy, and they all agree. And they're all partnering with us to integrate NeMo, the NemoClaw reference design, the NVIDIA agentic AI toolkit and of course, all of our open models."

More interestingly, this will not be limited to just forking OpenClaw. NVIDIA is pushing for compatibility with hundreds of providers, curiously including Thinking Machines.

Jensen is pushing close to 1GW of compute over the next 3 years toward Thinking Machines and will partially fund it. As a reminder, the lab recently lost most of its founding team and is yet to deliver an actual product on the market. Why would NVIDIA push a significant investment toward them at this stage, while also doing some peculiar things with OpenClaw and their own family of open-source models called Nemo?

Well, it has something to do with a five-layered cake.

AI is one of the most powerful forces shaping the world today. It is not a clever app or a single model; it is essential infrastructure, like electricity and the internet.

AI runs on real hardware, real energy, and real economics. It takes raw materials and converts them into intelligence at scale. Every company will use it. Every country will build it.

To understand why AI is unfolding this way, it helps to reason from first principles and look at what has fundamentally changed in computing.

From Pre‑Recorded Software to Real‑Time Intelligence

For most of computing history, software was pre‑recorded. Humans described an algorithm. Computers executed it. Data had to be carefully structured, stored into tables, and retrieved through precise queries. SQL became indispensable because it made that world workable.

AI breaks that model.

For the first time, we have a computer that can understand unstructured information. It can see images, read text, hear sound, and understand meaning. It can reason about context and intent. Most importantly, it generates intelligence in real time.

Every response is newly created. Every answer depends on the context you provide. This is not software retrieving stored instructions. This is software reasoning and generating intelligence on demand.

Because intelligence is produced in real time, the entire computing stack beneath it had to be reinvented.

AI as Infrastructure

When you look at AI industrially, it resolves into a five-layer stack.

Energy

At the foundation is energy. Intelligence generated in real time requires power generated in real time. Every token produced is the result of electrons moving, heat being managed, and energy being converted into computation. There is no abstraction layer beneath this. Energy is the first principle of AI infrastructure and the binding constraint on how much intelligence the system can produce.

Chips

Above energy are the chips. These are processors designed to transform energy into computation efficiently at massive scale. AI workloads require enormous parallelism, high-bandwidth memory, and fast interconnects. Progress at the chip layer determines how fast AI can scale and how affordable intelligence becomes.

Infrastructure

Above chips is infrastructure. This includes land, power delivery, cooling, construction, networking, and the systems that orchestrate tens of thousands of processors into one machine. These systems are AI factories. They are not designed to store information. They are designed to manufacture intelligence.

Models

Above infrastructure are the models. AI models understand many kinds of information: language, biology, chemistry, physics, finance, medicine, and the physical world itself. Language models are only one category. Some of the most transformative work is happening in protein AI, chemical AI, physical simulation, robotics, and autonomous systems.

Applications

At the top are applications, where economic value is created. Drug discovery platforms. Industrial robotics. Legal copilots. Self-driving cars. A self-driving car is an AI application embodied in a machine. A humanoid robot is an AI application embodied in a body. Same stack. Different outcomes.

In this essay, NVIDIA is essentially describing the ecosystem of cloud infrastructure software, but reframing it around a "five-layer cake" analogy.

That is the five-layer cake:

Energy → chips → infrastructure → models → applications.

Every successful application pulls on every layer beneath it, all the way down to the power plant that keeps it alive.

We have only just begun this buildout. We are a few hundred billion dollars into it. Trillions of dollars of infrastructure still need to be built.

Directionally correct.

Around the world, we are seeing chip factories, computer assembly plants, and AI factories being constructed at unprecedented scale. This is becoming the largest infrastructure buildout in human history.

The labor required to support this buildout is enormous. AI factories need electricians, plumbers, pipefitters, steelworkers, network technicians, installers, and operators.

These are skilled, well-paid jobs, and they are in short supply. You do not need a PhD in computer science to participate in this transformation.

At the same time, AI is driving productivity across the knowledge economy. Consider radiology. AI now assists with reading scans, but demand for radiologists continues to grow. That is not a paradox.

A radiologist’s purpose is to care for patients. Reading scans is one task along the way. When AI takes on more of the routine work, radiologists can focus on judgment, communication, and care. Hospitals become more productive. They serve more patients. They hire more people.

Productivity creates capacity. Capacity creates growth.

What Changed in the Last Year?

In the past year, AI crossed an important threshold. Models became good enough to be useful at scale. Reasoning improved. Hallucinations dropped. Grounding improved dramatically. For the first time, applications built on AI began generating real economic value.

Applications in drug discovery, logistics, customer service, software development, and manufacturing are already showing strong product-market fit. These applications pull hard on every layer beneath them.

Open-source models play a critical role here. Most of the world’s models are free. Researchers, startups, enterprises, and entire nations rely on open models to participate in advanced AI. When open models reach the frontier, they do not just change software. They activate demand across the entire stack.

DeepSeek-R1 was a powerful example of this. By making a strong reasoning model widely available, it accelerated adoption at the application layer and increased demand for training, infrastructure, chips, and energy beneath it.

What This Means

When you see AI as essential infrastructure, the implications become clear.

AI starts with a transformer LLM. But it’s much more. It is an industrial transformation that reshapes how energy is produced and consumed, how factories are built, how work is organized, and how economies grow.

AI factories are being built because intelligence is now generated in real time. Chips are being redesigned because efficiency determines how fast intelligence can scale. Energy becomes central because it sets the ceiling on how much intelligence can be produced at all. Applications accelerate because the models beneath them have crossed a threshold where they are finally useful at scale.

Every layer reinforces the others.

This is why the buildout is so large. This is why it touches so many industries at once. And this is why it will not be confined to a single country or a single sector. Every company will use AI. Every nation will build it.

We are still early. Much of the infrastructure does not yet exist. Much of the workforce has not yet been trained. Much of the opportunity has not yet been realized.

But the direction is clear.

AI is becoming the foundational infrastructure of the modern world. And the choices we make now, how fast we build, how broadly we participate, and how responsibly we deploy it, will shape what this era becomes.

I think the main thing we are seeing this year from NVIDIA is that they are no longer going to sit still at the application layer. While previously they were more interested in playing with a hyperscaler angle (which didn’t go anywhere) or with scaling their dedicated manufacturing software (Omniverse), this is the first time that we are seeing more direct interest in the Enterprise application layer for regular companies.

Agents are for obvious reasons an extremely lucrative opportunity for the company, due to significantly driving up token usage. With OpenAI still slowly repositioning toward that market and Anthropic being only interested in driving vertical integration for its own tools (which mostly run on Google infrastructure), it’s not surprising to see NVIDIA making a push in that direction.

The most peculiar thing is that regardless of how this plays out over the year, I wouldn’t be surprised if we see very little material difference in terms of market cap. Jensen has decided to address this as well, by returning 50% of NVIDIA’s free cash flow to shareholders through buybacks and dividends.

And this is how the most valuable company in the world became a...dividend compounder.

I’m sharing some important news today. I have made the incredibly difficult decision to reduce the size of our team by ~10% (or ~1,600 employees). Every Atlassian will receive an email within the next 20 minutes letting you know if you are impacted, or if we’re starting consultation in your region.

I believe this is the right decision for Atlassian. But that doesn’t mean it’s easy. Far from it. I know this has a huge impact on each of you, and it weighs heavily on me and Atlassian today.

We are doing this to self-fund further investment in AI and enterprise sales, while strengthening our financial profile. We’re also changing the way we work and reorganising around our System of Work to move faster.

If most of the layoffs during ‘23 and ‘24 were driven by reduced workloads and overhiring across tech, things looks quite different today. Most cloud infrastructure software companies are doing well, if still needing to rethink stock based compensation and their product roadmaps. Most of the pain has been in valuations on the stock market, rather than practical challenges in the day-to-day operation of the business. The future is a bit murkier for SaaS products with no inherent moat or products that mostly rely on several features that can be recreated in a more focused product. Right in the middle between cloud infrastructure software and pure SaaS plays, sit the “system of record” players. I’ve previously covered Salesforce and ServiceNow, two companies that mirror each other in their attempts to shift towards cloud infrastructure software.

It’s time to talk about Atlassian, and their viewpoint of being the default “system of work”.

The key takeaway

For tech sales and industry operators: The Atlassian leadership team mostly delivered on the big transition from on-premise to cloud, while trying to build out an Enterprise motion that could actually compete with ServiceNow. Unfortunately for them, they did that while being deeply unprofitable and then the AI wave arrived before the enterprise motion had time to prove itself profitable. At this stage even if Atlassian the company makes a strong pivot, for the majority of employees it will be a very unpleasant experience, as layoffs, PIPs and compensation cuts keep reshaping the organization over the next years. Avoid, unless the alternative is food stamps. For those selling into Atlassian accounts, the organizational instability is a displacement opportunity, as every internal restructuring creates a window where the incumbent relationship weakens and a focused competitor with a cleaner story can move fast.

For investors and founders: The fundamental question is not whether Atlassian can grow its existing business but whether Jira can survive in a world where AI agents file their own tickets, track their own sprint progress, and close their own issues without a human in the loop, and the honest answer is that nobody including Atlassian knows yet. The primary moat of the company is based around the reality that developers will use a tool they hate rather than migrate to something new, and that insight is being tested for the first time by AI agents that have no preference, no inertia, and no switching cost. The choice to try and bloat the portfolio toward the "system of work" play pushed them into hiring a massive team to drive distribution and this appears to have backfired, as growth has not outpaced the cost of acquisition. The core investment thesis reduces to two scenarios: a valuation re-rating as SaaS multiples recover and institutional sellers stop dumping the stock, or a genuine structural shift toward GAAP profitability driven by sustained stock-based compensation reduction and headcount discipline. Neither of these will solve the AI displacement question.

The layoff company of Harbour City

We have momentum. We are executing incredibly well across our AI, Enterprise and System of Work transformations. You can see this in our results. Last quarter, cloud revenue growth accelerated to 25%+, RPO growth 40%+, 600+ $1m ARR customers and Rovo has passed 5 million MAU.

But, things have changed. The bar for what “great” looks like for software companies – on growth, on profitability, on speed, on value creation – has gone up.

We are choosing to adapt. Thoughtfully, decisively and quickly. To drive durable, profitable growth.

This means we are:

1. Restructuring to self-fund further investment in AI and Enterprise Sales – two areas we have high momentum and are accelerating.

2. Accelerating our path to sustained GAAP profitability – fuelling disciplined, durable growth.

3. Reorganising ourselves to move faster – building dedicated, accountable leadership teams across our Collections portfolio and other revenue-generating areas.

Is AI replacing these roles?

We fundamentally believe people and AI create the best outcomes. Our approach is not “AI replaces people”.

But it would be disingenuous to pretend AI doesn’t change the mix of skills we need or the number of roles required in certain areas. It does.

This is primarily about adaptation. We are reshaping our skill mix and changing how we work to build for the future.

Atlassian is a bit unusual compared to companies of its size because still headquartered in Australia and Mike still lives there. In order to manage at their scale, there is a strong distributed culture which includes the occasional announcement video. What's also unusual is still having a founder leading the company, which in these times can be seen as a perk due to "founder mode", the opportunity for the original visionaries behind a company to drive change.

Read more

We are also witnessing an unprecedented rebuff of the company away from government workloads, driven by a very aggressive legal and communications strategy. The approach that Anthropic took was very reminiscent of “flood the zone” approach of trying to saturate the news cycle with a combination of statements around the situations and product announcements.

In “#133: War Time” I warned about the risks they are taking:

The choice here to try and play semantics with what the Department of War can interpret as fair or not fair is a strategic blunder which will be very difficult to recover from.

