AI Boom or Bust? – #122

Steve Hsu (0:03)

This message is for some friends at Mechanize, a startup that builds environments for training and evaluating frontier LLMs. Its customers include the top AI labs, and it has contributed to the breakthrough encoding capabilities of frontier models. Mechanize is hiring. See the links in the show notes. Compensation is extremely competitive for technical roles, $300,000 to $500,000 per year in salary, plus additional benefits and equity compensation. They are also seeking smart generalists. For example, a research engineer position focused on alignment. This person will build evals that test for misaligned model behavior, salary $500,000 a year. Puzzle maker. This person will design interesting and original puzzles that LLMs cannot yet solve. The salary component of compensation is $300,000.

Mechanize understands that my readership and my listenership on this podcast is highly selected. There is a very good chance you will be interviewed if you apply via the link in the show notes.

And you can see already by 2027, 2028, you're talking about revenues that have to be more than the trillions of dollars, okay? So this is an incredible amount of money. A good comparator is total software revenues. So all the software companies in the world together, you're talking about 1.5 trillion per year in revenue today. And what you're saying is that the AI companies or revenue really driven by AI itself has to more or less eclipse what we currently spend in total on all software by roughly 2027, 2028 That's the implied requirement if you want a decent return on existing and near future CAPEX. So these are the numbers that have everybody concerned to some degree. And it basically quantifies the hurdle that the AI industry has to hit.

This hurdle, these numbers in the next few years, or the CAPEX that's already been invested or currently planned for the next year or two, that will end up having a poor rate of return. So if we fall far short of these numbers for future AI revenues, then this investment will turn out when you look back, say it's 2030 and you're looking back at money that was invested in 2026 or 2027 You look back and say, wow, those guys got a terrible rate of return, maybe even a negative real rate of return from their investments. Then people will look back and say, this was a bubble.

Welcome to Manifold. This is a special episode. It will not be very long because I'm on the road. I was just at a big conference, not big in terms of size or number of attendees, but big in terms of the amount of capital, capital allocators represented here. As you can imagine, there was a huge amount of discussion about the AI CAPEX build out, the anthropic IPO, whether we're in an AI boom or an AI bubble, or in a sense, perhaps both could be true.

I thought I would share my views on this. In the slides I'm about to show you, I'm using figures that I myself have tweeted out or posted on X in the last, say, month or so, and you can go look at my timeline. One of the recurring themes I've been discussing on X is whether or not I believe that we're in an AI boom or bust, what is the evolution capabilities, how is this likely to translate into real productivity in the economy, revenues for AI companies, revenues for hyperscalers. Those are all the questions that I'm going to discuss in today's episode.

Before I get into the financial analysis, I just want to comment that the models themselves are improving very rapidly in capability. I think average people who are not themselves software developers or mathematicians or say theoretical physicists, wouldn't necessarily be aware of just how good they're getting. And in my own case, because I use the models heavily for research and also work on how to improve the model capabilities in areas like math and physics, I'm very aware of what's happened.

And I would say that for me, I'm living in a kind of science fiction world now, where I can talk to an AI or work with an AI that's actually better than, you know, most of the top level theoretical physicists in the sense that it has very, very broad understanding of our whole field. Literature that, for example, is outside of my own specialty, and so I haven't had time to dig deeply into that literature and master all the concepts and results. But it has. And then secondly, to do very, very powerful and complicated calculations very quickly. It's never been easier to do research in math and theoretical physics with the assistance of these AIs. And so that has been a huge change just from a year ago. So if you remember a year ago, I wrote my paper in which I had a result in quantum mechanics, in which the original idea actually came from GPT. And so I worked with GPT to produce a publishable result, which appeared in physics letters. And it was controversial at the time. A lot of people just were very negative about it because they thought that AIs were only capable of producing slop research. I think people's perceptions concerning that have changed within the last year. By now, probably everybody knows that some very important open problems have been resolved in mathematics using what I would call AI super intelligence. So I think it's pretty clear that they are superhuman in a bunch of different senses, even though their capabilities are quite jagged right now. I won't focus on this aspect in the current episode, but the math community is actually, I think, in a kind of trauma and adjustment phase, where they really have to adjust to the fact that they're in the presence now of things which are, they're not quite oracles, they can't solve every problem, but they can often make important advances on problems where humans have been stuck for a long time. I think the reason I just went through that discussion is so that you know that even though what I'm going to present here is a kind of bare case about the AI CAPEX boom and where the financial markets are likely to go with respect to AI.