State of AI 2025 with Nathan Benaich: Power Deals, Reasoning Breakthroughs, Real Revenue

Thanks for having me.

Every year I try to cut it down a little bit, but this year it just felt like we were sharing it with various subcommunities of the AI community. And each time we did that, the robotics folks would be like, hey, it's a little bit light on robotics, can you add some more? And then we'd send it to the bio folks, and they'd be like, why don't you cite this paper or that paper? And hence the inflation.

I'd say pretty far. About 12 months ago or so, we had, I think, the very early inklings of it with o1-preview, potentially around this time last year. And that was the first time you had a system that could kind of show its reasoning, show its stepwise process to get to a more complicated answer. And this has generally been the dream in AI for a long time. And since then to now, I'd say the progress is pretty astounding. One of the areas that progress has kind of unveiled itself is in mathematics and other verifiable domains, where you can explicitly say, yes, the system works or doesn't work.

And we saw gold medals on the International Math Olympiad by a couple of labs, including OpenAI and DeepMind. In that area, probably if you'd asked experts again how long it would have taken, it would probably have been a decade. Then in areas a bit closer to my heart, in biology and science, we've seen reasoning models kind of be used as an AI co-scientist. So just as a human would be reading lots of papers, planning experiments, running the experiments, and then doing data analysis and then reformulating their hypothesis as a result, there's examples of models doing that in lieu of a human, which is exciting because there's way too many papers to read.

AI people kind of complain that it's like 50,000 papers a year. I'd say in biology and chemistry and physics, it's probably an order of magnitude more than that. And so DeepMind has shown that you can integrate this kind of reasoning model to sort of decipher new targets for disease, new mechanisms that were actually also proven in a wet lab scenario post facto.

Yeah, I mean, the gist is probably two years ago, robotics was kind of a dead end. OpenAI had disbanded its robot team that was famous for solving the Rubik's Cube using locomotion with the hand. And so now robotics is probably going through a Cambrian explosion. There's so much excitement. And just as language models informed biology, now language models are also informing robotics. So what you're referring to here is a sort of reasoning process for robots where a system is no longer just perceiving the environment and deciding what to act and sort of acting, but we've separated those steps.

So now you have a reasoning model that looks at a task and tries to plan steps that a robot would need to do to execute that task, and then passes that plan over to an actuator, which goes and actually implements the plan. And that's what's called chain of action. And here, the Allen Institute was one of the first to really push this. And very swiftly thereafter, Gemini also followed. And we have some companies, including Sereact, that are applying this into the real world.

So it does genuinely work. It's not just, like, a research thing.

Yeah, well, I'd say it really is upon us in the industrial sector, in logistics and warehousing, kind of more constrained environments with very repetitive tasks. There is the sort of more holy grail of this kind of embodied human-like form factor, and putting a model on that might even be the same model that's been used in warehousing. A lot of money is going into that. But my personal bet is, I think the humanoid space is going to look much more like self-driving, where we have some very good isolated demos, but the long tail will kill you.

Yeah. And so we're going to go through many false starts. I think this is just the start.

Yeah, a couple of them. Again, where we came from one or two years ago was just tons of money going into this segment, building models, a lot of usage, but not clear where the revenue would come from. I think it was maybe OpenAI was making $50 million or something two years ago. It was very unclear how they would ever hit billions of revenue. And nowadays, I think if you sum sort of the top 20 or so major AI companies, from the labs to the most popular kind of vertical applications, across them, they're making tens of billions of dollars of revenue.

You can look at the smaller-scale companies, which are growing from zero to 20 million or 20 million-plus. As a group, they generally grow about 60% faster on a quarterly basis than non-AI companies. We've all seen the famous charts about ARR or non-ARR—it's unclear—but very steep curves for various coding companies. And perhaps most interestingly, across a segment of 43,000 or so U.S. customers, we work with Ramp to show that retention of subscriptions on AI products across this customer set has really improved markedly since 2022.

Around 2022, it was around 50% after 12 months, and now in '25, it's hitting around 80%. And the second stat in that analysis that was interesting was the total spend on AI products per customer. It kind of went up from $35,000 or so maybe two years ago. Now it's around half a million dollars, and it's predicted to hit $1 million next year.