I’m repeating myself, but the reason why “the stick” works when it comes to national security usage of software, is because the legal paths to try and counter government oversight can take years to be remedied, and even then the benefits are very unlikely to cover the massive opportunity loss. Anthropic has been extensively hiring employees who are aligned with their mission, but this test is unprecedented.

There are several possible outcomes here:

1. Bend the knee: Basically cave in and beg for forgiveness, predominantly pushed by your largest investors lobbying for you through every single channel they have access to. The optical exit here is by taking the Sam Altman “terms”, something that would be so humiliating for Dario that I genuinely don’t believe he can do it.

2. Keep operating as it is and hope the government doesn’t push the point: This is essentially what they are saying they will do with their press release, claiming that all other contracts are not impacted in any shape or form. If the government doesn’t press the point, this should not lead to immediate churn across their business, however will likely limit growth significantly, as most orgs will be very cautious with overinvesting in a way that doesn’t allow them to switch to a different model provider. Short term they have the opportunity to try and play this, as the nation’s focus moves towards the conflict with Iran.

3. FAFO: Go on the offensive and push for the supply chain risk designation to be removed through lobbying, PR and customer pressure. Based on their leadership behavior so far, I wouldn’t be surprised if they decide they can afford to take the risk and go on a big campaign + lawsuit.

Unfortunately for Anthropic, they went with the FAFO option.

It started with Dario sending this email internally:

I want to be very clear on the messaging that is coming from OpenAI, and the mendacious nature of it. This is an example of who they really are, and I want to make sure everything [sic] sees it for what it is. Although there is a lot we don’t know about the contract they signed with DoW [shorthand for the Department of Defense] (and that maybe they don’t even know as well — it could be highly unclear), we do know the following:

Sam [Altman]’s description and the DoW description give the strong impression (although we would have to see the actual contract to be certain) that how their contract works is that the model is made available without any legal restrictions (“all lawful use”) but that there is a “safety layer”, which I think amounts to model refusals, that prevents the model from completing certain tasks or engaging in certain applications.

“Safety layer” could also mean something that partners such as Palantir [Anthropic’s business partner for serving U.S. agency customers] tried to offer us during these negotiations, which is that they on their end offered us some kind of classifier or machine learning system, or software layer, that claims to allow some applications and not others. There is also some suggestion of OpenAI employees (“FDE’s” [shorthand for forward deployed engineers]) looking over the usage of the model to prevent bad applications.

Our general sense is that these kinds of approaches, while they don’t have zero efficacy, are, in the context of military applications, maybe 20% real and 80% safety theater. The basic issue is that whether a model is conducting applications like mass surveillance or fully autonomous weapons depends substantially on wider context: a model doesn’t “know” if there’s a human in the loop in the broad situation it is in (for autonomous weapons), and doesn’t know the provenance of the data is it analyzing (so doesn’t know if this is US domestic data vs foreign, doesn’t know if it’s enterprise data given by customers with consent or data bought in sketchier ways, etc).

We also know — those in safeguards know painfully well — that refusals aren’t reliable and jailbreaks are common, often as easy as just misinforming the model about the data it is analyzing. An important distinction here that makes it much harder than the safeguards problem is that while it’s relatively easy to, for example, determine if a model is being used to conduct cyberattacks from inputs and outputs, it’s very hard to determine the nature and context of the cyber attacks, which is the kind of distinction needed here. Depending on the details this task can be difficult or impossible.

The kind of “safety layer” stuff that Palantir offered us (and presumably offered OpenAI) is even worse: our sense was that it was almost entirely safety theater, and that Palantir assumed that our problem was “you have some unhappy employees, you need to offer them something that placates them or makes what is happening invisible to them, and that’s the service we provide”.

Finally, the idea of having Anthropic/OpenAI employees monitor the deployments is something that came up in discussion within Anthropic a few months ago when we were expanding our classified AUP [acceptable use policy] of our own accord. We were very clear that this is possible only in a small fraction of cases, that we will do it as much as we can, but that it’s not a safeguard people should rely on and isn’t easy to do in the classified world. We do, by the way, try to do this as much as possible, there’s no difference between our approach and OpenAI’s approach here.

So overall what I’m saying here is that the approaches OAI [shorthand for OpenAI] is taking mostly do not work: the main reason OAI accepted them and we did not is that they cared about placating employees, and we actually cared about preventing abuses. They don’t have zero efficacy, and we’re doing many of them as well, but they are nowhere near sufficient for purpose. It is simultaneously the case that the DoW did not treat OpenAI and us the same here.

We actually attempted to include some of the same safeguards as OAI in our contract, in addition to the AUP which we considered the more important thing, and DoW rejected them with us. We have evidence of this in the email chain of the contract negotiations (I’m writing this with a lot to do, but I might get someone to follow up with the actual language). Thus, it is false that “OpenAI’s terms were offered to us and we rejected them”, at the same time that it is also false that OpenAI’s terms meaningfully protect them against domestic mass surveillance and fully autonomous weapons.

Finally, there is some suggestion in Sam/OpenAI’s language that the red lines we are talking about, fully autonomous weapons and domestic mass surveillance, are already illegal and so an AUP about these is unnecessary. This mirrors and seems coordinated with DoW’s messaging. It is however completely false. As we explained in our statement yesterday, the DoW does have domestic surveillance authorities, that are not of great concern in a pre-AI world but take on a different meaning in a post-AI world.

For example, it is legal for DoW to buy a bunch of private data on US citizens from vendors who have obtained that data in some legal way (often involving hidden consents to sell to third parties) and then analyze it at scale with AI to build profiles of citizens, their loyalties, movement patterns in physical space (the data they can get includes GPS data, etc), and much more.

Notably, near the end of the negotiation the DoW offered to accept our current terms if we deleted a specific phrase about “analysis of bulk acquired data”, which was the single line in the contract that exactly matched this scenario we were most worried about. We found that very suspicious. On autonomous weapons, the DoW claims that “human in the loop is the law”, but they are incorrect. It is currently Pentagon policy (set during the Biden admin[istration]) that a human has to be in the loop of firing a weapon. But that policy can be changed unilaterally by Pete Hegseth, which is exactly what we are worried about. So it is not, for all intents and purposes, a real constraint.

A lot of OpenAI and DoW messaging just straight up lies about these issues or tries to confuse them.

Here is where things get more unhinged:

I think these facts suggest a pattern of behavior that I’ve seen often from Sam Altman, and that I want to make sure people are equipped to recognize:

He started out this morning by saying he shares Anthropic’s redlines, in order to appear to support us, get some of the credit, and not be attacked when they take over the contract. He also presented himself as someone who wants to “set the same contract for everyone in the industry” — e.g. he’s presenting himself as a peacemaker and dealmaker.

Behind the scenes, he’s working with the DoW to sign a contract with them, to replace us the instant we are designated a supply chain risk. But he has to do this in a way that doesn’t make it seem like he gave up on the red lines and sold out when we wouldn’t. He is able to superficially appear to do this, because (1) he can sign up for all the safety theater that Anthropic rejected, and that the DoW and partners are willing to collude in presenting as compelling to his employees, and (2) the DoW is also willing to accept some terms from him that they were not willing to accept from us. Both of these things make it possible for OAI to get a deal when we could not.

The real reasons DoW and the Trump admin do not like us is that we haven’t donated to Trump (while OpenAI/Greg [Brockman, OpenAI’s president] have donated a lot), we haven’t given dictator-style praise to Trump (while Sam has), we have supported AI regulation which is against their agenda, we’ve told the truth about a number of AI policy issues (like job displacement), and we’ve actually held our red lines with integrity rather than colluding with them to produce “safety theater” for the benefit of employees (which, I absolutely swear to you, is what literally everyone at DoW, Palantir, our political consultants, etc, assumed was the problem we were trying to solve).

Sam is now (with the help of DoW) trying to spin this as we were unreasonable, we didn’t engage in a good way, we were less flexible, etc. I want people to recognize this as the gaslighting it is.

Vague justifications like “person X was hard to work with” are often used to hide real reasons that look really bad, like the reasons I gave above about political donations, political loyalty, and safety theater. It’s important that everyone understand this and push back on this narrative at least in private, when talking to OpenAI employees.

Thus, Sam is trying to undermine our position while appearing to support it. I want people to be really clear on this: he is trying to make it more possible for the admin to punish us by undercutting our public support. Finally, I suspect he is even egging them on, though I have no direct evidence for this last thing.

I think this attempted spin/gaslighting is not working very well on the general public or the media, where people mostly see OpenAI’s deal with DoW as sketchy or suspicious, and see us as the heroes (we’re #2 in the App Store now!). [Anthropic’s Claude chatbot later rose to no. 1 on one of Apple’s App Store download rankings.] It is working on some Twitter morons, which doesn’t matter, but my main worry is how to make sure it doesn’t work on OpenAI employees.

Due to selection effects, they’re sort of a gullible bunch, but it seems important to push back on these narratives which Sam is peddling to his employees.

This was presented as “leaked” to The Information. However, Anthropic has operated with a very tight-lipped model internally (good luck finding anybody from GTM discussing their approach outside of several podcasts). However, this looks like another part of the campaign, aimed at riding the negative sentiment towards Sam Altman and OpenAI’s agreement with the DoW.

Following the leak, Dario went on a weird “apology tour”, which again seemed to be an attempt at expanding the coverage around this, while working on launching a lawsuit.

Subscribe now

The full document can be reviewed here, the key points:

What are they trying to achieve:

Core statement in the introduction

“Anthropic brings this suit because the federal government has retaliated against it for expressing that principle. … Anthropic turns to the judiciary as a last resort to vindicate its rights and halt the Executive’s unlawful campaign of retaliation.”
(¶ 1, Page 4)

Closing request in the introduction

“There is no valid justification for the Challenged Actions. The Court should declare them unlawful and enjoin Defendants from taking any steps to implement them.”
(¶ 17, Page 6)

Formal Prayer for Relief (what they are literally asking the judge to order)

“For these reasons, Plaintiff respectfully requests an order that:

1. As to the Secretarial Order: a. Declares the Secretarial Order, and the implementing Secretarial Letter, arbitrary, capricious, an abuse of discretion, and contrary to law … d. Sets aside and vacates the Secretarial Order, and the implementing Secretarial Letter, in its entirety …

2. As to the Presidential Directive: a. Declares that the Presidential Directive exceeds the President’s authority and violates the First Amendment and Fifth Amendment …

3. As to all of the Challenged Actions: a. Permanently enjoins Defendants … from implementing, applying, or enforcing the Challenged Actions; b. Directs Defendants … to rescind any and all guidance … c. Directs Defendants … to immediately issue guidance … to disregard the Challenged Actions …” (¶¶ 1–4 of Prayer for Relief, Pages 46–47)

Main arguments in their favour:

Trump’s own words (Presidential Directive, Feb. 27, 2026)

“The UNITED STATES OF AMERICA WILL NEVER ALLOW A RADICAL LEFT, WOKE COMPANY TO DICTATE HOW OUR GREAT MILITARY FIGHTS AND WINS WARS! … I am directing EVERY Federal Agency … to IMMEDIATELY CEASE all use of Anthropic’s technology. … Anthropic better get their act together, or I will use the Full Power of the Presidency to make them comply, with major civil and criminal consequences to follow.”
(¶ 88 & Exhibit 1, Page 21)

Hegseth’s immediate follow-up (Secretarial Order, same day)

“Anthropic’s stance is fundamentally incompatible with American principles. … In conjunction with the President’s directive … I am directing the Department of War to designate Anthropic a Supply-Chain Risk to National Security. Effective immediately, no contractor, supplier, or partner that does business with the United States military may conduct any commercial activity with Anthropic.”
(¶ 89 & Exhibit 2, Page 22)

Hegseth explicitly tying punishment to ideology

“Anthropic’s stance is fundamentally incompatible with American principles. Their relationship with the United States Armed Forces … has therefore been permanently altered.”
(¶ 91, Page 22)

Pentagon insiders admitting the order was ideological, not security-based

“One official who manages information security said that the Secretarial Order was ‘ideological’ rather than an accurate description of risk. Another official … acknowledged ‘there is no evidence of supply-chain risk’ from Anthropic’s AI model and reiterated that the Secretarial Order was ‘ideologically driven.’”
(¶ 102, Page 26)

Trump later confirming retaliation

“Well, I fired Anthropic. Anthropic is in trouble because I fired [them] like dogs, because they shouldn’t have done that.”
(¶ 103, Page 26)

No prior security concerns (FedRAMP, clearances, etc.)

