Brad Gerstner on Whether AI Is a Bubble

Picture two lines on a chart moving in near-perfect lockstep. One line is how much money Microsoft, Google, and Amazon are spending to build AI data centers. The other is the free cash flow flowing into the chip and memory companies that supply them. Investor Brad Gerstner, founder of the investment firm Altimeter, put that chart in front of a room full of investors and asked what they noticed. The two lines are almost identical. Hyperscaler capital spending, meaning the money big cloud companies spend on infrastructure, is converting into chipmaker profit at nearly a dollar-for-dollar rate (04:42). That one image does a lot of work in explaining what has happened to markets this year.

The scoreboard, as Gerstner lays it out, looks odd at first glance. The Nasdaq is up fifteen percent this year and thirty-nine percent since January of last year, despite tariff fights, geopolitical noise, and constant chatter about AI regulation (02:29). Gold, which plenty of people bet would rip, is flat. Bitcoin is down ten percent. Meanwhile Nvidia's revenue has doubled, hyperscaler capital spending has doubled, OpenAI and Anthropic valuations have doubled, and SpaceX is up two and a half times. Dell is up five times in eighteen months. Memory maker SK Hynix is up nine times. Those are venture-capital-style returns showing up in public, liquid stocks, and Gerstner's explanation is simple: the infrastructure market is so tight that ordinary industrial companies are behaving like startups.

This is where Gerstner pushes back hardest on the word everyone keeps reaching for: bubble. His argument rests on one distinction. In a bubble, prices rise faster than the businesses underneath them. Gerstner says that isn't what's happening now. Market earnings are up twenty six percent this year, largely on the back of AI infrastructure spending, while the price investors pay for those earnings, the multiple, has actually gone down. Nvidia, the chipmaker whose processors run most large AI systems, trades at fourteen times next year's fully taxed GAAP earnings, meaning earnings calculated under standard accounting rules after taxes, which is well below where the Nasdaq and the broader chip sector, known as the SOX index, have historically traded (03:16). As Gerstner puts it:

"This is not about multiple expansion. This is an earnings-driven market expansion." (Brad Gerstner, [03:16])

The catch is concentration. Semiconductors alone account for seventy percent of the Nasdaq's total return this year (04:11). Huge parts of the market, consumer discretionary, software, financials, have barely moved. Gerstner calls this both good and bad: good because the earnings behind the rally are real, bad because so much of the market's fate now rides on a small number of chip and infrastructure names.

The Only Number That Matters

Gerstner traces the current cycle back to a specific, slightly awkward moment. In October of last year he interviewed OpenAI CEO Sam Altman and Microsoft CEO Satya Nadella, and he asked Altman a blunt question: how do you commit to a trillion dollars in capital expenditure when your company has thirteen billion dollars in GAAP revenue? Altman's answer, Gerstner says, was to tell him to sell his shares (05:16). The question didn't go away. It got answered instead by the numbers themselves. AI company Anthropic's monthly revenue reportedly went from two billion dollars in January to four billion in February to eleven billion in March, coinciding with the release of its Claude Code product (05:42). That jump lit the fuse for a rally in April and May. Then in June and July, Anthropic revised its expected annual run rate down from a rumored seventy five billion to sixty five billion, and the market stalled while investors worried whether open-source AI models were catching up. Gerstner's read on the whipsaw:

"We are on parabolic double exponential curves around these revenues." (Brad Gerstner, [06:31])

To put that in context, he notes that reaching a billion dollars in software revenue within four or five years used to put a company in the top five percent of the industry. Anthropic and OpenAI are now debating whether their monthly revenue will be four billion or eight billion. Gerstner estimates the combined run rate across the three largest labs, which he names as Anthropic, OpenAI, and SpaceX, sits around one hundred billion dollars based on rumors circulating in July, and he thinks they collectively need to reach one hundred and eighty billion by year end just to keep the AI trade intact (06:58). The reason this matters so much is structural: Microsoft, Google, and Amazon aren't paying for their own data centers, they're building them to rent out. Somebody else has to generate the offtake revenue, meaning the actual paying demand for that rented compute, to justify the spending. Gerstner estimates the industry needs to go from roughly two hundred billion dollars in current run-rate revenue to four hundred fifty billion, then eight hundred billion, then a trillion, just to keep pace with a trillion and a half dollars a year in projected capital spending (07:52).

Power, Regulation, and the Flight Path

If revenue is one constraint, electricity is the other. This year the US added about nineteen gigawatts of AI compute capacity, with about seven gigawatts going to the two leading labs. Forecasting firm SemiAnalysis, run by analyst Dylan Patel, projects forty three gigawatts of new compute for next year, with fourteen gigawatts going to the top labs (09:02). By 2028, Gerstner says, over half of all US compute will be controlled by just two labs (09:56). He argues the demand side is real: total knowledge work, meaning the combined market for consumer AI, advertising, coding, and white-collar business tasks, is so large that AI companies only need to capture about four percent of it, or roughly one point two trillion dollars, to pay for the infrastructure being built (10:14). As supporting evidence, he points to enterprise AI spending growing seventeen times over eighteen months, forty seven quadrillion tokens produced this year, and usage of the coding tool Codex up forty times in eight months (10:46). He also argues this shows up as margin expansion rather than mass layoffs: from 2015 to 2025, Nasdaq earnings growth ran about ten percent a year, split between six percent revenue growth and thirty eight basis points of margin expansion. Companies like Uber and Snowflake are telling him they'll grow revenue twenty to thirty percent without growing headcount, which is exactly how that margin number could move higher (11:38).

The bigger risk, in Gerstner's telling, is that the US won't actually be able to build forty three gigawatts of power capacity in a single year. Total US compute today is under forty gigawatts, and new projects face permitting fights, grid interconnection delays, labor shortages, and sold-out equipment (13:55). He points to a cautionary precedent: the US shut down sixty seven fission reactors under activist pressure, a decision he calls a unilateral disarmament against China's energy build-out.

"We unilaterally disarmed against China. It's been a disaster for the country." (Brad Gerstner, [12:36])

His own guess is that the US will stand up closer to twenty five gigawatts next year, not forty three, with roughly half going to Anthropic and OpenAI. He thinks that's still enough to hit revenue targets, since Anthropic is reportedly generating on the order of a hundred billion dollars in revenue off just a gigawatt and a half of compute today (14:16). Layered on top of the power question is interest rates: Gerstner puts the odds of a rate hike the next day at over ninety percent, which raises the cost of the debt funding all this data-center construction and makes safer, risk-free returns more competitive with stocks (15:05). He borrows Warren Buffett's framing for why that matters:

"Interest rates are to stocks what gravity is to matter." (Brad Gerstner, [15:05])

Gerstner closes with what he calls a flight path for managing a portfolio through this. If monthly AI lab revenue climbs toward eight billion dollars, he treats that as a sign of takeoff and expects an IPO before year end (15:54). If rates and oil prices stay contained and regulatory noise doesn't delay Anthropic's expected public offering, the trade holds. If any of those wobble, he'd trim exposure. The framing he leaves the room with is a shift in mindset from the last two years to the next one:

"If you were in the AI trade, you made money. That is not where we are in twenty twenty-six." (Brad Gerstner, [17:01])

From 2023 to 2025, in his words, investors needed to get exactly one thing right: that AI would be the biggest technology cycle in history, and then simply hold on. That phase, he argues, is over. Everybody already knows AI is a big deal, and it's mostly priced into stocks already. What comes next depends on facts, not conviction: whether monthly revenue keeps compounding, whether the grid can deliver the gigawatts, and whether rates stay low enough for all that borrowed capital to keep paying for itself.