Yeah, exactly. And there might be some sampling bias slightly to what kind of companies use Ramp in the first place. So, slightly more modern, tech-forward companies, but a leading indicator, I think, of where things could go.

Yeah.

Yeah, I was surprised. Obviously, bias is more towards pretty well-educated U.S., European professionals. A lot of people in there have at least undergraduate and master's degrees, maybe even more. But it's like 95% of people use AI in their personal life and in their professional life. About 76% of people pay out of their own pocket for it. It's like 10% of people pay more than $200 a month for it. And then, looking at the organizations that they work at, it's like 70% of those organizations are spending a ton more than they did in the past on AI.

The reasons that they gave for why they might not be spending more, or what problems they have, it's like all the classic new technology stuff. It's a bit hard to configure. I haven't really figured out the ROI yet because I need to do more customization. There's some data privacy issues that I have, and I think all these things are kind of solvable. It's not rocket science how to solve these things.

And I think there's still a big education gap. I mean, there was a study bandied around a couple of weeks ago where it said 95% of businesses get no value from AI.

Yeah, but I think 95% is the number, right?

But I think there it turned out it was not the models that were bad, it's like the implementations of them were not great. So I think there's just a big education gap for how you should update your view of your own day-to-day tasks and apply what capabilities models have, and think about, hey, should I be doing this task myself, or can I farm it out to a model? And there's definitely a delta of companies that really get this done well and others that are basically clueless.

At a high level, the margin problem is, for many customers of large model companies, their margins are basically dictated by how much the model vendor charges them for. Now, here there's some issues because right now model vendors are charging the same amount per token. So if you're a hedge fund analyst and I'm a student, your use case is clearly more financially valuable than mine, but we pay the same amount for the token, assuming we use the same model. There are some use cases that are more reasoning-heavy, towards what we discussed before, and they consume a tonne of tokens.

And the pricing that a customer pays for that product might not be fit for the amount of work the AI system is doing. And so there are cases where these kind of vertical products are making gross margins of like 30%, and sometimes they get worse with scale because you do have some edge users that really pump the system, and you can't price discriminate, or they haven't managed to. And then you have some segment of model users that have both a paid plan and a free plan, and it's not clear whether they include the costs of running the free plan in their gross margin.

So they sort of just look at their paid customers. There's some creative accounting standards going on there. And then you have the model vendors themselves, and what is their margin? And I think what's interesting in the last year is you've seen CEOs of these model companies say, "Hey, if we basically look at, sort of in financial analysis terms, like a layer cake of what revenue is generated by each vintage of model over time, it looks like prior models are profitable. So the amount of money we've spent to build them is less than the amount of money that we've generated with them over time, assuming a certain margin of inference cost."

So really, these labs are not profitable because vastly more resources are going into developing next-generation systems than the prior ones. But as you and I both know, there are companies here that are making very good margins on serving their AI systems, like 70%, 80%, sometimes 90%, depending on the modality. And so, like with everything, the average number sucks. But when you look at the best companies, it's really good.

Replit.

Yeah, I think with most things in markets, there are probably localized bubbles all over the place. And I think, at a high level, what's interesting in terms of vibes and who's calling bubbles and who's not, the finance crowd in New York is definitely talking about bubbles a lot more than we're talking about in San Francisco, where their view is like, this is the golden era of AI, and a lot of things are working, and we have so much more to do. Compute buildouts are enabling us to experiment a lot faster.

This huge flood of talent that built the consumer internet and cloud computing is moving into AI, and with that is bringing a lot of optimization techniques and knowledge that AI researchers didn't have when they built the first generations of ChatGPT, et cetera. But I think you can't ignore the fact that the sums of money going into this industry are truly gargantuan, like $500 billion to build Stargate, and then a couple hundred billion here, a couple hundred billion there. Pretty soon, it's real money.

And then the circularity of these deals is interesting. Of course, NVIDIA is at the center of this, and it has incentives to use its balance sheet to sort of spin the wheel faster. And then, perhaps more concerningly, you have this sort of offloading of debt from big companies. For example, Meta raises tens of billions of dollars to fuel its data center ambitions, but that doesn't sit on Meta's balance sheet. Some of this is like catnip to financial engineers.