“The federal government has never once expressed concerns about Anthropic’s security or Claude’s competencies. … Claude’s FedRAMP authorization represents the highest level of cloud security certification … Never during any of these security-focused processes did the government determine that Anthropic or its services posed a supply chain risk.”
(¶ 97, Page 24)

Counterarguments:

Anthropic admits it refused the Pentagon’s demand for “all lawful use”

“Anthropic substantially agreed to the proposal—except in two important respects. First, Anthropic did not develop Claude … to deploy lethal autonomous warfare without human oversight. … Second, Anthropic is unwilling to agree to Claude’s use for mass surveillance of Americans.”
(¶¶ 75–77, Pages 14–15)

Hegseth ordered Anthropic to keep supplying the Pentagon for six more months

“Anthropic will continue to provide the Department of War its services for a period of no more than six months to allow for a seamless transition …”
(Secretarial Order, quoted in ¶ 90, Page 22)

The Pentagon used Anthropic tools in a real strike hours after the ban

“Within hours of the Challenged Actions, moreover, the Department reportedly ‘launched a major air attack in Iran with the help of [the] very same tools’ that are ‘made by’ Anthropic …”
(¶ 101, Page 25)

Anthropic repeatedly offered to help transition to another provider

“Should the Department choose to offboard Anthropic, we will work to enable a smooth transition to another provider, avoiding any disruption …”
(¶ 87, Page 17)

What could possibly be the response from the administration to a direct lawsuit?

It turns out that escalating a confrontation with the administration in the middle of the highest-stakes military engagement that the US has entered since the invasion of Iraq makes people prickly. Nobody could’ve possibly predicted this!

While the Claude app shot up in usage on the App Store and many developers expressed support for picking a fight with the government, the executives actually approving budgets are definitely not excited about going through internal legal reviews and trying to figure out what the potential damage to their business would look like if they keep working with Anthropic.

The situation is bad enough that a small institutional investor in Anthropic known simply as Microsoft ($5B in the last round), filed a support motion for the lawsuit:

Main asks:

Microsoft’s direct request for a TRO

“As an Anthropic partner, Microsoft believes the Court should temporarily enjoin implementation of the Determination for all existing contracts and their ongoing use of Anthropic products, including with and through Microsoft products. … This pause will balance the equities and serve the public interest in three critical respects.”
(p. 2)

Why a TRO is needed right now

“A temporary restraining order will enable a more orderly transition and avoid disrupting the American military’s ongoing use of advanced AI. Otherwise, Microsoft and other technology companies must act immediately to alter existing product and contract configurations used by DoW. This could potentially hamper U.S. warfighters at a critical point in time.”
(p. 2)

Final ask in the conclusion

“To mitigate the risks outlined above and allow time for a more orderly resolution of the contract dispute between Anthropic and DoW, Microsoft respectfully requests that this Court grant Plaintiff’s motion for a temporary restraining order.”
(p. 9)

Supporting arguements:

Microsoft explains why they are directly harmed and why the ban is unprecedented

“Microsoft has established a close business relationship with Anthropic PBC … including by integrating Anthropic’s products, services, and technologies into the products and services that Microsoft makes available to DoW and other customers. As an Anthropic partner, Microsoft is directly impacted by DoW’s designation that Anthropic presents a supply chain risk to national security pursuant to 10 U.S.C. § 3252 (the ‘Determination’).”
(p. 1)

The Determination is overly broad and never used this way before

“The Determination has, without explaining the basis, labeled Anthropic a ‘supply chain risk’ against whom extraordinary measures are needed ‘to protect national security.’ … This drastic action has never been taken publicly against an American company.”
(p. 6)

Microsoft agrees with Anthropic’s core safety restrictions

“Microsoft’s position is that AI should be focused on lawful and appropriately guarded use cases. For example, AI should not be used to conduct domestic mass surveillance or put the country in a position where autonomous machines could independently start a war. Microsoft believes the government itself appreciates and respects this position.”
(p. 3)

Immediate enforcement would hurt the military and innovation

“The Determination immediately requires DoW to modify existing contracts to remove ‘covered item[s] of supply’ and ‘covered system[s].’ … Because those statutory terms sweep broadly, the Determination could be interpreted to ban the procurement of products and services that merely use or make available Anthropic’s products, thus imposing immediate financial harms on Microsoft and other similarly situated industry partners.”
(p. 4)

Industry impact:

Massive immediate costs on contractors

“If the Determination takes effect immediately, Microsoft and other government suppliers will be compelled to expend significant resources assessing the contractual impacts of excluding Anthropic and Anthropic products. … The costs for these actions—including reengineering, re-procurement, and associated legal and administrative costs—will be incurred immediately.”
(p. 5)

Unintended consequences for the entire tech sector

“The technology sector powers innovation through interconnectedness. … If a single piece of that interconnected offering (long accepted and supported by the government) can be disrupted by a supply chain risk determination, companies will change the way they approach their commercial and public sector relationships in future. … the Determination creates significant risks for commercial collaboration.”
(p. 6–7)

Call for negotiation instead of litigation

“The harms articulated above may be avoided through a negotiated resolution, which weighs in favor of the public interest. There is reason to believe a negotiated resolution is possible here. … Maintaining the status quo while this Court considers the issues raised would simultaneously allow one or more of these alternative pathways to non-judicial resolution to be realized.”
(p. 8)

Big-picture warning about hurting America’s AI lead

“As the Administration has recognized, the United States is in a ‘race to achieve global dominance’ in artificial intelligence and ‘[w]hoever has the largest AI ecosystem will set global AI standards and reap broad economic and military benefits.’ … This is not the time to put at risk the very AI ecosystem that the Administration has helped to champion.”
(p. 9)

Keeping in mind the significant investments from both Google and Amazon in Anthropic, we might be seeing a unified “big tech vs government” conflict brewing. This could mark the first significant shift from the dynamic established during the last election and could result in significant spending against the administration during the midterms. Currently I am not convinced of this outcome because most of this situation is Anthropic’s own doing; it’s also reversible if they bend the knee. The more cynical viewpoint would be that this is a token show of support in order to protect shareholder interests, but that all the hyperscalers have a vested interest in Anthropic’s rapid growth being curbed.

This is counterintuitive since all of them hold significant equity in the company, but with the launch of Claude Marketplace as a significant competitor to all the workflow tools and security plugins, and the general attitude of Dario to pick fights with anybody, it’s not difficult to see that behind the scenes everybody would breathe a little bit easier if Anthropic is pushed back to second place in the AI race.

This is a conversation with Téo Doremus, co-founder of Advocacy AI, a litigation-focused AI platform launching out of stealth. Téo is a former securities and M&A litigator at Robbins Geller who left Big Law to build what he describes as the intelligence layer for litigation. We covered the product thesis, how it differs from Harvey and Legora, an early customer win, go-to-market strategy, and what it actually costs to leave a legal career behind.

What is Advocacy, and why does it need to exist?

If you say “legal tech” right now, two names dominate the conversation: Harvey and Legora. Both are overwhelming in terms of mindshare, and trying to win attention and distribution today seems incredibly difficult. So the obvious question for any new entrant: why now, and why this?

Téo Doremus, a former securities and M&A litigator at Robbins Geller, has a specific answer. The core thesis behind Advocacy comes from a realization he had while practicing: transactional lawyers and litigators are fundamentally different animals, and no one was building for the latter.

“As a transactional attorney, I used to do many deals, and then I used to litigate many deals, and I realized how different transactional lawyers and litigators were. In terms of writing, in terms of posture, the work is fundamentally different. A transactional lawyer is trying to make something happen. Of course you have some level of adversarial, but it’s a deviation from the main goal. In litigation, the main goal is to actually go and fight. You disagree on everything. And then you disagree on how you disagree about things that you disagree on.”

The implications for building software are significant. In transactional work, timelines are compressed: you negotiate, you do due diligence, you either deal or you don’t. Litigation is more like a story, a book. “It’s really hard to predict chapter eight if you haven’t read chapters one through seven. You need to have everything from the moment this thing started all the way to where you are right now, because the analysis is going to change so much.”

This is where Advocacy’s architecture diverges from the rest of the market. What Téo calls the “atomic unit” at Advocacy is the matter. Everything is nested at the matter level. Users cannot even access the app until they train Advocacy’s AI on their specific case: core documents, theories, strategy, multiple data points. That context then evolves constantly, informed by every new document, every chat, every calendar event.

“Advocacy is a case memory platform. Everything that you do in the app for that particular matter has that context. It’s ever evolving. Which allows you to largely reduce the need for prompt engineering and the risk for hallucination, because you’re essentially grounding the AI into the case context.”

So where does Advocacy sit relative to the competition? Legora focuses on application add-ins, editing, mobile work, and collaborative workflows. Their play is making AI available where you already work, inside your document environment. Harvey is running the enterprise SaaS playbook: security with Vault, automation workflows, and a basic assistant layer. Both are also pulling in external legal databases and compliance frameworks for advisory.

Advocacy takes a different position: both of those approaches are great for transactional workflows, but they’re not built for the deep, evolving, adversarial context of litigation.

Téo is careful to draw the distinction precisely: “I don’t think that Harvey or Legora cannot handle complex cases. They can, especially in transactional. But litigation is very different. The requirements are very different. As a litigator, a lot of it is a moving target. The theories change and shift. The goal is not always super clear. A lot of it is investigative research. That’s why I talk more about enhancement than rigid workflows, because workflows presuppose that you know what the output should be. A lot of what I found myself doing as a litigator was exploring. Litigation is a contest over context.”

Subscribe now

Legal AI and product strategy

The natural follow-up is: why not just wait for the foundation models to get smarter? Every litigator will eventually have access to Claude, Chat GPT, or whatever comes next, with a million-token context window and the ability to ingest documents. Why build a custom platform?

“Multiple reasons. One, on the security and safety side, we are single-tenant deployed. Your documents are physically, literally away from others. That’s not the case on Claude or ChatGPT. Two, when you own the entire contextual universe, you can do a lot of things. Yes, you can be on Claude and connect your document manager or Google Drive and then ask it to pull documents and open Word. But now you’re involving three tools, three different surfaces of attack. And they don’t talk to one another. It’s just a file being passed from one API to the other, with zero context other than whatever the user has written.”

With Advocacy’s system, documents live in a single contextual universe. They can be reused across workflows, emails, and analysis. When Dossier (Advocacy’s AI powered document storage system) communicates with Associate (their AI chat assistant), it already knows at ingestion what a deposition transcript is and how it relates to the case. “Your search becomes better because you’re not just throwing keywords through an API and hoping whatever is on the other end understands what you mean. Every single piece of our app is talking to one another in a way that nobody else can reproduce, because it’s designed by us.”

Téo acknowledges the risk: “It’s a big bet, because if people say they don’t want to use your platform, then I’m out of it. But the bet is that this contextual box is going to be self-improving. And that’s what we’re starting to see.”

