What’s Next for AI Progress?

This article was originally published on Goldman Sachs Insights, which features analysis and perspectives on the global economy and markets from across Goldman Sachs.

There are signs that artificial intelligence (AI) is exceeding even the bullish expectations of 2024. Usage of the most popular models has risen substantially, while the amount of investment in AI from the largest hyperscale technology companies is forecast to be about double what was predicted, says George Lee, co-head of the Goldman Sachs Global Institute. AI is helping to build more advanced generations of itself and has the potential to accelerate invention.

In short, technical achievements in AI are racing ahead. But even so, some important questions remain about the technology’s ultimate impact on the economy and corporations’ bottom line. Enterprise adoption of AI still must prove itself and overcome corporate inertia, and reasonable questions remain about the return on the extraordinary level of investment, Lee says in an interview with Tony Pasquariello, global head of Hedge Fund Coverage, on the Breaks of the Game podcast.

Pasquariello spoke with Lee, about investor sentiment toward AI, the impact of tariffs on the AI investment landscape, and the future of AI innovation.

This conversation was originally recorded on May 24, 2025. Below are edited excerpts from the discussion.

Pasquariello: It's been a little bit over a year since we last discussed the AI revolution. So let's start with just a quick mark to market about what has changed. 

We last talked on Breaks of the Game in the middle of March 2024. And in some ways that feels like yesterday. Yet in AI terms, it feels like that was the Paleozoic era. So much change, so much progress. And yet, as I'm sure we'll get to, there are some persistent questions that remain. So, I went back and wrote down a few comparisons from that March 2024 period to today—and AI may or may not have contributed to this—but here's just some interesting data.

OpenAI’s weekly average users when we talked last was approximately 180 million. Today, that's north of 800 million, probably north of a billion*.

The price of API tokens from leading models has declined about 100x* since we talked last.

Back then the 2026 capex expectations for the four principal hyperscalers was about $207 billion. Today, it's $405 billion*.

In March 2024, most observers hadn't heard of DeepSeek, Stargate, reasoning models, Elon's Colossus, and on and on.

It's a really extraordinary rate of change. But there are still some big questions: questions about enterprise adoption, reasonable questions about the ROI on all this capital spend, questions about what the winning business models will be.

There’s one question I get asked a lot about, which is what killer app has emerged in this period of time. I have an answer to that, which I’m sure we’ll get to.

* Figures mentioned are as of record date May 24, 2025.

Pasquariello: Those are all eye-popping statistics. The capex, from $207 billion to $405 billion, is probably the most intuitive for interlopers like me. But would you say when you take stock of it all, and again when you look back to March of last year, would you say the collective pace of advancement has exceeded what I think were already your high expectations?

Lee: In the dimension of technical progress, absolutely. We're finding our way into the future here. Sitting a year, or year and a half ago, it was unclear whether we'd run into some glass ceiling or unforeseen stopper of progress, and even during the intervening period, there have been questions about the scaling laws being over, etc. And we've just continued to motor through barrier after barrier. The rate of technical change is higher and more sustained than I might have expected.

I think enterprise adoption is probably fairly consistent with my expectations, because, as we talked about then, that's sticky. There are a lot of inertial factors. But I wouldn't say enterprise adoption has outperformed expectations. I would say it's largely on track and starting to inflect up. In general, though, the train continues to accelerate down the tracks.

Pasquariello: I want to get your sense for sentiment on AI today. There's been a lot of ups and downs in the world. When it comes to AI, it’s specifically been in the form of DeepSeek, but more broadly, of course, tariffs and more. You're based in the Bay Area. But you spend a ton of time visiting with our clients all over the world. What are you hearing?

Lee: I said when we chatted last that almost every observer I talked to or met with was uniformly bullish, with a few exceptions. I'm sure, and as you know, at the same time I had predicted that there would be troughs of despair or concern along the way. And there’s an Economist article out this week with that in the title. So, I suppose that was slightly predictive.

I think overall opinion is slightly more measured now. You find plenty of bulls, but also people with real questions and doubts. As I said, I think the technical progress is undeniable. I think we have found product-market fit in the area of coding assistants and coding agents. The progress there, and the impact of that, is pretty hard to dismiss.

I think sentiment is perhaps more measured, perhaps more balanced, but still generally bullish. Interestingly, I'm just back from Europe, and I had an interesting experience. I was talking to a large room of investors, and I asked people to raise their hands if they believed that the advent of chatbots had interrupted the number of Google searches they're doing.

And if you ask that question in Silicon Valley, every hand will go up in that room. I would say 10% to 15% of the hands went up with that audience in Europe. So, I think we're in a world that's not a flat world. It's a little bit spiky. There are places where this is more impactful, where there's more adoption, more bullishness, where there's just less engagement.