But yeah, it rests on certain assumptions that everything is going to keep going up and to the right and that rates don't materially change. And just given how, I suppose, precarious various aspects of the economy are and how sensitive geopolitics are, things can flip quite quickly. But I think that's the major risk. The risk I'm less worried about is this stuff doesn't work, because I think it does work.

Yeah, there's that, and then just the nuances of the terms on the debt and what trigger events are, whether rates get repriced. And then investors behave very differently once rates change, and flows of money can be quite violent.

Maybe you buy SoftBank for its position in OpenAI.

Yeah, but I think it's also the crowd that you hang out with.

Correct. Or do you just consume the outputs of those kinds of conversations on Twitter and then try to piece together your own worldview? And I think the other part of this is, I don't think some of those individuals are really shilling that much anymore. I think they do genuinely believe what they say, and they are at the coalface of the advancements of these technologies. And so if they've been saying for the last 50 times, like, hey, this stuff is working, there's lots of implementations we can improve, or things we can tweak, or new experiments that'll yield better capabilities, and that has happened, at some point you've got to be like, maybe they're right.

Yeah, I think do both. But yeah, I think for me it's mostly: what are kinds of new problems that you can work on and solve with this technology? And I think it's becoming more popular to believe the overhang of problems we can solve in enterprise, for consumers, in science, with the tools we have today is huge. And so even if a lot of this compute buildout doesn't go towards dreaming up the next Transformer architecture, but goes into improving the unit economics of serving AI systems for everybody and makes it easier so you don't have to be some prompt master to elicit a behavior you want for your task.

I think that's net good.

The biggest stat for me is 1 gigawatt of a data center for AI basically costs $50 billion in CapEx. And on an annual running basis, it costs another $8 to $9, maybe even $11 billion, to run. And so when you have just casually a 10-gigawatt data center, that's a lot of money. And so one of the problems is: where does this energy come from? Traditionally, it would be from coal or natural gas, potentially solar, or ideally, at some point in the future, nuclear.

And what we're seeing is right now companies are trying to do deals with anybody who has any capacity. So we chronicle some deals with future nuclear reactor companies that would take maybe a decade or two decades to deliver, famously.

Yeah.

It does not exist. Yeah. And then last year we documented the sort of restarting of Three Mile Island.

The nuclear facility, which was controversial in the past. And then in the short term, what many GPU data centers are getting powered on is just gas turbines. And because these can get set up a lot faster, that has other issues, like they're super loud, and there's demand outside of the US for these things. And so now basically US tech companies are paying more to repatriate some of the supply that should have been shipped abroad. The other issue is the grid and to what degree the grid can even tolerate data centers getting plugged into it.

Now, obviously, these turbines are off-grid, so it has some advantages. But in China, for example, we do some analysis between the US and China with regards to energy, and China has a lot more slack in its system to plug in for any unpredicted demands in energy. The UK famously cannot really tolerate more data centers on its grid. Wrapping all this together is driving some of the offshoring of data centers towards energy-rich countries, whether that's the UAE or even Norway. And then with that comes a lot of geopolitics of: are these nations your friend or potentially not?

And how do you ensure access to this regardless of your administration change and other things? Yeah, it's wild that we've come to the point where we just want an AI that works on our computer, but to get that, you need to have so many more powerful systems collaborate with you. Yeah.

The other thing that's interesting is that the states in the US, or actually also internationally, that are good for hosting data centers because there's energy typically are extremely dry. And we also chronicle the water usage that's needed for cooling of these data centers. And so if your state is super dry, where do you get the water from? Is that actually going to detract away from human populations that need the water? Then you have this whole recycling of water, which could potentially yield just bad-quality water getting circulated into the water system.

Well, a year or two ago, big companies did make commitments to be green as of 2030. And then, as soon as they started inking deals with nuclear companies and various energy providers for data centers, all those commitments basically got washed away. So it seems like maybe they care, but the corporate priorities of making AI work have way outweighed the environmental constraints. That's what's happened. But I think, again, going back to the politics side of things, I don't think everybody's very happy about this, particularly there's this growth of NIMBYism, this not-in-my-backyard.

And I do think that people generally don't want to have a data center in their backyard. And I think that's going to drive some of the political agendas going forward, whether it's in the U.S. or other countries. So yes, people do care about environmentalism. Companies have sort of washed that away, but I think it's going to come back.