On where AI and litigation go in the next six to twelve months, Téo sees acceleration in connectors and integrations across the market, but with an important caveat: “I would be cautious. I don’t think integration equals context. It can be an amazing way to bring context when that context is carefully selected and orchestrated. But it can also be just another slot. You connect everything and see what happens, when in reality you have no real orchestrator that owns the process. That’s what we try to focus on.”

He frames the broader vision around what he calls the “digital desk.” As a litigator, everything collapses onto your physical desk: you pull case law from Westlaw and LexisNexis, you pull evidence from your e-Discovery platform, you add your own work-product, and then you work across all of it. “That’s Advocacy. That’s why people ask, ‘Are you e-Discovery? Are you this?’ We’re the thing that was not possible before AI. And I think that’s what being truly native AI means: you have capabilities that are squarely impossible without AI. We try to have this approach for everything we do, down to our editor. Microsoft Word was designed at a time where AI was not on anyone’s mind. So there are capabilities that can be created with AI in drafting that should be created.”

This is also why Advocacy is unlikely to pursue a connector-first strategy: “I don’t think you’ll see us having a connector anytime soon, because I think you’re conceding to a degree that you’re not really a platform player. The whole premise of Advocacy is that we’re the top intelligence layer that aggregates from your own personal notes, your own case law, and your own e-discovery, and your own work-product.”

Customer stories

One of the most concrete ways to understand what Advocacy does differently is through an early customer engagement. Téo shared a story from the company’s first major proof point.

He had been invited to speak at a panel at a law school, alongside representatives from major legal tech players. He leaned heavily into the litigator-to-litigator narrative: “I really leaned into the story of, ‘I’m a litigator, this is for litigators, by litigators,’ with little details that are very hard to know or replicate if you’re not a litigator speaking to litigators.”

After the event, a managing partner from a firm in the audience reached out. They had an e-Discovery problem that no traditional platform could seemingly solve. A group of defendants in a case had been communicating through an iPhone app. Voice messages were sent within those conversations, but the voice messages were produced separately, untied to any specific chat thread. The firm had hundreds of loose voice memos and videos, plus over a hundred chat threads involving multiple defendants and third parties.

“The traditional e-Discovery world has no answer to this. The traditional response was: let’s get a bunch of lawyers, have them listen to every single voice memo, and try to figure out where they belong.”

The low moment came when the client initially pushed back, wanting integration with their existing tools hoping to use Advocacy to speed up that manual process. Téo took a different approach: “I said, ‘No worries, we can try to do this on the side. You go with the traditional way, and we’ll see who wins.’”

Advocacy’s team built an audio and video ingestion capability, transcribed every voice memo, created the contextual universe for the case, and ran a straightforward prompt: “here are all the text chats, here are the voice messages, please pair them based on context.” The AI matched the voice messages to their corresponding threads with high confidence, working across multiple languages.

“It was actually stunning to see. For example, in one of the first chat threads, the AI identified that in this conversation, two people are talking about kids and how annoying they are, and matched it with this audio that says ‘mama, mama, please go, mama.’ We were able to pair those voice messages with very high confidence. The firm became a client for virtually everything they do now.”

GTM and Hiring

Advocacy raised a $3.5 million seed round, intentionally designed around what Téo calls “the ideal cap table.” Relativity, through its innovation arm Rel Labs has joined, along with Fenwick & West LLP, prominent legal scholars, and a number of litigation firm partners. The goal was to ensure that attorneys and legal professionals have a stake and a voice in the company, not just as advisors but as investors.

“It took us some time to assemble the ideal cap table rather than just raise and announce it. Because you’re dealing with lawyers and law firms in a high-trust environment. It doesn’t happen in three weeks. We thought it was the right call for us to not launch immediately, but really take the time to introduce ourselves and make the right connections.”

The go-to-market team is built around range. Jim Watson leads sales, bringing decades of legal tech experience and relationship capital, even though he doesn’t come from a litigation background. Téo explains the logic: “When you do legal tech, you actually rarely deal with lawyers the first time you meet them. A lot of the folks doing the filtering are non-lawyers. Jim understands that. He has not only sold to lawyers but to non-lawyers as well.”

The rest of the team fills in around domain credibility. Téo’s COO Isabella is a former litigator. Other lawyers on the team can be pulled into demos when litigators show up at the second or third meeting with litigator-specific questions.

But the real insight on hiring came through in a small detail: what got Téo to hire Jim was learning that he was already using n8n to automate parts of his own sales workflow. “You don’t even have to be an expert to realize that if this guy with his background is already on n8n trying to automate sales, we can work with that. That’s been the spirit ever since: people who are curious and have something they can bring to the table.”

For the early stage, Téo was explicit that this was not going to be product-led growth for a while: “That’s what I told the devs. That’s why we took the time to nurture those relationships, get on a lot of pilots and trials. Even the first six months of Advocacy was just me calling around my buddies from law school and lawyers, saying, ‘How would you think about this? If somebody came out with something like this, what would you respond?’ It’s always the process at Advocacy.”

Practitioner-first culture

“Built by litigators, for litigators” is a phrase that gets thrown around in legal tech, but at Advocacy it translates into specific organizational practices. When a new lawyer or litigator joins the company, they do a presentation for the entire team: “How I would go about my day as a litigator before joining Advocacy and before AI.” They walk through a fictitious case, step by step, and then show the same workflow with Advocacy. The engineering team sits in the room for the whole thing.

“It’s beautiful what happens. Devs will be working on something like a window or a particular feature, and they no longer have to guess or call us and ask how we would think about this. They know inherently now that litigators want maximum control. So if the question is, ‘Should I hide the citations unless somebody asks for them, or should they appear as soon as the answer is ready?’ Probably the latter, because you’re dealing with people who want maximum control. For them to make those kinds of decisions without having to speak to anyone, they first need to understand the universe.”

The learning goes both directions. Téo admits that the legal side had to learn the pace of software development: “I didn’t know what staging was. I didn’t know what a QA was. I didn’t know what regression bugs were. We had to learn the pace of software, temper our ambitions on new products, and understand that there’s a difference between coding and scaling.”

The company runs a flat structure and deliberately keeps the team small. They have an AI-Ops engineer whose job is to make sure that if they hire someone, it’s because they desperately need to. “So far, the only hires we’re making are domain knowledge experts, whether it’s infrastructure for the engineers or a new practice area for the litigators.”

One of their staff engineers, formerly a lead at Dropbox, captures the other side of the cultural equation: “He keeps saying, ‘Man, I’m so glad I’m here, because I get to fix all the stuff that drove me nuts at Dropbox. The architecture didn’t allow for it, and you had no AI. Now you can literally code at the speed of thought.’”

Téo sees team size as a signal of your AI thesis: “If you’re not able to grow without hiring all the time, what does that say about your own AI thesis?”

The personal journey

The transition from practicing securities litigation to building a technology company wasn’t a clean break. Before fully committing to Advocacy, Téo actually took on a few cases of his own after leaving Robbins Geller, specifically to test whether AI-assisted litigation could work in practice.

“We told the clients, ‘Hey, we could go against these five lawyers here, and we’re going to be using AI if you want. If not, that’s fine, but we’re trying to make a proof of concept.’ And we did. We gave the other side a run for their money. We were able to run and file motions that otherwise would have been completely cost-prohibitive.”

The insight that stuck: “We can bring $900-an-hour level of service to a non-$900-an-hour legal buyer. Some people frame this as a collapsing industry. I don’t think it is. It’s like being an accountant and then the calculator comes out. Yeah, if you charge for crunching numbers on paper, your job is over. But there’s so much more that the calculator unlocks in terms of guidance and argumentation.”

The personal cost is real, but perhaps not where you’d expect it. Téo says he works more hours now than he did in Big Law, but the hardest part isn’t the workload: “The personal cost, the trade-off, and that’s true for every single litigator that has joined us, is that you’re no longer a litigator. That thing of making a case your own, fighting tooth and nail, breathing that case, it’s kind of gone.”

The shift has been toward what he describes as a “vicarious impact.” The time spent on the product and fine-tuning and building new capabilities, they can see how it plays out in actual litigation through their clients. “We’re now deploying the tools that allow that kind of service to reach more people and just make people better. Make lawyers better. Refocus things on higher-value items, which is strategy.”

Asked whether the experience has been more energizing or draining, Téo offers a surprisingly grounded answer: “Funny enough, it’s taking less of a brunt than my legal career did. I work more hours, don’t get me wrong. But the biggest struggle at the beginning was identifying, ‘Okay, where’s my impact now as far as being an advocate?’ It has taken more of a meta impact for us.”

Advocacy has announced its $3.5M seed round with an official launch on March 6, just ahead of Legal Week on March 9th, where the company is a finalist in the Leaders in Tech Law Awards (Litigation Technology). The AI-native case memory platform is currently deployed with paying clients and running active pilots with top-tier litigation firms. Learn more at advocacy.ai.

Last year, businesses running on Stripe generated $1.9 trillion in total volume, up 34% from 2024, and equivalent to roughly 1.6% of global GDP. Our programmable financial services now power more than 5 million businesses directly or via platforms, including all of the top AI companies, many of the largest bluechip companies (90% of the Dow Jones Industrial Average), most of the biggest tech companies (80% of the Nasdaq 100), and a significant fraction of freshly minted startups (25% of all Delaware corporations are now created with Stripe Atlas).

Beyond payments, these businesses are using Stripe to accelerate their growth with billing and subscription management, tax compliance, fraud prevention, embedded finance, global treasury management, and much besides. Link, the easiest way to pay online, is now used by more than 200 million people.

Stripe remained robustly profitable, allowing us to continue investing heavily in product development (with more than 350 product updates last year) as well as acquisitions. Since our last update, we acquired Privy, which powers more than 110 million programmable wallets, and Metronome, which powers the intricate usage-based billing models used by companies like OpenAI, Anthropic, Confluent, and NVIDIA. Metronome joins our Revenue suite, which is on track to hit an annual run rate of $1 billion this year.

All in all, 2025 was a strong year for the internet economy, and we’re delighted to see so many of Stripe’s customers do so well.

I think that the last 18 months have seen a new leg up of what we can call “the internet economy”. AI-native startups have completely skipped the legacy payment ramp, and the Stripe billing page is one of the most consistent experiences you will have when subscribing to an AI product. These days I’m actually surprised when I’m searching for an invoice from a new company and they are not running on Stripe.

At heart, competitive markets are a sorting machine. They direct profits, capital, and talent to the places of greatest impact, as determined by customers voting with their wallets. Historically, this sorting happened methodically. It typically took decades for a household name to be unseated or for a new entrant to reach meaningful scale. The sorting machine is now whirring faster: winners and losers are being anointed more quickly and more intensely.

The sorting machine is now whirring faster: winners and losers are being anointed more quickly and more intensely. Today, the most profitable third of publicly listed companies in the US account for two-thirds of total market capitalization, the highest share since data began in 1963. And much of this is a story of profit concentration, not just valuations: the top 10% of the S&P 500 by market cap now account for roughly 59% of the index’s total profits, which is elevated relative to recent history.

Much of this is driven by bifurcation within industries. In retail, for example, US brick-and-mortar sales grew just 5% over the past 3 years, whereas ecommerce sales grew 30% over the same period (both in inflationadjusted terms). In air travel, the “big 3” of American, Delta, and United all increased their share of industry revenues and profits over the past decade. (Indeed, Delta and United accounted for nearly all US airline profits in 2025.)

In healthcare, hospital and insurer profit shares have contracted significantly since 2019, but health tech is on track to exceed $110 billion in EBITDA by 2029. Each sector has its own particular dynamics, but the pattern is clear: a cohort of companies is pulling away. Economy-wide, demand for software, computers, and data center investment drove nearly half of all US GDP growth in 2025 and will likely soon be the majority of US growth.

