A Critical Time for AI

0:04 What a great time for this conversation because it's the most critical time for the AI story as an investor, thinking about macro, thinking about politics, and thinking about what we're going to do with Bridgewwater. We are at the most critical time. So, it's really important to be thinking and paying attention and having good insight on what's going on as AI has become sort of half of everything. If you take my job as running pure alpha, it's half the conversation. It's more important than the fed etc. And and if you think about running bridgewater is most important. And of course in i labs, this is all we do and think about. So as we lay out these perspectives, I want to talk about how we do it. Of course, we're thinking as we always have for 50 years about the impact of AI from a perspective of productivity, inflation, understanding the macro and market impacts of that, but that's informed by the fact that we're also practitioners.

1:00 We're in the lab working with these tools at a cutting edge level all the time, which helps really understand what's next and how things are going to play out. And it's kind of those two pieces connected that drive what we do. And and so the insights that I'm going to share today are related to both both sides of that. And let me start with us as investors and and looking at the world and again this is fed by what we know as practitioners as well and and as the relationships we've built in the labs and with the critical people. so when I dig into here the state of AI today the basic picture is we're at this critical moment because scaling laws continue to hold and the models have gotten extremely capable particularly with a tremendous amount of depth in certain areas well superhuman and more dangerous such that things are going to change fast and they're going to change at a critical time. A regulation is needed and will come. I'll discuss that in more detail. The environmentals the environments meaningfully shift in the sense that we are at a critical time for the ecosystem to raise capital and we're at a more dangerous time to do that because Frontier Labs in our view revenue growth is going to slow to some degree. is you went through the Claude Claude code moment that got this huge spurt in revenue growth. And while cloud code continues to get better, coding continues to get better, the breath of adoption not quite keeping up with the depth of adoption, which means revenue growth rates in our view are likely slowing. On top of that, regulation already in place, starting to get in place to slow how fast better models can come out and increasing competition from open weighted near frontier models going to hit revenue growth to some degree.

2:53 Not that big a deal except for that's happening at the time when the labs are going to need capital potentially IPO slowing revenue growth into a massive need of capital at a time where regulation will be looming is why we worry there's a risk of an air pocket that's starting to play out in markets. So let's get into what's going on. Most importantly the power is unbelievable. The scaling continues. As you well know, this is one of our favorite measures, the EPOP capabilities index, which is measuring capabilities across a bunch of fields.

3:30 If you take the capabilities on compute and stuff, they're off the chart. We can't even measure them anymore because we're so far past human level in so many ways. But a more general measure of breath shows that the models keep getting more powerful. That's what we're seeing in the lab. And you're seeing breakthroughs where original thinking by these models are happening. You're getting new discoveries in math like nearly every day. some major problems that humans had not been able to solve have been solved with little or no guidance to AI. A big deal although when you look at those there is a sort a certain type to them. The broad creativity that will change the physics of the world and so on are not quite there. and and so some more needs to happen. But basically looking at this, you continue on track, which means it you're probably going to push to the next level of investment. Of course, to continue the scaling, the amount you of compute you need continues to grow exponentially and is a massive challenge from a funding perspective, but but the progress continues and recursive self-improvement is kicking in. So for all the negatives for the labs I hit on, I do think they're on the forefront of the intelligence. They are able to use their models to train future models. They are able to use their harnesses to code a lot of the work that they're doing. They're probably the best use cases for AI is happening in the AI labs and that has the potential to really kick in a recursive self-improvement that keeps them ahead in terms of intelligence in an important way. Now the flip side of that is the danger.

5:12 getting stronger also means you're stronger in hacking. You're stronger in biological weapons. You're stronger in many things. And and it's extremely dangerous if you haven't paid that much attention to the open AI internal model escaping it sandbox and hacking into hugging face. You should you change the players a little bit. I if it didn't hack hugging face you hacked something else, a government, a foreign government, etc. A much bigger problem. and it could have the fact that OpenAI had been struggling to keep its model in the sandbox illustrates the challenge that it is and the need essentially that given how dangerous these models could be that they could go out and break the law and potentially create geopolitical problems. The you will need to see regulation not just of models but of the environment in labs and the environment of people that are are generating these things. It's a big deal and it's coming and while it's complicated because within the US there's obviously a lot of people pushing to allow for very low regulation because of the race with China in the end the situation so dangerous that those things in my view are going to change and probably quickly particularly because politically AI is so incredibly unpopular and so this stuff is coming together in a way that's going to lead to big changes.