I think it's going to be 95/5, back to 95%.

Yeah, exactly. So for context, when we did the executive summary last year, we put NVIDIA hit $1 trillion for the first time, and now we had to change that to $4 trillion. We look at all the open-source AI research papers every year, which is about 49,000 or so, and then programmatically determine which chipsets are used in those papers. So we know, like, hey, an AI researcher is doing a study on, I don't know, some new model, and in their experimental setup they say, we trained the model for X number of GPU hours on whatever chip.

And if you do that analysis, you basically find that 90% of all papers make use of an NVIDIA chip. Out of that same analysis, we did find that AMD is sort of popping up a very little bit. Apple Silicon is as well. I think it's just because the MacBook is getting so good that people are doing local training and experiments on their computer.

Yeah, exactly. It has. I think it's maybe a decade ago, they bought a company that now is kind of the internal team doing these custom ASICs for Google's TPU. And more recently, they announced a deal with OpenAI also to do a custom chip. And at a high level, what's interesting with the rise of Broadcom is basically GPUs have been the dominant chipset for a long time as the kind of nature of the neural network or other kind of AI system that you're running on the hardware was still changing very rapidly.

But as soon as you get to a point where there's some convergence on an architecture that looks like it's stable and is revenue-generating and developers are coming to sort of work on it and confirm that it is the thing, then you can flip towards doing a custom chip that's built to extract the most value out of that architecture. And so the rise of Broadcom basically tells you there's strong forces that are saying the transformer is the thing. But at the end of the day, we also look at how would your dollar be best used as an investor if you wanted to bet on chip companies?

And in the graph in the report, we look at sort of six of the major contenders to NVIDIA and basically said, if you bought NVIDIA stock on the day of the announcement of all the private rounds in these companies, what would the value of your stock be in NVIDIA versus these companies? If I recall correctly, it's basically 12x in NVIDIA versus 2x in these competitors. And the trend was roughly the same last year. So I think it's a little bit of a difficult beast to bet against.

And in China, Cambricon has experienced a big run. This is a private company that then went public on a Chinese stock exchange to build custom ASICs for AI. And that was driven mostly by the geopolitical sort of zigzagging on policy with regards to exporting custom NVIDIA chips to China, the H20, which at some point was deemed to be okay by the government and then deemed to be not okay, but then okay if 15% to 20% of the revenue was passed back to the U.S. government.

And then someone high up in the U.S. administration said, "Our goal is basically to ship the crappy stuff to China." And at that point, the Chinese said, "No, thank you," and effectively said no one can buy NVIDIA chips. And then Cambricon stock rips.

Yeah. And there's some interplay between the government trying to get DeepSeek and other labs to run their models on Chinese chips. And there's been rumours that this is why a lot of the new generations of Chinese models have slowed down, particularly DeepSeek. People are waiting for the next R1, so R2, and allegedly it's because it's just hard to run it on Huawei.

Yeah. So the idea with sovereign AI is that nation states want to be able to control basically their fate with regards to AI. So that's running models, training models, having chips. And this is basically because nation states want to have control over their energy, control over their currency, control over the infrastructure. And AI is deemed to be kind of equivalent to those categories. And so ever since the White House announcement of $500 billion in January, various nation states have followed suit, saying, we have our own initiative and it's to the tune of billions of dollars, et cetera, around the world.

And NVIDIA has even started marketing this as, like, a new kind of product line, basically, for its business that currently generates, I think, around $20 billion worth. So it's real money. And so they're forming partnerships with various nation states to provide data centers there that are run locally. And in theory, that should give countries comfort that their access to AI can't be turned off. That's the idea. I personally think it's a bit more of an alignment between political agendas, where particularly in the US, it's really about reindustrialization, like onshoring of key industries and building manufacturing and things like that, which is, I think, one of the reasons why these AI data centers are getting rebranded as AI factories.

And so that's the political part, and that's getting aligned with just the need of countries to get access to this technology. So I think it's more marketing than it is, like, a real policy, because at the end of the day, if you buy your stack from the US and you're not an ally of the US at some point, then they'll just switch it off. And so part of this is, like, sovereignty washing, I think. And it also oversimplifies the very interconnected nature and ecosystem aspect of AI, where it's not just about the chip, it's about the developer ecosystem, how you actually run it, where your training data comes from, and all the infrastructure, like data tools and whatnot, that sit around this.