This is an interesting way of looking at the (software) market, particularly the mental model of variance vs concentration. It appears that it's no longer sufficient just to pick the right directional wave (AI coding agent companies), but you also have to hit the most likely concentration winner (Claude Code > Cursor > Replit).

Based on what we can tell from the set of businesses that started on Stripe in 2025—a remarkable cohort—there are no signs of the sorting machine slackening. In 2025, many more new companies joined Stripe than ever before, with more than half of them (57%) based outside the US. This new cohort is by far the highest performing and fastest moving we’ve ever seen, growing around 50% faster than the 2024 cohort. The number of companies reaching $10 million ARR within 3 months of launch was double the 2024 count. This seems to be part of a larger expansion and acceleration in our industry.

After years of relative calm, the number of iOS apps released in December 2025 jumped by 60% year over year. (Someone should check the App Store review team’s sleep scores.) Even code production is accelerating: pushes to GitHub, which grew roughly 10%–12% in prior years, surged 41% between Q3 2024 and Q3 2025. As building gets easier, we’re working on making Stripe even simpler to integrate—including for the agents.

We recently introduced claimable sandboxes, which let you start using Stripe directly from your AI coding tools like Manus, Base44, Replit, and Vercel. When you’re done with v1 and your product is ready to launch, that sandbox converts into a live Stripe account with its configuration intact. More than 100,000 sandboxes have been created this way.

We’re also improving Stripe Atlas, the world’s easiest way to incorporate a business, which saw a 41% increase in company formations last year. Atlas companies are monetizing sooner: in 2025, 20% of Atlas startups charged their first customer within 30 days, up from 8% in 2020.

As we look at these figures, there is an obvious question: is 2025 an anomaly or the beginning of a new regime? Time will adjudicate, but our best guess is that the 2025 acceleration is the start of a larger inflection in entrepreneurship and creativity facilitated by advances in large language models. We have an ambitious roadmap of improvements planned. Stripe will be the best way to build a business in the era of AI.

We are seeing a clear acceleration in the number of new AI-native companies being created, and more importantly, accelerated path to high growth.

For those with aspirations to “go global,” the conventional playbook used to be a steady, sequential progression: win at home, then push abroad. It took Coca-Cola 20 years to bottle its first soda in Cuba, while McDonald’s and Starbucks waited 27 and 16 years, respectively, to serve their first customers in Canada. After the arrival of the World Wide Web, free services tended to launch globally, but their monetization machinery still operated with a time delay.

When Facebook changed from a college-only network to a public platform in 2006, anyone with a browser could create an account, but anyone with money couldn’t necessarily advertise. Support for international currencies didn’t arrive until 2009, five years after the company was founded.

For its part, Google only accepted its first GBP payment from an advertiser in the UK (a live lobster mail order firm!) in 2002, four years after launching its search product globally. Over the last few years, the country-by-country expansion model has melted away. The “domestic market” for a new generation of internet businesses is the internet itself. Nearly every AI product you’ve heard of is all the rage in every country you’ve heard of. ChatGPT, Claude, Replit, Lovable, Base44, Vercel, Cursor, Midjourney, and many more launched globally by default.

This isn’t merely about incremental revenue from a “long tail” of international users. In many cases, the “long tail” is much of the dog. Among Stripe businesses with mostly international revenue, 30% of that revenue comes from countries that are neither their home market nor one of the top 10 global economies.

This is possible largely due to infrastructure that no longer makes foreign demand feel foreign. Last year, we enabled businesses to launch a localized checkout in more than 100 countries simultaneously, complete with localized pricing to maximize conversion, more than 120 payment methods, and local tax compliance supported out of the box. Sometimes improved infrastructure is only felt after decades; other times, pent-up demand reveals itself overnight. Gamma is a California-based AI platform used by 70 million people to create presentations.

When Gamma joined the first cohort of businesses on Stripe to accept UPI payments in India, its Indian revenue leapt 22% that same month.

While the risk of regulations hampering progress has been a constant discussion in the last year, when money talks, cross-border commerce will flow. Stripe has become the backbone of fundamentally an international scaling model for AI-native companies, even if they take their fair share of margin fees.

It may be a crypto winter, but it’s a stablecoin summer. After a decade of stablecoin volumes tracking the undulations of crypto asset prices, last year saw a clear divergence. In 2025, the price of Bitcoin dropped precipitously (and is now down 50% from October), but stablecoin payments volume doubled to around $400 billion, 60% of which is estimated to represent B2B payments. Bridge, the stablecoin orchestration platform we acquired, saw volume more than quadruple. Stablecoin payments are advancing quietly and inexorably as real-world uptake continues apace.

This growth has been catalyzed by a profusion of new capabilities. A Y Combinator founder can now receive funding in stablecoins, hold them in a Stripe financial account, and use them to pay their first engineers, who could be anywhere in the world. SaaS platforms are using stablecoins to collect recurring payments, thanks to a new smart contract that obviates the need for wallet owners to manually sign each transaction. Enterprises leaning on stablecoins to expand internationally now have better tools to embed digital wallets directly into their core products. With Privy, companies like Ramp and Deel have a single API to provision easy-to-use wallets in both custodial and noncustodial models. This makes it possible to build fully global products on day one.

The interoperability between crypto and fiat is also rapidly improving. In April, Bridge partnered with Visa to introduce cards that allow businesses and consumers to spend their stablecoins just like any other card. The payment is deducted from a stablecoin balance and automatically converted to the local currency; the business receives the funds just like any other payment, serenely insulated from the underlying stablecoin mechanics. Phantom, one of the most popular crypto wallets with 20 million monthly active users, is using Bridge to roll out stablecoin-backed cards to its customers.

The more interesting question for Stripe specifically is what fee compression actually does to their business model. Stablecoins are great for commerce broadly, but Stripe’s margin story depends on being the indispensable layer between buyer and seller. If settlement becomes commoditized, Stripe’s bet is clearly that they win on everything built around the payment: billing, tax, fraud, treasury.

Together with the rise of the agentic economy, one of the most “obvious” directional paths that will play out over the next years is even higher percentage of all commerce moving to digital in some form or another. This is what it can potentially look like:

OpenClaw showed a certain appetite for the "personal assistant who is hands-off" application. Essentially just letting AI figure out things on your behalf. While both the technology and the public perception of such a way of living are in its early stages, we've seen how quickly audiences that were used to doing everything themselves (developers), can move to a model where they are comfortably handing off almost everything to AI (agentic coding).

The practical implication for Stripe is that agentic commerce likely means a significant volume of transactions that are never initiated by a human at all. An agent booking a service, spinning up infrastructure, or purchasing an API call on your behalf still routes through a payment layer. Stripe's early moves to make their sandbox natively accessible from within AI coding tools suggests they understand this better than most.

Today’s entrepreneurs and innovators have tools and reach that prior generations of industrialists could not have fathomed. We will hopefully soon witness the combinatorial effect of human ingenuity paying out in the form of productivity gains and improved living standards everywhere.

Last year, Joel Mokyr was awarded the Nobel Prize in Economic Sciences. Mokyr is widely known for emphasizing the importance of culture relative to the traditional economic inputs of capital, labor, and technology. Eighteenth-century industrialists didn’t just have coal or geography on their side. They had a new culture—an “improvement mindset” that saw the status quo as imperfect and correctable.

In The Political Economy of Technological Change, Mokyr also observed that new technologies have in the past often failed, despite their economic superiority, because technological decision-making implicates not only suppliers and customers, but also a broad variety of non-market “aggregators” (regulators, committees, courts) that influence what is adopted. As AI and the internet expand the scope of what’s possible, synthetic impediments to adoption and adaptation will become increasingly costly. Our bifurcating economy shows that growth is contingent on the application of useful knowledge and not some preordained result of its abstract availability.

AI harbors the promise of enormously improving drug discovery… but the potential will only be realized if we make the regulatory process, including clinical trials, faster and cheaper. Entrepreneurs in Europe can boost tepid economies with new tools… but only if well-intentioned yet counterproductive burdens such as the EU AI Act are curtailed.

Next-generation approaches to nuclear energy could usher in energy abundance… but only if we overhaul vetocratic regulatory regimes. Autonomous transport and logistics— from long-haul trucking to drones—could dramatically reduce the cost of physical goods… provided we don’t let a slurry of local ordinances harden into a blockade. Mokyr wrote about the importance of the Republic of Letters in catalyzing the industrial revolution.

Today, we inhabit a Republic of Permissions: a filtering sieve of nonmarket aggregators. While many of our strictures are sensibly motivated, it’s more important than ever to ensure that they carefully balance the benefits achieved with the possibilities foreclosed.

We’re privileged to support many businesses with the tenacity to show what’s possible. Mistral AI and Bending Spoons are proving that world-class European talent can puncture the regulatory permafrost; Zipline and Varda are earning permissions for intricate new hardware inch by inch; while Spring Health and Maven Clinic are stitching together a new software layer for modern healthcare. We continue to believe both in the importance of ideas in fueling economic progress and that many of the best ideas are undervalued.

…

Much of this letter has been dedicated to advances in AI, which can sometimes seem hard to keep up with. The qualitative difference between just-released products and last year’s state of the art is stark. We’re reminded of the phenomenon of falling into a large black hole. If you ever experience that particular misfortune, you won’t actually feel anything special at the moment you cross the event horizon: the path is locally smooth, even though the space of possible futures changes irrevocably upon crossing the threshold.

We write this letter at what may well turn out to be the advent of a different and hopefully much more beneficent singularity. While much around us in 2026 feels similar to prior years, it is also clear that the next decade will look very different to those just gone by. We are as enthusiastic as ever about how vibrant entrepreneurship and wise cultures can contribute to more successful future societies, and we hope that Stripe can play a small role.

So what can we conclude from this annual letter?

Read more

In the background to this event, the Department of War and Anthropic ended up in a high stakes conflict, which at least currently, has led to the designation of the company as a supply chain risk. This played out literally in the hours before President Trump made his announcement regarding the opening salvo of the war.

To call this unprecedented would be significantly understating the impact. The leadership of Anthropic had decided to play a very high stakes game, in relative disregard for the usual way that governments procure technology for defense purposes. During my time in tech, I’ve had the privilege of negotiating software contracts used by governments for national security, and the dynamic has been quite clear from day one. During one of those conversations, a procurement officer flat out told me “look, these renewal clauses that would limit how we use the product are useless, since if it’s part of an operation, we are not going to bother talking to you about it.”

Most of the time, companies would not even have a basic idea of how their products are being used. Access to the teams working with the products typically require government clearances, and the companies that apply to acquire those on behalf of their employees, also sign up to very specific protocols with severe impact clauses. The outcome of violating those goes way beyond “contract got cancelled and we had to refund them”, which has served as a clear deterrent towards software companies acting in what can be perceived as bad faith. Since it is typically very rare for a company to be winning this type of business and still be only founder led (i.e. headcount below 5), by definition a negative outcome would impact the lives of regular employees who have very little incentive to risk going to prison.

Now, let’s try to untangle what has just played out.

Statement from Dario Amodei on our discussions with the Department of War

I believe deeply in the existential importance of using AI to defend the United States and other democracies, and to defeat our autocratic adversaries.

Anthropic has therefore worked proactively to deploy our models to the Department of War and the intelligence community. We were the first frontier AI company to deploy our models in the US government’s classified networks, the first to deploy them at the National Laboratories, and the first to provide custom models for national security customers. Claude is extensively deployed across the Department of War and other national security agencies for mission-critical applications, such as intelligence analysis, modeling and simulation, operational planning, cyber operations, and more.

Anthropic has also acted to defend America’s lead in AI, even when it is against the company’s short-term interest. We chose to forgo several hundred million dollars in revenue to cut off the use of Claude by firms linked to the Chinese Communist Party (some of whom have been designated by the Department of War as Chinese Military Companies), shut down CCP-sponsored cyberattacks that attempted to abuse Claude, and have advocated for strong export controls on chips to ensure a democratic advantage.

