Brad Gerstner：没有AI泡沫，半导体吞噬纳斯达克，AI的起飞难题

1. Trump Accounts, Every Child a Capitalist & The CAC Scan

Thank you, guys. And thank you for so much love yesterday, especially on the Trump Accounts. So many people came up. Everybody gets what this means for America. We’re in a battle for the soul of America. Seventy million kids are going to be made direct owners in America. That is how we beat the scourge of socialism: We make every child a capitalist.

And thank you to all the people yesterday who came up and took the CAC scan—the heart scan—from the Center for Heart Attack Prevention that we started. There’s no doubt, based on these results, we’re going to save some lives, even from yesterday’s scans. This is the highest-ROI thing you can do in health care. Every cardiologist I talk to does this for themselves, their family, and their friends.

It’s $100 and 15 minutes. Get it done. If we turn this into the mammogram for the heart, we’ll save 50,000 lives a year in this country. It’s like ending the war in Ukraine in America every year. So, go get your CAC scan done if you have it.

But today is not about those 2 moonshots. Today is kind of a throwback to what I used to do on the pod with these guys, which is a market check, a tech check, and a state of the market. Where are we? Where are we going? What do we have to believe to be true in order for the market to continue to work? We’re going to do a bit of a speed round here, so bear with me. Get your cameras out and get your notes out. There’s some good chart candy in here for you guys.

Markets are up 15% this year and up 39% since January of last year, despite all the concerns about tariffs, geopolitics, and AI regulation. The scoreboard: We have a lot of people in the bestie group who said gold was going to be off the charts this year. It’s flat. Bitcoin is down 10%. But look: We have NVIDIA revenue up 2x, hyperscaler capex up 2x, OpenAI and Anthropic valuations up 2x, and SpaceX up 2.5x in a pretty nasty backdrop.

This is not about multiple expansion. This is an earnings-driven market expansion. We’ve seen multiple contraction this year. Earnings are up 26%, driven a lot by AI infrastructure, but the multiple on the Nasdaq and the S&P is actually down. Look at NVIDIA trading at 14 times next year’s fully taxed GAAP earnings. This is no bubble like it was in 2000. The Nasdaq, S&P, SOX, and NVIDIA are all trading well below their average multiples.

The Magnificent 7 is basically in line with its average multiple, but not everybody is winning. At the bottom here, consumer discretionary, software, and financials—huge sections of the market—have barely moved. This is a market being driven by the largest capex buildout, the largest supercycle in the history of technology. Semiconductors are 70% of the Nasdaq’s return—70% of the return. That’s both good and bad, and we’ll get into that.

So who’s making the money? The makers of the tokens are making the money, and the buyers of the tokens are basically going along for the ride because of the tightness in the infrastructure market. We have massive public companies that look like venture capital returns. Dell is up 5x, up 9x in just 18 months. I love this chart. In blue, you have the hyperscaler capex. In orange, you have the free cash flow of the semiconductor companies. Do you notice anything? Their capex is almost dollar for dollar the free cash flow to the infrastructure companies.

2. Can AI revenue pay for the CapEx?

Where were we at the start of the year? This is a really important framework to get your head around. You may remember a certain podcast I did with Sam Altman and Satya Nadella in October of last year. I asked Sam a very basic question that was on everybody’s mind: How can you commit to $1 trillion in capex when you have $13 billion of GAAP revenue? I thought this would be a way to clear up some confusion in the market. Instead, he told me to sell my shares.

But then, in the beginning of December, we had Claude Opus 4.5 and Claude Code. In January, Anthropic’s revenue was $2 billion; in February, it was $4 billion; and in March, it was $11 billion. Do you notice anything that happened there in March? People started figuring out what Anthropic’s revenue was. We had an exclamation-point answer to the question, “Will the AI revenue show up?” It showed up in a massive way, and that’s why we got this historic run in April and May. We ripped off the bottom because the fuse was lit by Anthropic’s monthly revenue.

Then, in June and July, we had some consolidation. Why? Because Anthropic came out and said, “Our annual run-rate revenue is $65 billion.” People thought it was $75 billion, so they had to revise down their estimates a little bit. We also had some concerns about open source. Is open source catching up? Is Anthropic going to continue to be able to generate those revenues? So we’ve kind of moved sideways since then.