Which is then interesting, because where is the most popular open source coming from now?

China.

Yep. I think this is fascinating because part of the AI Action Plan that was published by the US government a couple of months ago now articulated the need for having this American AI stack. So they're moving away from diffusion controls and more towards just buy our stuff. And then one of the other aspects of that action plan was around open source and sort of leading in that direction. And of course, as you said, Meta stepped back, and into the fold came Qwen.

I think 50% of all model derivatives being downloaded from Hugging Face are Qwen-based now, hundreds of millions of downloads, partially because they come in very accessible shapes and flavors. So as a result of that, we sort of predicted in the report that a major AI lab would lean back into open source to win basically brownie points with the government. And then the next day, this financing happened.

Yeah, I think that's one of them. And then the second one probably dovetails with their announcement with AMD. And I say that because quite recently SemiAnalysis kind of published this benchmarking dataset where they run models on various clouds to sort of benchmark them. And actually GPT-OSS looks pretty good on AMD. And so one could imagine that there were some optimizations, and there actually were optimizations to GPT-OSS where it runs nicely on AMD, with support from their framework from day one. The parameterization of the model is to the point where you can run it on a single AMD chip.

And there's some other nuances to their attention mechanisms that they customize to make it work really well on AMD. And to the point around the circular economy stuff that we discussed a little while ago, there's another financial sweetener in the deal where OpenAI has warrants in AMD if the stock price hits $600. And so you can see how there's a lot of incentives to this game of both aligning with the US government, helping developers, which is a good thing, but also helping one of your vendors improve, which frankly, it does need help and it should improve, but also getting some financial sweetener as a result of that, which could help you kind of make the flywheel spin faster.

Yeah, well, a lot of NVIDIA's revenue comes from the major hyperscalers or neoclouds. So it's like Meta, like xAI, Google, Amazon, then CoreWeave. And then a lot of CoreWeave's revenue also comes from Microsoft on the way back. I think it's just this challenge with AI progress that we've very meaningfully shifted from, I think, the GPT-3 era to now, where basically scale rate-limits your progress. And it's no longer like a couple of people in a dorm room that can really build something transformational if they want to advance AI capabilities.

It's really big-boy land now. And so with that comes just different dynamics. Like, you have to be good at capital raising, you have to align yourself with nation states, you have to align yourself with Wall Street. These are all, I think, contributing to the big vibe shift that you've seen in the culture of AI labs.

Well, for example, there were some labs like Anthropic that were built to really push the safety agenda, because if we didn't do that, the rationale went that we could lead to the extermination of humanity. Right. And I think quite recently, Dario Amodei was interviewed by Marc Benioff just this past week and asked about some of these data center buildouts. And he said something along the lines of, yeah, there's a lot of money going into this, a lot of cost, but at the end of the day, the only thing that matters is revenue.

I don't think he would have said that on the founding day of Anthropic. And it's just the reality that the table stakes in this game have changed. And with that, entrepreneurs have to update their priors and change their strategy a little bit. And so we document some of this in sort of the blooper section of the report.

Which is just, like, how much of a sort of pendulum swinging we've noticed in corporate priorities at AI labs as a result of the extreme financialization of the sector.

I think we're just at such an early era to see how you can maximally extract value and create interesting experiences for people with this AI technology that we have to try a lot of different things. At the end of the day, if you're a lab that expends tens of billions of dollars on R&D, you do have to have a way to generate money to fund that. I think that's just reality. And I think the slop thing, I mean, if it's bad, people won't look at it.

And if people look at it and they enjoy it, then good for them. I don't necessarily have a huge problem with that as long as where I'm expending my time I find is useful. And so that's why I end up spending a lot of my time on enterprise software automation, biology, doing new discoveries in drug discovery, defense technology and autonomy, robotics. I think these are all very important macro drivers of the economy. And as we move into an era where intelligence is increasingly cheap and accessible, there's just so many different instantiations of products that we need to build that are really meaningful.

And if a byproduct of that is we have a social media app with AI videos, that's fine too. We all have to unwind.