One area where there's more engagement is China. A number of our colleagues have just returned from visits to mainland China, where they are struck by the degree of proliferation, confidence in, and impact of AI systems. You see it everywhere, from facial recognition and palm payment systems that basically remove queuing from the world to the advent of robotic taxis. AI seems to have permeated that commercial environment at a really accelerated rate, which I think has meaningful implications for the US and for the world at large.

Pasquariello: Maybe we'll jump off the China point. We have this 90-day tariff pause that expires in July. Is this just kind of short-term macro noise that shouldn't distract from the big ball in AI land? Or could it? Could the tariffs have a meaningful impact on the AI names and AI cycle?

Lee: I think in the first order, I would say the tariffs and trade wars, if you want to call them that, have had minimal impact on the trajectory of adoption, investment, etc. I would have expected in some ways that this economic uncertainty would have played through more. I just personally haven't seen it.

But I think there's a more pernicious thing beneath the surface there, which is, these systems emerge from extremely complex and intricate supply chains. Jensen Huang said recently that their current NVL72 system has about 600,000 parts, which is extraordinary. He further said that their Khyber rack, which is coming in roughly 2027, will have approximately 2.5 million parts. He didn't disclose how many of those are foreign-sourced. But there are things like interconnects, wiring, cooling, power systems, etc.

As you know, chips right now are at least temporarily exempt from import duties. But all these other subsystem parts are often manufactured abroad, often not available from multiple geographies, hard to replicate, and will be subject to duties.

This is an interesting thing. We've got an administration in Washington that, on the one hand, is extremely competent, bullish, and positively inclined toward AI and sees it as a critical dimension of US competitiveness. On the other hand, they're imposing some headwinds here to progress in that area, and it's interesting to see how that all plays out.

Pasquariello: I'm reading a headline that you published recently: “When AI builds AI: The next great inventors might not be human.” Can you elaborate on that thesis?

Lee: It's from an article that I wrote for Fortune. The essence of the article is that I've been really struck by the degree to which AI systems are contributing to the development, refinement, and advancement of their successors. And observers like Satya Nadella and Mark Zuckerberg and others have commented on the degree to which these systems are central to the advancement of next-generation AI.

It comes in three or four different areas. One is that we've largely run out of human-generated data to train these machines on. And so, the machines are now being used to generate synthetic data that advances their pre- and post-training. That’s very clearly AI building AI.

Second is this element of post-training, which in the era of reasoning models has become more and more central. Machines are doing reinforcement learning for versions of themselves. They are generating hypotheses, evaluating the quality of those hypotheses, and doing what's called rejection sampling. They're throwing out bad answers in favor of good answers and then using that to reward the model. That's obviously the use of these technologies to advance themselves.

And then, finally, one of the things people I've heard a lot about are these small language models. Small language models are born from very large models. Very large models are then distilled by the models themselves down into smaller form factors that are maybe more domain-specific or more limited in their functionality. So, you start to add that up, it's just extraordinary the extent to which these machines are contributing to this already steep pace of progress in the technology itself.

The inventor analogy, I think, sprang from one paragraph I had in the article which talked about a paper out of Google that described a system they've built called Co-Scientist. Co-Scientist is an agentic framework that is designed to help inventors accelerate the pace and breadth of what they can achieve. And it's a fascinating paper. I recommend it to everyone. It hints at a world in which these agentic frameworks can marshal evidence, organize the evidence, assess hypotheses critically, analyze the output of those hypotheses, and at least advance good theories for change or invention.

But probably most frontally, they're just an extraordinary tool for humans to elevate the inventive capacity we already have.

Pasquariello: We'll end with an informal question, as we always do. If AI could take over one routine task for you, what would it be?

Lee: I'll give you a business one and a personal one in business.

You and I are both generationally appropriate for an X-Files reference. One of the rallying cries in “The X-Files” is, “the truth is out there.” And inside Goldman Sachs there is such a plethora of data, you could answer almost any question you might have from information resources that exist across our firm. That ability to have a central place where I can go, ask, get answers and get access to that incredible breadth of data and information and perspectives, that to me is the ultimate for a knowledge worker, the ultimate killer app. And we're making progress toward that, with the appropriate controls and guardrails. You'll be greeted with a blinking chatbot at your desktop, and you can ask anything and get access to any piece of information that exists around the firm, if it's compliance appropriate.

Personally, I think the thing I'm looking forward to would be really ambient systems. So rather than being anchored to my phone or my PC, having something that's sort of ever-present around me, that can do things like remind me of things and prompt me for things, or course correct things. I think that breaking free of the tyranny of our devices and having more ambient sources of intelligence would be super cool. That's many years in the future. But that’d be very fun and very useful.