It’s hard to get your head around the fact that these revenues have never happened before in the history of capitalism. We are on parabolic, double-exponential curves around these revenues. The collective run-rate revenue of the top 3 labs—Anthropic, OpenAI, and SpaceX—is about $100 billion, based upon all the rumors coming out of July. I think they need to collectively get to at least $180 billion by the end of the year—adding another $80 billion across those 3 labs—just to keep the AI trade intact.

So this is the single most important data point in the market today: Is Anthropic’s monthly revenue, or OpenAI’s monthly revenue, going to be $4 billion or $8 billion? It’s almost hard to get your head around. Most of us who have been in venture capital for a long time know that if you added a billion dollars—if you got to a billion dollars in software revenue over 4 or 5 years—you were in the top 5% of software companies. But this is what the world is now pricing in.

Why is this so important? It’s this slide. If you’re going to build $1.5 trillion a year in capex, somebody has to pay for it, right? Microsoft isn’t paying for it. They’re building it to rent it. Google isn’t paying for it. They’re building it to rent it. Amazon is building it to rent it. Well, who is the person renting it? We have to have the offtake revenues in order to pay that rent.

If we exit this year around—let’s call it—$200 billion of run-rate revenue, I think you have to go from $200 billion to $450 billion to $800 billion or $1 trillion just to keep up. The blue bar is the expected capex just from the Magnificent 5, and the orange bar is the offtake revenue—the gap in offtake revenue that we need to see in order to keep this trade intact over the course of the next few years. Otherwise, we can’t build this much capex.

3. The Build Out Issue: Gigawatts, TAM, Token Growth, and Margin Expansion

The labs are increasingly—like in the conversation I had with Sam last October—aggressively expanding compute. Why? All the things you heard yesterday: We’re heading into recursive cycles. They’re seeing incredible demand for the product, and so they’re building out compute.

In 2026, the total amount of compute added is about 19 gigawatts, and about 7 of those gigawatts went to the 2 leading labs. Next year, this is SemiAnalysis—Dylan Patel’s forecast—the compute we’re going to add will be 43 gigawatts, and about 14 gigawatts will go to the leading labs. We’ll come back to this question of whether we can really stand up 43 gigawatts of compute next year.

Notice that the amount we’re adding next year is as much as the cumulative compute we have in the United States this year. So that’s what the market is anticipating. What happens if that doesn’t happen? And by the way, if you added those bars up, by 2028, to David Sacks’s point yesterday, over half of the total compute in the country is controlled by 2 labs.

So does the TAM exist? Does the TAM exist? If we look at the total TAM of knowledge work, this is a massive category. You’ve got consumer, ads, coding, and all these white-collar workflows, plus millions of enterprises. I would argue it’s the largest TAM in the history of the world. You only have to get to about 4% of that TAM, or $1.2 trillion, in order to pay for the capex. So I don’t think it’s a TAM issue.

Obviously, Jensen was on our pod, and he talked 2 years ago about how inference was going to go 1 billionx. Remember all the people saying he was full of it, that there was no way this could go 1 billionx? That’s exactly what we’ve done in the age of agents. We’ve had this exponential token growth this year: 47 quadrillion tokens are going to be produced.

So it’s not a question of demand. Codex users have grown 40x in the last 8 months, and knowledge work at enterprises—if you look at the median amount that enterprises are spending—is up about 17x over the course of the last 18 months. I’ve talked to many people in the audience here: small businesses, medium-sized businesses, large businesses, businesses like Altimeter. We can’t operate our business without buying AI, so this is not overly surprising to me.

And when we think about the productivity dividend to the economy, if you look at 2015 to 2025, EPS growth was about 10%. That was 6% revenue growth plus about 38 bps of margin expansion every year in the Nasdaq.

So, here's the question: Can we turn the 38 bps into 100 bps of margin expansion because of AI? The answer is obviously yes. Every company I talk to—Uber says, “We're going to grow 20%; we're not going to grow headcount.” Snowflake says, “We're going to grow 30%; we're not going to grow headcount.” That's what's happening. That is margin expansion.

The single largest cost input to every one of these companies is humans and engineers. It's not that they're going to fire everybody; they're just not going to hire them at the rate that they hired them before. Then, of course, we're going to have consumer agents in everybody's pocket. You may remember this bet I had with Bill: When are we going to be able to book a hotel using your consumer agent in your pocket? I think we've just gotten that with Muse and with Instinct. This could be another trillion-dollar category, but it's definitely going to consume massive tokens.