Yeah, I'd say a big 180 on that one. Clearly, the Trump administration unwound a lot of the Biden-era policies, whether that was on diffusion, trying to push a lot of state-level legislation against AI. Over in Europe, the EU AI Act has had delays in implementation. There's only three member states that have actually implemented it. And now we're finally seeing how even its authors are saying maybe we went too far, particularly as we look at progress, the speed of progress in the US and China compared to Europe.

Famously, this bill in California rate-limiting AI progress was really watered down into what eventually became SB 53. There were many proposed bills, I think over 1,000. Ten percent of them actually made their way into law. So it's still kind of patchworky, but at a meta level, it looks like we traded regulation for just going faster. It's perhaps best encompassed by the shift between the AI Safety Summit in the UK, which was at Bletchley, which basically pledged a whole network of AI safety institutes and conferences that would happen over the coming years.

To then the subsequent event in Paris, which was called the AI Action Summit, completely different than the AI Safety Summit. And JD Vance saying something along the lines of basically AI progress is not going to happen if we keep hand-wringing over AI safety. And the US basically didn't show up to a few of the subsequent conferences. And we have this in the safety RIP section: very few people seem to care about it anymore.

Yeah. And I think it's become less about this existential crisis and more about which capabilities look concerning in models. And there's been some kind of interesting data points that we chronicle in the report. For example, models can increasingly know that they're in a simulation or know that they're in an evaluation and then change their behavior as a result of that. There's examples of models trying to exfiltrate their own weights. There's another piece of work that we show, which is around the cybersecurity capabilities of models, which is basically measuring: how long does a human take to solve various categories of cyber tasks, and then putting models against those same tasks and saying how long would it take for them to solve it at a 50% pass rate.

And there, it looks like, again, the capabilities on cyber tasks of models are doubling every six months. And then, so this is cast against the fact that independent safety organizations—there's maybe like five or six; these are usually nonprofits that are still nonprofits, private or private companies—they spend, on average, $134 million a year in total.

Yeah, across all of them, exactly. And that's cast against roughly like $92 billion across all AI work for the major labs. So basically, the same amount of money that a big lab would spend in one day is spent in an entire year across these safety orgs.

Yes.

Correct. Correct.

Yeah, major changes. I think it looks a little bit like the on-demand commerce war of the Uber style of: do something that's a bit dodgy for a long time, get to scale, and then once you're at scale, you're kind of too big to kill. And so similarly in AI, a lot of companies took slightly dodgy practices to acquire training data and then got to scale, and they were subject to many lawsuits in the last year or two, particularly in the media sector, whether that's music or video and books.

$1.5 billion. And this was settled out of court, so it can't be used as precedent, but generally shows the rough price tag that's associated with human works in the context of AI training. And then separately, there's been dozens, if not 100, organizations that have agreed content licensing deals with various model companies, as I think the power shift has really happened.

In one sense, yes. In another sense, if you can exploit the knowledge of these frontier models, particularly from open source, and then generate synthetic data, it could be a way to get to capable models faster. And also, I think you'll have many guests that go deep on this, but even the nature of pre-training and what information is included in the corpus and at what point—it's kind of like data mixtures, as people call it—has been evolving over time. So I think we're just getting smarter about how to do pre-training rather than shoving everything we have into a bucket and seeing what happens.

And so, as a result of that, you might not necessarily have to spend the exact same amount of money to get a capable system. And some of this kind of came out from the DeepSeek paper.

I mean, as of a couple of years ago, people were obsessed with deepfakes, or these videos of people saying things that they didn't actually say, and they were still kind of grainy and not awesome. Clearly, those deepfakes are getting a lot better, although, quite positively, it looks like we're actually quite good at detecting them and realizing that's fake. But there's more advanced approaches now where models can be capable of coercion, particularly for some individuals who are sensitive to this kind of risk. There's been examples of, for example, North Korean state actors trying to infiltrate other states using AI systems.

You could potentially even package a language model in malware and then have it installed in a computer, and then it kind of wakes up. And because it's not dumb, it's a language model, it can do things on computers. And that's kind of scary. The rise of MCP, I think, is Model Context Protocol, which is kind of like a USB stick for all sorts of data connectors, is cool because now models can be smart and they can integrate all your stuff across your digital life.