6:30 So speaking quickly about the revenue growth of course part of the ability to raise the valuation of these relatively new companies open AI and anthropic to a trillion dollars to some of the biggest companies in the world has been on the back of what is unprecedented revenue growth. So if you look on the bottom left here, the anthropic revenue grow revenue line massive and we think growth is still going to be impressive, but I do expect it to slow down for the reasons I've talked about because the tightening regulatory environment is going to slow model releases. Enterprise adoption is still narrow and it's starting to be managed. When you began this surge, a lot of companies were allowing people to use cloud code etc without real bottlenecks. Today there's budgets, there's the attempt to use competition openweight near frontier models that are cheaper and at the same time while the depth is getting deeper in terms of what it could do in coding and some other related area as a math the breath of usage is still relatively modest and not growing exponentially. So those are concerns for the break neck pace of revenue growth likely to slow.

7:43 and that's happening into a growing fear of well what will the space be when other models are catching up in terms of what they've released and are cheaper. There's a few things here. I don't think the picture is quite as bleak as it can appear on the surface. those models tend to do the the sort of Chinese models tend to do better on benchmarks than in practice. but there's some real things going on. I mean the first line shows probably the most misleading characteristic is what does it cost per token right? Claude Fable 5 very expensive $7.70 per token per million tokens. GPT 5.6 4.35 per million tokens and let's say GLM 5.2 90 cents or Kimmy K3 231. But what you really care about when you're paying is well how much does it cost to finish a task and how well do you do that task right? The next two numbers really show that which is because the US models and the frontier models tend to be more smarter they tend to do things to finish a task be in a better way such that GPD 5.6 soul is about a dollar4 per task fable 5 2.75 etc. those are closer to the competition than that and the numbers above show kind of how well they do and they generally do it better.

9:07 but still the competition is real and some tasks those models are roughly equivalent as as Bradley will get into with with certain types of forecasting you can use those models and get decent results although they come at it in a different way. The last thing I'd say is we're internally benchmarking many of the different tasks that we're doing and if you look at the Bridgewater Investment Systemization benchmark Claude Fable 5 is in the lead right now.

9:35 now that 42% just to differentiate that GLM 5.2 is at 28% and that's a huge difference in terms of usefulness at GLM 5.2 is lagging behind in terms of planning tasks etc. lagging behind even the older models like GPT 5.5. So that's kind of how we see the land the landscape right now which is look there's a place for these cheaper models. There's still not the cutting edge of intelligence and the cutting edge of intelligence is going to have value. but this is nipping at the heels in a certain way and from the psychology of markets when so much is priced in it's particularly dangerous and that's happening at a time when capital is needed. This is what I'm calling crunch time for capital for the AI markets. What's happening in terms of the growth in the need for capital is illustrated here on the bottom left. I mean it's incredibly fast growth. it's going to be difficult to get it and this is just what's needed to fulfill what's priced into equity markets today.

10:43 This isn't on top of that. So right now if you look at what we've got, we've got 567 billion came into the ecosystem in 2025. 62% of that though was from operating cash flow from the companies in the ecosystem. 214 billion from outside the ecosystem. 643 billion of capital is needed this year and we think that may be a challenge. that that the you've been able to fund the first half of that to some degree but more is needed in the second half and that is already crunch time and next year you need a trillion and it matters some of the companies open AAI's kind of the poster child for this they need the capital now we don't have a perfect model of this so this is roughly correct though that open AI has a huge burn rate they're burning down they will burn out of their cash without a new capital by Q127. Now, they'll probably get new capital, but how they get it is going to matter a lot. Can they IPO somewhere near what they were hoping or will that be difficult given the different challenges that they face in terms of revenue growth, etc., and and expenses and so on. So that just hits that they need capital, they need it now. And that might be occurring into an air pocket where people are starting to get nervous about the competition, about the tokens, about the funding simultaneous.

12:06 now none of that should take away that the capabilities keep growing. We see it's going to transform Bridgewater. We're using in our lab. We think we're very close such that 3 to 5 years from now we do expect Bridgewater in ways we can't even predict now will be extremely changed and those processes that we're using are making tremendous amount of pro progress. We'll talk about these with scientists in a minute but but basically the idea of harnessing cutting edge intelligence and taking external models and open models and training them reinforcement learning training them for your purposes those things are working for us huge deal at the same time adoption challenges regulation everything that I talked about into a market that is pricing in a tremendous this kind of bumpless ride is built a bit of a risk of an AI air pocket. Although I'd say we we're in the midst of that and that may play out relatively quickly and create actual opportunities to buy.

13:11 so AI will transform the world but because of regulation and those other things where the value occurs will shift and the ride will be bumpy in significant ways. So that's the backdrop.