4. The risks: AI regulation, the nuclear precedent, power limits, and rising rates

Here are the 3 risks and challenges. Sorry I'm going so quickly, but I want to keep it moving: regulation—we heard a lot about this yesterday—power, and what's going on with interest rates. This is the regulation tug-of-war, right? I've heard from a lot of people, “You're on both sides of the issue.” Here's the fact: The answer is not going to be on one end or the other. We're going to have to have common-sense, pragmatic solutions that get my mom, my sister, and my brother off the cliff, right?

We need to give comfort and confidence to the people who elect our representatives that it's safe. You heard Elon yesterday give a great suggestion around peer review. I'm confident that we're going to get there, but it's kind of messy—the sausage-making along the way. We have a prior history of excess regulation. When people get scared, when activists start pushing an agenda, we shut down 67 fission reactors in this country. We unilaterally disarmed against China. It's been a disaster for the country.

All the clean energy we could have gotten—instead, we've gotten non-clean emissions because we had a group of activists who were hell-bent on shutting down nuclear. We can't allow this to occur with AI. Regulation is a threat. Atoms and energy are hard. Getting back to whether we can stand up 43 gigawatts of compute, our total compute in the country is less than 40 gigawatts. To do this in 1 year, we've got to overcome permitting and local opposition. You guys see all of that: grid interconnection delays, skilled labor shortages, and power equipment being sold out.

It's the largest buildout in the history of the country. I would suggest Dylan's forecast of 43 gigawatts next year is too aggressive. I don't think we're going to get there. I think the total amount we're actually going to stand up is somewhere closer to 25 gigawatts. I think of those 25 gigawatts, half will be for Anthropic and OpenAI.

I think that's enough to generate the revenue. Remember, Anthropic's revenue reportedly this year is $100 billion–$110 billion. If they do that, they're doing it with 1.5 gigawatts of compute. So, if they add another 4 or 5 gigawatts of compute, that's certainly enough to add another $100 billion in revenue. I don't think we need more gigawatts to get to the revenue targets for next year. This is my hunch: We're not going to get to 43.

Then, of course, we're going to hear more tomorrow. I think rate hikes are coming. I think there's now over a 90% chance that we're going to have rate hikes tomorrow. Why does this matter? Because all of this is now borrowed money, right? We're borrowing money in order to stand up these data centers, so the hurdle rate for that money is going up.

That's not only a challenge for data centers, but remember, as Warren Buffett says, “Interest rates are to stocks what gravity is to matter,” right? If you can earn 5.5% or 6% on your money without taking equity risk, then it's going to be a challenge for stocks.

Here's where I think we are, and we're going to end on this slide. Then we'll bring the guys out and chop it up a little bit. This is the flight path. This is how I think about managing the portfolio, right? The Nasdaq's up about 15%, but as I sit here and think about the risk, do I want to be small, medium, or large? This is how I think about the fan of potential outcomes.

If the monthly AI lab revenues are closer to that $8 billion number, I think it's takeoff. I think we are going to see an IPO this year. I'm paying very close attention to what, in fact, those monthly revenue numbers are. I think the trends are intact, but we will see. That's going to be the single most important thing as to whether or not, between now and the end of the year, we have liftoff.

The second one is rates, the election, and oil prices. Obviously, rates are following oil prices to a certain extent. So, what happens there? If rates were to go to 5.5% on the 10-year, that's going to be a big burden on the equity market. Then, finally, regulation and the Anthropic IPO. We saw a trade-down yesterday because people are concerned that maybe we're going to have a halt or postponement. That would obviously be a major issue. I don't think that's going to happen, but again, that would be a concern.

So, that's the fan of potential outcomes. From here, we're up 15%. I think we could go higher through the balance of the year, or we could go lower. I'll leave you with this: The period from 2023 to 2025, you only had to get 1 thing right—that AI was going to be the biggest supercycle in the history of technology. You needed to shove your chips into the AI trade. That's it, right? If you were in the AI trade, you made money.

That is not where we are in 2026. Everybody knows about AI. It's all priced now. It's about facts and circumstances. Stay mentally flexible. Follow the facts. Don't YOLO. Don't go 4x levered like our friend up north who gave all his money to Citadel, right? 4x levered in this market is very dangerous.

In my estimation, we're medium-positioned. We're mentally flexible. If we see those revenues come in big over the next few months and we see oil prices retreat, we're going to put more chips on the table. If not, we'll reserve the right to go even smaller. With that, thank you all. Thanks for having me.