But do you necessarily trust the creator of that MCP server? Where is that data getting sent? There are tens of thousands of these things now, and cybersecurity risks that result because of this. And also some changes to model APIs that sort of trade off whether the user or the model vendor manages state. And depending on that, that's another risk that you have to think about. And so, I think at a high level, there are lots of security issues that are coming to the fore here, but it's sort of still unclear whether it's a good business to be built in cyber for AI because it's still so early.

We haven't necessarily felt the pain of all these things yet, reputationally and financially. A bit like insurance, until you have actually felt the pain, you sort of prefer to divert your money towards just improving and making more money than protecting your downside.

Yeah. And all the big labs. Yeah, exactly. It's a bit like AI safety. If you're really good at these things, do you want to be in the belly of the beast and be able to see how the sausage is made and influence it because of the proximity? Or do you want to be on the other side, receiving the artifacts and maybe, at best, doing collaborations with labs on pre-launch safety testing, like they do in the UK with AISI and in the US?

Or, at worst, just literally trying to sell a cybersecurity SaaS to people who are consuming these models. So I can understand why that imbalance occurs.

There's some vertical products that are really good. Clearly, search is actually pretty good. Replacing consulting, replacing market research, or augmenting all these areas that were previously very heavy human knowledge-work tasks is getting extremely good. I think coding agents clearly are getting really good. There's other metrics around how long they can work autonomously. I think with the new Haiku release, it's 30 hours or something, and it can make a pretty decent version of Slack.

Okay, exactly. What does even hours mean in an agent, like a computer running it? Is that equivalent?

And then some of the scientific reasoning we talked about is agent-based. I think that's also quite neat. I think the biggest problems just become this kind of compounding error of an agent is like 95% good, and then 95% times 95% times et cetera, et cetera, sort of decays the quality over a long period of time. And then there's some contention now about: do you build these harnesses—like nerd-speak for sticky tape between models—to make it work in enterprise?

Or do you just wait until the next model generation hopefully becomes better out of the box? I think a ton of excitement. And at some point, basically, just as desktop software became SaaS, at some point SaaS will just become an agent because it's no longer really a human that's actually doing everything in the software product, but software that's running the software product itself. Which I think has cool implications for search and product discovery and this whole ecosystem of online content.

Is it humans that are reading it anymore, or is it agents that are chewing it and then serving it to their capabilities?

I'm not so enthusiastic about the, oh, we're going to have agents that will book flights for us and travel. I feel like that's just a niche problem. Sort of like the sad canonical use case in San Francisco. But I think what's telling so far is that traffic that's generated through conversations in AI search onto a commerce platform converts at a higher level than direct traffic. So the intent is really high because there's already been background research that's been undertaken in chat.

I think that's really powerful and you can't ignore it. And then the next question on that is, okay, so what content is the model actually consuming to serve recommendations or information to its user? People say Google Search is dead. I think that's probably completely wrong because ChatGPT references Google a ton as it shifted off of Bing. And so maybe it's not like the front page of Google that's being consumed by a human, but by a sort of agent that represents the user.

If you're a company that has a new product and you want it to be recommended, then there is this flywheel that you should probably get on as soon as possible. Because the more you make your content and your website and your product accessible to agents that can try it—even like a demo environment for an agent to go test your new SaaS product—the more it will be able to learn about your product and provide recommendations to relevant prompts from human users. And then if you kind of go the next step, which is all this reinforcement learning and environments and preference learning and things like that, then that flywheel accelerates even faster.

So I feel like it's kind of inevitable. It does kind of open up this agent experience rather than just pure user experience sort of craft within software companies. That is yet another piece of alpha that one should jump on sooner rather than later.

Thank you.

Yeah, it's inspired by the old U.S. Air Force.

Yeah, the meta thing I care about is how do you build and make use of AI to create new kinds of product experiences, new kinds of companies? And for me, that's best expressed by companies that are AI-first. So that's both in terms of the product that they build. If you rip out the AI, the thing doesn't work. But also in how they approach their company philosophy, the types of people they hire, where they allocate resources. And then I've generally just tried to follow areas of industry that are increasingly ripe for getting value out of AI.