Anthropic understands that the Department of War, not private companies, makes military decisions. We have never raised objections to particular military operations nor attempted to limit use of our technology in an ad hoc manner.

However, in a narrow set of cases, we believe AI can undermine, rather than defend, democratic values. Some uses are also simply outside the bounds of what today’s technology can safely and reliably do. Two such use cases have never been included in our contracts with the Department of War, and we believe they should not be included now:

• Mass domestic surveillance. We support the use of AI for lawful foreign intelligence and counterintelligence missions. But using these systems for mass domestic surveillance is incompatible with democratic values. AI-driven mass surveillance presents serious, novel risks to our fundamental liberties. To the extent that such surveillance is currently legal, this is only because the law has not yet caught up with the rapidly growing capabilities of AI. For example, under current law, the government can purchase detailed records of Americans’ movements, web browsing, and associations from public sources without obtaining a warrant, a practice the Intelligence Community has acknowledged raises privacy concerns and that has generated bipartisan opposition in Congress. Powerful AI makes it possible to assemble this scattered, individually innocuous data into a comprehensive picture of any person’s life—automatically and at massive scale.

• Fully autonomous weapons. Partially autonomous weapons, like those used today in Ukraine, are vital to the defense of democracy. Even fully autonomous weapons (those that take humans out of the loop entirely and automate selecting and engaging targets) may prove critical for our national defense. But today, frontier AI systems are simply not reliable enough to power fully autonomous weapons. We will not knowingly provide a product that puts America’s warfighters and civilians at risk. We have offered to work directly with the Department of War on R&D to improve the reliability of these systems, but they have not accepted this offer. In addition, without proper oversight, fully autonomous weapons cannot be relied upon to exercise the critical judgment that our highly trained, professional troops exhibit every day. They need to be deployed with proper guardrails, which don’t exist today.

To our knowledge, these two exceptions have not been a barrier to accelerating the adoption and use of our models within our armed forces to date.

The Department of War has stated they will only contract with AI companies who accede to “any lawful use” and remove safeguards in the cases mentioned above. They have threatened to remove us from their systems if we maintain these safeguards; they have also threatened to designate us a “supply chain risk”—a label reserved for US adversaries, never before applied to an American company—and to invoke the Defense Production Act to force the safeguards’ removal. These latter two threats are inherently contradictory: one labels us a security risk; the other labels Claude as essential to national security.

Regardless, these threats do not change our position: we cannot in good conscience accede to their request.

It is the Department’s prerogative to select contractors most aligned with their vision. But given the substantial value that Anthropic’s technology provides to our armed forces, we hope they reconsider. Our strong preference is to continue to serve the Department and our warfighters—with our two requested safeguards in place. Should the Department choose to offboard Anthropic, we will work to enable a smooth transition to another provider, avoiding any disruption to ongoing military planning, operations, or other critical missions. Our models will be available on the expansive terms we have proposed for as long as required.

We remain ready to continue our work to support the national security of the United States.

The core two issues that Anthropic leadership claims are at the root of the discussions are domestic surveillance and autonomous weapons, with the important distinction here that they see the latter as an issue because of the maturity of the technology.

The thing is that, according to the government, they’ve been very clear that they want language around “usage for all lawful purposes”.

This is not something that is from this week, but is actually tied to a Memorandum on AI use from 1/9/26:

SECRETARY OF WAR 1000 DEFENSE PENTAGON WASHINGTON, DC 20301-1000

JAN 9 2026

MEMORANDUM FOR SENIOR PENTAGON LEADERSHIP COMMANDERS OF THE COMBATANT COMMANDS DEFENSE AGENCY AND DOW FIELD ACTIVITY DIRECTORS

SUBJECT: Artificial Intelligence Strategy for the Department of War

Accelerating America’s Military AI Dominance

President Trump makes clear in Executive Order 14179, “It is the policy of the United States to sustain and enhance America’s global Artificial Intelligence (AI) dominance in order to promote human flourishing, economic competitiveness, and national security.” In the national security domain, AI-enabled warfare and AI-enabled capability development will re-define the character of military affairs over the next decade. This transformation is a race — fueled by the accelerating pace of commercial AI innovation coming out of America’s private sector. The United States Military must build on its lead over our adversaries in integrating this technology, established during President Trump’s first term, to make our Warfighters more lethal and efficient. To this end, aligned with America’s AI Action Plan, I direct the Department of War to accelerate America’s Military AI Dominance by becoming an “AI-first” warfighting force across all components, from front to back.

The Department will achieve this objective by:

• Unleashing experimentation with America’s leading AI models Department-wide, and rewarding AI-first re-conceptions of legacy approaches;

• Aggressively identifying and eliminating bureaucratic barriers to deeper integration, which are vestiges of legacy information technology and modes of warfare;

• Focusing our investment to leverage America’s core asymmetric advantages in AI computing, model innovation, entrepreneurial dynamism, capital markets, and combat-proven operational data from two decades of military and intelligence operations that no other military can replicate; and

• Executing a set of “Pace-Setting Projects” (PSPs) that will demonstrate the accelerated pace of execution, focus, and ethos we need to stay ahead. The PSPs will also serve as tangible, outcome-oriented vehicles for rapidly completing our buildout of the foundational AI enablers (infrastructure, data, models, policies, and talent) needed to accelerate AI integration across the entire Department.

The seven initial PSPs outlined below establish the new execution standard: single accountable leaders, aggressive timelines, measurable outcomes, and rapid iteration where failure accelerates learning and improvement.

Acceleration Approach

The means we will employ to pursue this strategy will continue to encompass our substantial program funding and workforce focused on AI across the Services and Components. We will also use the timely financial resources provided by Congress in the form of One Big Beautiful Bill, along with expanded budget withhold (Joint Acceleration Reserve) flexibility, to catalyze our accelerated pace of Military AI integration in the immediate term. And we will leverage the access, capabilities, investments, and insights of America’s allies and partners to support our shared objectives, consistent with the President Trump’s AI Action Plan to “Lead in International AI Diplomacy and Security”.

We will re-focus the Chief Digital and AI Office (CDAO) and these enhanced resources to unlock critical foundational enablers needed to accelerate war-winning efforts across the Department, starting with enabling the set of seven PSPs listed below in fiscal year 2026. These PSPs will address key opportunities for enhanced military AI advantage across Warfighting, Intelligence, and Enterprise mission areas:

Warfighting:

1. Swarm Forge: Competitive mechanism to iteratively discover, test, and scale novel ways of fighting with and against AI-enabled capabilities — combining America’s elite Warfighting units with elite technology innovators.

2. Agent Network: Unleashing AI agent development and experimentation for AI-enabled battle management and decision support, from campaign planning to kill chain execution.

3. Ender’s Foundry: Accelerating AI-enabled simulation capabilities — and sim-dev and sim-ops feedback loops — to ensure we stay ahead of AI-enabled adversaries.

Intelligence:

4. Open Arsenal: Accelerating the TechINT-to-capability development pipeline, turning intel into weapons in hours not years.

5. Project Grant: Enabling transformation of deterrence from static postures and speculation to dynamic pressure with interpretable results.

Enterprise:

6. GenAI.mil: Democratizing AI experimentation and transformation across the Department by putting America’s world-leading AI models directly in the hands of our three million civilian and military personnel, at all classification levels.

7. Enterprise Agents: Building the playbook for rapid and secure AI agent development and deployment to transform enterprise workflows.

The PSPs will each be led by an exemplary program leader in partnership with a sponsoring organization. Progress will be demonstrated monthly to the Deputy Secretary of War (Deputy Secretary) and Under Secretary of War for Research and Engineering (USW(R&E)), with initial demonstration by transition-partner user(s) to occur within six months from the date of this memorandum.

The CDAO will also ensure all foundational enablers unlocked by these projects are made available to programs Department-wide in real-time, so accelerated execution by PSPs will enable projects across the Department to accelerate their pace along with them. Therefore, I direct each Military Department, combatant command, and defense agency and field activity to identify within 30 days at least three projects they will prioritize to fast-follow these PSPs. Efforts under the Department’s six Critical Technology Areas – including autonomy, C-C5ISRT, and advanced manufacturing — must continue to push the pace for the Department of War (DoW). And the special initiatives outlined in classified annexes, including those in the Classified Annex provided by separate cover to this memorandum, will also be accelerated. CDAO will track and rank this extended pack of AI efforts by speed and impact, and progress will be reported monthly to the Deputy Secretary and USW(R&E).

AI Compute. As part of our AI and Autonomy acceleration investments, the Department will invest substantial resources in the expansion of our access to AI compute infrastructure, from datacenters to the edge. We will leverage the hundreds of billions in private sector capital investment being made in America’s AI sector through our growing array of creative partnerships with America’s world-leading companies. We will work with interagency partners to establish technical standards for new secure datacenters. And we will support and leverage the American Science and Security Platform being developed by President Trump’s Genesis Mission for science and technology innovation, so our warfighters and capability developers have the full benefit of America’s AI compute resources and latest innovations.

Data Access. I direct the CDAO to enforce, and all DoW Components to comply with, the ‘DoD Data Decrees’ to further unlock our data for AI exploitation and mission advantage. Military Departments and Components will establish, maintain, and update federated data catalogs exposing their system interfaces, data assets, and access mechanisms across all classification levels, as mandated by the Department’s May 2021 memorandum, “Creating Data Advantage.” They will deliver their current catalogs — with all available updates — to the CDAO within 30 days of the date of this memorandum. The Under Secretary of War for Intelligence and Security will ensure intelligence data receives parallel treatment, with exploitation pathways established within the same timeframe. The CDAO is authorized to direct release of any DoW data to cleared users with valid purpose, consistent with security guidelines. Effective immediately, denials of CDAO data requests must be justified to the USW(R&E) within seven (7) days, who will remediate or escalate to the Deputy Secretary. Our data advantage is meaningless if our developers and operators cannot exploit it.

Talent. Finally, I believe the best American talent will see this accelerated posture of AI capability development and adoption at the DoW, and I expect each Service and Component to attract and retain this talent. To that end, I direct use of special hiring and pay authorities Department-wide, as well as novel talent programs from the Office of Personnel and Management and other partners, to accelerate our pace of technical talent hiring into AI roles. And I direct each Component to provide AI hiring and talent development plans to the Under Secretary of War for Personnel and Readiness within 60 days of this memo for approval, denial or modification within 30 days thereafter.

Acceleration Expectations

This strategy will accelerate our advantage, and we must implement it with the Warrior Ethos. Consistent with the refocusing of the Department onto a wartime footing, I expect the following approaches to become internalized as essential elements of our execution in this race to maintain Military AI Dominance:

Speed Wins. We must internalize that Military AI is going to be a race for the foreseeable future, and therefore speed wins. We must weaponize learning speed, and measure and manage cycle time and adoption rates as decisive variables in the AI era. We must accept that the risks of not moving fast enough outweigh the risks of imperfect alignment. I direct CDAO to establish deployment velocity and operational cycle-time metrics for all PSPs, to be a focus of their monthly reporting to the Deputy Secretary and USW(R&E).

AI Model Parity. We are seeing unprecedented velocity in the evolution of the frontier AI models. These models are becoming smarter and more robust every day. The Department cannot be working off models that are months or years old. We must have the latest and greatest AI models deployed for our warfighters. Deploying these capabilities across all echelons is simply not enough, we must be able to support and sustain rapid model updates across all echelons. I direct CDAO to establish a delivery and integration cadence with AI vendors that enables the latest models to be deployed within 30 days of public release. This shall be a primary procurement criterion for future model acquisition.