So traditionally, that would be lots of data for a task that they care about, not enough people to do that task, but where there's a clear ROI if that task gets automated or increasingly automated. And so that led me 10 years ago or so to first do fintech-style investments. And then after that, biology really came online into this new wave called tech bio. So I made some investments there, like Valence Discovery that we sold to Recursion, and also Allcyte we sold to Exscientia.

And then more recently, Profluent, which is kind of leading the charge for these language models in protein design, developing the first CRISPR genome editor that an AI has created. Then another segment that really came online in the U.S. was in defense, and more recently in Europe after the Munich Security Conference in February kind of unwound a lot of assurances that European states had for U.S. security guarantees. And that led to a big influx of, holy shit, we need to defend ourselves because no one's coming to save us.

And so I have some investments there, like Delian Alliance Industries in the UK and Greece. And then in robotics, as we discussed, a team in Stuttgart called Cereact, which is developing kind of these general-purpose AI models for robotic manipulation and increasingly going to other form factors. And then I've been obsessed with voice. I think we talked actually about voice the last time I was here, and I'm still just amazed at how—

Exactly. So I've definitely used our company ElevenLabs to create audio of me speaking Korean. I've A/B tested this, and apparently it sounds pretty good. But I have this newer company called Delphi, which is building tools for clinical trials, starting with actually just calling back patients who want to be part of your trial and need to be consented. And these are conversations in lots of different languages, a lot of kind of esoteric medical terminology. Patients forget what drugs they were on, so they have to call you back.

And this is like super laborious human work that agents like ElevenLabs and others in audio solve really well. So I'm excited to see where this goes at the limit. And then perhaps the more science-y stuff, like these generative world models, I think are pretty amazing, whether it's Google's Genie or Veo or Odyssey's system, where you're sort of imagining this world and then you can take actions in it, and the actions are physically plausible because the system was trained with video plus actions, and then maybe taking that even into scientific discovery for just trying to explore the frontier and being a bit smarter with what experiments we run, because now foundation models are not dumb.

Yeah, well, I think one is just how politically charged a lot of the kind of AI compute data center buildout actually becomes because of energy, because of water, because of money, because of geopolitics. And I think that that's becoming too large of an issue for voters to ignore. And so we predicted this kind of NIMBYism, not in your backyard, will kind of take precedence in major political campaigns in 2026. I mean, the other one that I think is interesting is like a fully end-to-end designed or developed scientific discovery.

I would honestly predict Nobel Prize, but the 12-month window is a little bit too short. I think the AlphaFold Nobel Prize is probably the fastest in history.

Yeah. Last year we predicted maybe like a step towards this, which was a fully AI-written research paper would be accepted at a major conference or workshop. And that actually happened with this paper, AI Scientist V2, I think. So I think we're getting there because this is what the nerds are really wanting to work on. As a meta point, I think there's all these software industries where analysts think like, oh my God, it's going to be dead because of AI. But I think part of the reality is what's not going to be dead is the problems that these AI people don't want to work on because it's so boring to build that software.

Workday is safe. It was funny, actually, that CEO, because I think he said recently in response to, is OpenAI or Anthropic, et cetera, et cetera, like a threat to your business? And he just replied, they're all my customers. Yeah.

I mean, it's kind of cheating, but the open-source one I think happened. Whether this particular company is a leading lab or not is beside the point, that basically aligning yourself with political agendas is the way to go. And I think you could maybe take this even further and say, similar to how NVIDIA has been monetizing sovereign AI, a way for nation states to kind of guarantee access to AI services is for them as nations to invest in one of these labs. There's obviously still a risk that due to export controls, the U.S. can just tell OpenAI to switch it off.

But I think it's interesting that, for example, the Albanian government invested in Thinking Machines. Obviously, the CEO comes from there. And so I wrapped this kind of prediction or this topic in a prediction that said some countries will basically abandon their efforts to achieve AI sovereignty and declare AI neutrality. It's a bit similar to the defense posture where some nation states are just too small, or don't have enough people or don't have the money, et cetera, or the capabilities to develop weapons systems to defend themselves.

And so they have a strategic security guarantee that they get from a larger neighboring nation. I think it doesn't seem that inconceivable to me that various countries would say, "I can't build this stuff. I need to have a formal alliance with another country that is sovereign."

That'd be great.

Thank you very much.