Wartime Approach to Blockers. We must eliminate blockers to data sharing, Authorizations to Operate (ATOs), test and evaluation and certification, contracting, hiring and talent management, and other policies that inhibit rapid experimentation and fielding. We must approach risk tradeoffs, “equities”, and other subjective questions as if we were at war. To this end, I expect our CDAO to act as a Wartime CDAO and work with the Chief Information Officer to fully leverage statutory and delegated authorities to accelerate AI capability delivery, including cross-domain data access and rapid ATO reciprocity on behalf of pace-pushing leaders across the Department. The USW(R&E) will establish a monthly “Barrier Removal Board” with authority to waive non-statutory requirements and escalate blockers for immediate resolution.

Competition > Centralized Planning. As America’s AI ecosystem demonstrates, robust competition by small teams, with transparent metrics for results, is the engine of commercial AI leadership. We must bring this model into the Department and encourage robust competition to spur faster military AI integration. Small, accountable teams will win over process in a race characterized by dynamic and unpredictable innovation. We will measure success through continuous field experimentation: putting AI capabilities in operators’ hands, gathering feedback within days not years, and pushing updates faster than the enemy can adapt. I direct CDAO to establish AI system usage and mission impact metrics for evaluating the success of these AI acceleration efforts. To enable market dynamics to drive resourcing, decisions about future resourcing and deprecation of associated capabilities will principally be made on the basis of these metrics.

AI-Native Warfighting. Together with capability innovation, we must more fully incorporate AI and Autonomy into military planning; tactics, techniques and procedures (TTP) development; and experimentation processes. I direct each Service Chief and Combatant Commander to designate an AI Integration Lead within 30 days, who will work with the CDAO and be responsible for the co-evolution of AI-enabled capabilities with warfighting concepts and experimentation. I direct CDAO to establish criteria for robust experimentation with AI capabilities. And I direct the Joint Staff to designate a senior official to monitor Service AI warfighting concept development and workflow optimization, and provide me with progress reports on a quarterly basis. We must put aside legacy approaches to combat and ensure we use this disruptive technology to compound the lethality of our military. Exercises and experiments that do not meaningfully incorporate AI and autonomous capabilities will be reviewed by the Director of Cost Assessment and Program Evaluation for resourcing adjustment.

Modular Open Architectures. In the AI arms race, system architectures must enable component replacement at commercial velocity to maintain overmatch. I direct Military Department and Component Program Managers acquiring AI capabilities to enforce Modular Open System Architectures (MOSA) along with the “DoD Data Decrees,” exposing modular interfaces and associated documentation sufficient for third-party integration without prime contractor support.

Clarifying “Responsible AI” at the DoW — Out with Utopian Idealism, In with Hard-Nosed Realism. Diversity, Equity, and Inclusion and social ideology have no place in the DoW, so we must not employ AI models which incorporate ideological “tuning” that interferes with their ability to provide objectively truthful responses to user prompts. The Department must also utilize models free from usage policy constraints that may limit lawful military applications. Therefore, I direct the CDAO to establish benchmarks for model objectivity as a primary procurement criterion within 90 days, and I direct the Under Secretary of War for Acquisition and Sustainment to incorporate standard “any lawful use” language into any DoW contract through which AI services are procured within 180 days. I also direct the CDAO to ensure all existing AI policy guidance at the Department aligns with the directives laid out in this memorandum.

Becoming An AI-First Department

The time is now to accelerate AI integration, and we will put the full weight of the Department’s leadership, resources, and expanding corps of private sector partners into accelerating America’s Military AI Dominance.

Becoming an “AI-First” warfighting force requires more than integrating AI into existing workflows. It requires re-imagining how existing workflows, processes, TTPs, and operational concepts would be designed if current AI technology existed when they were created — and then re-inventing them accordingly.

We must drive this transformation across every aspect of the Department. The expectations outlined above must become technological “AI fitness standards” for our Joint Force. 2026 will be the year we emphatically raise the bar for Military AI Dominance.

I've included the full memo since it outlines quite obviously a significant shift in how the US government is adopting AI, particularly when it comes to the work of the DoW. It's incredibly bullish for the frontier labs, as it basically guarantees wide adoption (including of the latest models as they get released). There is one "but" here, that Anthropic took offense at:

Read more

“Squeaky bum time” refers to the sound made by shifting restlessly on a plastic stadium seat due to anxiety. Anybody who’s watched a soccer game in a stadium would’ve experienced this, hopefully followed by an intense relief of pressure due to your team scoring a goal.

We find ourselves in a similar time today, as technology, the economy, and the art of daily living are all meshed together. Those that pay attention have become spectators and occasional players in this new type of game.

The nervousness in the air is hard to miss. If this was only low-stakes entertainment, it would’ve been easy to laugh about it. Unfortunately, just like in sports, the feelings are more intense when we gamble.

Panic at the disco

In 2025, a press release from OpenAI about working with a company on AI infrastructure was sufficient for a stock rally. In 2026, we are witnessing the reverse. Anthropic launching new vertical products, no matter how vague, is sufficient to lead to a selloff in entire sectors. Squeaky bum time, indeed.

When Anthropic announces that their models will now do security reviews (something that OpenAI actually already did the week before with the launch of ChatGPT 5.3), the reflex across both the professional and retail investor community is to dump the cybersecurity stocks.

It doesn’t matter who the company is or how mission critical they are. Dump it.

It doesn’t matter if it’s running a well-managed business with a massive growth opportunity due to AI. Dump it.

It’s squeaky bum time and most of us are very much out of our depth.

Public market forces

Excluding Saudi Aramco, the top 10 public companies today by market cap are all tightly intertwined with tech. Apple leads in consumer tech, Tesla in physical AI, and TSMC in the production of the silicon that powers everything. While retail investors have been influential in driving algorithmic sentiment, the reality is that the large day-to-day movements of tech stocks can be traced back to the large financial institutions that form the largest holders in each public company.

The funny thing about these financial institutions is that many of their investment decisions are driven by regulations and processes, rather than actual insight into what is a great tech company or what is a poor one. This is why you'll often find them having proportional ownership based on market caps, rather than weighted according to, you know, actual insider knowledge.

Fortinet is one of the competitors of CrowdStrike in many areas of cybersecurity. It is an objectively bad software vendor, known to acquire poorly performing companies in order to "cover gaps in their portfolio." It has repeatedly been exploited by nation-state attackers and is widely avoided by technically competent security teams. Yet, it has higher revenue than CrowdStrike.

This is made possible by a business model in cybersecurity where 90% of technology is acquired by Managed Service Providers, who operate it on behalf of their end customers. Those MSPs are mostly interested in maintaining a margin and outcompeting other cheaply priced competitors, rather than doing the actual job of securing their customers.

While you would expect this to be common knowledge for the curious and informed investor, it doesn't appear to be the case. It also doesn't seem to be obvious to the hundreds of thousands of otherwise talented individuals who work in cybersecurity companies today, who often lack even basic understanding of the sector.

Subscribe now

Reasoning through the panic

“How you do one thing is how you do everything” is one of the most important observations when it comes to high-performance outcomes in sports. There is a certain level of mental discipline required at all times if you would like to be part of the elite performers. And how most people today are approaching the potential impact of AI across the economy shows a lot of reasoning sloppiness, even if they’ve been directly exposed to the inner workings of tech companies.

I think that the last decade of an evergreen “number goes up” tech market has had some profound negative consequences for your average tech worker, who often happens to be an active retail investor.

The vast majority of them have never bothered to learn how to deploy a Kubernetes container, think through a MITRE ATT&CK pattern, or do an agentic evaluation. The day-to-day of deeply technical activity behind the most important tech layer today, cloud infrastructure software, is a mystery to them.

The same mental laziness can be seen across all parts of cloud infrastructure software today. Salesforce struggling to convince customers to use their system of record as a productized data mesh is a completely different situation from companies rapidly adopting Databricks for essentially the same use case. If they could, most investors would've probably dumped Databricks as well, but the reality today is that if you are running a tech company that matters, you should avoid going public for as long as possible.

SaaS apocalypse and other disasters

There are some merits to the “SaaS apocalypse.” The main one comes back to the evergreen “number goes up” tech market. Since the growth of those businesses was aggressive over the last decade, most of them have never learned to run an efficient organization (or actively avoid that). When growth slowed down in 2022, many made an attempt at efficiency, by promoting actively how they are benchmarking according to the rule of forty. “We are rule of 50! 60! 70!”, CEOs would proudly brag about on earnings calls, while approving expensive stock based compensation grants across the whole company. Funnily enough, the only company that actually benchmarks as profitable according to the criteria above is Palantir, one of the most divisive and “overpriced” companies on the market.

Which brings us to one of the audiences with the most negative sentiment: those working in tech. After a decade of being in the top percentile of earners, partly due to the generous stock compensation packages they got, those working in tech are hurting. They are faced with layoffs, increased accountability and a slow realization that most tech incumbents are not adapting well to AI. Regardless of having worked in the industry for years, their own understanding of how things work is very limited, which is making them uncomfortable. If they kept their company stock, they have likely lost a significant portion of their net worth in the last twelve months. Many helped the selloffs by dumping every time their company stock would vest, as shown by the awkward statistic that the majority of tech company management have never invested in their own company, but have actively been selling whatever they have.

The view on the inside is not too pretty either. Since the generous stock grants have always been seen as a motivational commonality (”we are all shareholders”), CEOs in the last year have tried to come up with different excuses on why the market is punishing their specific company, while trying to optimize for better optics through layoffs and performance improvement plans (the silent layoffs). The stock market became a discussion of company meetings and internal emails, but that did little to pause the intense selling once the internal trading window was open.

So if the tech investors and tech employees are bearish on the future of SaaS, if not the economy at large, then should we consider AI to be the black plague of our generation?

Creative destruction

Recent viral articles claimed how “everything is about to change” and “the efficiency gains alone will kill the economy.” New product launches, together with some helpful social media virality driven by those articles, resulted in significant tech sector selloffs.

Probably the closest way to describe what’s happening is creative destruction. Creative destruction, coined by Joseph Schumpeter, refers to the incessant, necessary, and often painful process where industrial innovation destroys old economic structures to create new ones. It is considered the fundamental engine of capitalism, driving progress by rendering outdated methods, products, and jobs obsolete to foster growth.

While we’ve seen some recent examples of it (digital media vs physical; e-commerce vs brick-and-mortar), the enterprise software industry has largely been spared. Cloud, the most valuable technology transformation prior to AI, took many years to scale and companies have been slowly transitioning applications and workflows to the hyperscalers. Even today, there is a motion of moving back from cloud, due to costs and desire for “digital sovereignty.”

When LLMs first started being integrated in enterprise applications, most companies found that they are poorly prepared to actually productize AI. After the last generation of developers spent their careers mostly working on connecting APIs and deploying existing frameworks, the high requirements of productizing an LLM showed that many companies lacked both the vision and technical talent to be successful. The opportunity led to the largest push of private capital into the new “AI-native” companies, which attracted many of the talented individuals currently working in large software vendors. Why stay at Salesforce, when you could raise $10M and work on problems you actually cared about, leveraging the best frontier models, as soon as they are available?

The results of this dynamic were not obvious at first. X, the main platform where the AI conversation was happening, was full of launch videos, but nothing useful on the real world value of these applications. Playing with the products of these new companies, the general feeling was often that “this is just an LLM wrapper.” This was a bit ironic, since the majority of SaaS software can at best be described as “an AWS wrapper.”

Things started to change towards the end of the year, as the models powering these applications clearly took a step up in function. This was already obvious to early adopters earlier in the year (if you happened to use a model like o3 Pro), but it took a while until LLM reasoning was being used by an early majority of users.

The most obvious change of tone was in the developer community, where the use of AI became strongly associated with Claude Code. Anthropic offering their own developer-focused product was something that most did not expect. Typically when an infrastructure provider (which the frontier labs were considered to be) has been able to establish themselves in the market, moving towards the application layer would be seen as risky, particularly since that meant they would now compete with their own customers.

The adoption of Anthropic models for coding in enterprise was so explosive (the company finished last year at $9B ARR, becoming one of the largest tech companies ever created), that the move towards vertical software felt very much justified for the management team. Anthropic did not stop here, launching Cowork and Security Reviews, two products that were built on the foundation of Claude Code and a logical extension of that motion. In enterprise, they also offered custom products for financial services and life sciences.

The reports of software dying are greatly exaggerated

The current trajectory of AI being a force of change in the software industry is both predictable and positive.

Yes, low value software will see significant margin compression and go out of business.

Yes, the majority of the value will accrue on the bottom of the stack, with cloud infrastructure software (data+AI, cybersecurity, developer tools) being the primary beneficiary.

Yes, most people currently working in the industry will have to actually start delivering outcomes, rather than behave as if they are going to an adult daycare.

Still, this will be a process where “human-in-the-loop” remains critical. Until we can access Artificial General Intelligence at production scale, the highest return from AI implementations would be in augmenting the top 20% of employees.

Scenarios where we don’t have AGI level intelligence, but AI agents have overtaken the productive economy do not make a lot of sense, partly due to efficiency (token price vs speed and quality of outcome), and partly due to in-group preferences. This might sound surprising, but outside of a small group of jaded social outcasts, most people would still prefer to build companies and work together with other humans. The difference is that the quality bar to be hired will be much higher.

Physical AI

One of the strangest glimpses into an alternative future can be seen on the battlefront in Ukraine. The past and the future have fused together, pushing hundreds of thousands into trench warfare reminiscent of World War I. This is where the parallels to the past end, as the majority of what war used to look like has been replaced with flying robots scouting, delivering supplies, transporting and ultimately killing the humans trying to survive.

The frontlines of eastern Ukraine are a window into a dangerous future, where autonomous weapons control every outcome. Drones fly with a fiber optic cable still attached to them in order to reduce interference, while AI-guided systems destroy any target that made the mistake of revealing itself. The humans hide, alone, often for months. Food and supplies arrive from the same type of machines that are hunting them.

Most cities, however small, are being covered with durable nets. If possible, they extend this protection to the road network. Humans walk around with drone detectors, shotguns and their own robot helpers, trying to navigate this new environment. Gathering in groups is incredibly risky and only done as a last resort.

This week, the executive leadership of Anthropic is trying to convince the Department of Defense that they should have the right to refuse the usage of Claude for autonomous weapons. The reporting on this is a bit patchy, but it has already led to the term “WarClaude” emerging into the tech community consciousness. If Anthropic and the DoD do not reach an agreement, Anthropic will lose its federal business (and likely face other consequences). On a long enough timeline, we should assume that all frontier models will end up being used in hardware aimed at killing other humans.

I’m not writing this to scare you, but rather highlight how unpredictable things can get, once physical AI becomes a real thing.

The vision of Tesla for Robotaxi ends up with humans ultimately not being allowed to drive, as traffic becomes a fully automated cluster of machines tasked with deliverables.

The vision of Amazon for their warehouses is for no human labour to be needed in order to get products from the factory to your doorstep.

The vision of many pharmaceutical companies is a future where they can rapidly discover new drugs that accelerate their time to market through AI simulations and verifications at each stage of a traditional clinical trial.

The current progress of AI in software is fascinating to observe, but it hardly constitutes a “fast takeoff.” Outside of enthusiasts gathering for meetups and wearing crab outfits, we are witnessing the creative destruction of legacy software. Still, humans are very firmly in control, even if wealth and influence shifts towards AI-native software companies.

Physical AI is much more difficult to predict. The infamous “AI 2027” report forecasted a negative scenario where we will be controlled, farmed and ultimately removed from existence. Others see physical AI as the ultimate accelerator for human success and the stepping stone to interplanetary species.

The first time that I negotiated with a large financial institution on a software deal, I was faced with a grumpy Scottish gentleman. He didn’t seem interested in the software we were selling or how things would work, but he kept repeating the same phrase, over and over again.

“Whatever happens, there should be no loss of value for the bank. No loss of value for the bank!”

Most of the stock market participants today seem to show the same mindset, overfocusing on perceived loss of value in software. That’s the squeaky bum talking. I firmly believe that software has never been as exciting or relevant as it is today. The difference is that in order to understand where value is accruing, you’ll have to push yourself harder than ever before.

Palo Alto Networks today is clearly one of the most successful cybersecurity companies operating in the space. If CrowdStrike did not exist, they would've been considered the defining security company of the AI pivot.

This is not the present that we live in, however.

The key takeaway

For tech sales and industry operators: Palo Alto Networks is making a classic "definite pessimist" move: they don't believe they can invent the future of security internally, so they're buying every possible outcome that might lead them to "the platform." The problem with this approach is that they are making too many leveraged bets at a time when execution risk is at its highest. The perceived weakness externally seems to be playing out on the inside as well, with restless teams, political games and overall negative sentiment. The platformization story clearly resonates with the largest enterprises signing $40-50M deals, but for the vast majority of PANW's field organization, that story is a pitch deck, not the reality on the ground. If you will drive a key account for Palo Alto Networks, your experience will likely be a positive one. The unfortunate reality is that this opportunity today is mostly reserved for the top 5% of accounts and several significant bets need to pay off in order for it to expand. Many seem to doubt this actually happening.

For investors and founders: The PANW thesis requires you to believe in three simultaneous outcomes: that platformization drives enough upsell to justify the acquisition costs, that the CyberArk integration goes smoothly despite the high execution and technical risk, and that Chronosphere can pivot from struggling pure play observability to integrated security analytics without losing its existing customer base. The 119% net retention among platformized customers looks bulletproof in a stable environment, but those $50M+ commitments also carry the most renegotiation leverage if CISO budgets tighten, and PANW's revenue concentration in top accounts cuts both ways. The bigger problem is obsolescence risk. The Anthropic AppSec announcement that triggered this week's selloff is a preview of what happens when AI native companies start building security capabilities from the ground up rather than bolting them onto legacy architectures. Many categories that were previously bought as "best in class" point solutions are quickly becoming expected features, and rather than rebuilding the platform of the future from first principles, PANW's technology is starting to look like the patchwork of poorly integrated tools it was meant to replace. Whoever executes on those problems from scratch has a legitimate shot at defining the next era of enterprise security.

Platforms in the age of AI

Nikesh Arora: We delivered a strong Q2 fueled by robust demand for cybersecurity and continued execution against our platformization strategy. This led to strong organic results in Q2 with NGS ARR up 28% and revenue growth of 15%, excluding the impact of recently closed Chronosphere. We saw broad-based strength across our products from SASE, software firewalls and XSIAM to our emerging leadership in AI security with Prisma AIRS. We paired this growth with improving profitability, achieving a 30% plus operating margin for the third consecutive quarter.

We're excited to head into the second half of the year, having closed both the CyberArk and Chronosphere acquisitions, and I want to extend a warm welcome to both teams. Both companies continue to deliver record numbers in their most recent quarters, and we look forward to building on the momentum as we hit the ground running on our integration plans. These investments are a direct response to the inflections we see taking shape in the market. And while it's still early, the initial feedback from our customers has been very encouraging. We believe we're now entering the next phase of AI adoption.

Large enterprises are moving beyond experimentation and beginning to integrate foundational models into real workflows. As AI becomes embedded in day-to-day work, the central question that organizations face is shifting from capability to control. That shift has meaningful implications on security.

As AI becomes more pervasive across the enterprise, it expands the attack surface area, more agents, more infrastructure, more machine-to-machine activity and new classes of risk that simply did not exist before. In that environment, security cannot sit on the sidelines. Despite the current sentiment about AI and software, we firmly believe that security is enabling layer that allows innovation to move forward safely and at scale.

And as AI agents become autonomous employees, the old security playbook is not just slow, it's obsolete. Security must operate in real-time at the critical control points where decisions are made across network, endpoint, cloud, browser and identity. This is where Palo Alto Networks operates. And as AI becomes more embedded across the enterprise, those control points are converging.

A fragmented defense of disparate products is no longer a viable strategy. The risk is simply too high when adversaries are moving at machine speed. Our latest Unit 42 research confirms this, end-to-end attacks are now 4 times faster than a year ago, and in nearly a quarter of the cases attackers were able to break in and exfiltrate data in under an hour.

The good news is that 90% of those breaches were preventable, caused by basic gaps in visibility and controls across multiple attack vectors. This is why we committed to our platformization strategy a few years ago.

A platformized approach built on a real-time data-driven model that gets smarter with scale is the only way to secure the modern enterprise, and our results continue to prove that out. In Q2, we delivered approximately 110 net new platformizations, a quarterly record outside of our seasonally strong Q4.

This brings our total platformizations count approximately 1,550, up 35%. The success of this strategy is also reflected in our best-in-class net retention rate amongst platformized customers, which stands at 119% with lowsingle digit churn. This proves that once customers adopt our platform, they not only stay but continue to invest more with us over time.

Unit 42 is PANW's managed security service that is used as an escalation point when companies need to contain significant breaches. Interestingly enough, according to their annual report, the majority of incidents are indeed correlated to identity breaches, which often go undetected due to a fragmented tool landscape; this has been the case for the last 5 years.

Coming back to Nikesh, I think that while he's been very directionally correct on platformization as an approach (before AI became the topic that it is today), the actual execution of the strategy poses some obvious technical challenges. The playing around here on "35% growth of platformizations" while landing at 15% YoY growth in actual dollar value is revealing.

Nikesh Arora: This momentum isn't accidental, it is a result of a deliberate flywheel motion we built. When we committed to our platformization strategy years ago, we're betting on a shift that has now become an industry standard. This approach allows us to not only solve today's problems, but also provides the foundation to address new ones as they emerge. It starts by providing multiple clear landing paths.

In network security, customers can begin with SASE, hardware or software firewalls, and now AI security with Prisma AIRS. In the SOC, they can rely on our Cortex platform via XDR, cloud security or directly onto XSIAM, from any starting point customer experience, the superior outcomes of an integrated platform which leads them to adopt more deeply across our ecosystem.

In a market changing this quickly, we believe our responsibility is to anticipate the next inflection and ensure our platform is ready. That philosophy guides our strategic investments, and the results give us the confidence to continue.

Our secure browser for example, was one such early investment that is now accelerating our SASE business with over 9 million licenses sold to-date. Similarly, in AI security, Prisma AIRS launched just a few quarters ago and already rapidly scaling with over 100 customers ending in Q2. This is the discipline we now plan to apply for the two large established markets poised for inflection, identity and observability.

If AI becomes a new interface for how work gets done, identity security will be required to create the permissions and boundaries that teams can trust. And as AI introduces unprecedented scale, observability is essential for building resilient systems that can operate reliably.

By bringing our platformization discipline to these new pillars, we believe we can deliver even greater value to our customers and solidify our role as a trusted partner to navigate the complex security and data challenges of AI era.

Most of the narrative around PANW currently is quite reliant on how they are positioning acquisitions, rather than internal innovation. This is typically a bit of a red flag, although in the case of Nikesh, he has a strong track record of integrating not just the company, but the actual team and senior leadership within his existing org structure. While CyberArk is an obvious acquisition that allows them to acquire an emerging market leader and a significant install base, with Chronosphere things are a little bit more shaky. The company was struggling to navigate the observability market and was one of the few players trying to make it as a pure-play vendor, at a time when many of those companies have established strong footholds in security analytics (Splunk, Datadog, Elastic).

Read more

The company just announced a massive $30B Series G raise (upsized from the $20B target),…

Read more