Brad Gerstner's test for the AI trade: $4 billion a month or $8 billion a month

At the All-In Summit, Brad Gerstner did what he used to do on the show before he stopped: a market check, run as a speed round, with the deck doing the arguing. His subject was the trade everyone in the audience already owns, and his first move was to separate what has driven it from what people assume has driven it.

The numbers on his first slide: the market up 15% this year and 39% since January 2025, through tariff scares, geopolitics and the threat of AI regulation. Earnings up 26%. Multiples on the Nasdaq and the S&P down. Nvidia at 14 times next year's fully taxed earnings. Semiconductors responsible for 70% of the Nasdaq's return 1.

His conclusion from that slide is the sentence the rest of the talk unpacks: "The makers of the tokens are making the money, and the buyers of the tokens are basically going along for the ride."

Gerstner dates the current phase of the trade to a question he asked Sam Altman in October of last year: how do you commit a trillion dollars of capex against $13 billion of GAAP revenue? He had expected arithmetic. Altman told him to sell his shares.

The market's answer arrived in the first quarter of 2026. Anthropic's monthly revenue ran $2 billion in January, $4 billion in February and $11 billion in March, he said, and the AI trade took off from there — the fuse, in his telling, lit by the month the number crossed. Then it stalled. Anthropic revised its annual run rate down from what investors had assumed, questions accumulated about open-source models catching up, and the market moved sideways through June and July.

What he is watching now is the same series at a different order of magnitude. Across the three leading labs he puts the collective run rate at roughly $100 billion as of the July rumours, and he says the trade needs about $180 billion by the end of the year — another $80 billion across those labs — to hold. Expressed monthly, that is the choice he framed for the audience: whether a lab's monthly revenue arrives nearer $4 billion or nearer $8 billion is, in his words, the single most important data point in the market today.

The capex arithmetic is the part that makes the growth rate matter. Gerstner's framing is that Microsoft, Google and Amazon are not paying for the build-out; they are building to rent it out, and the rent has to come from somewhere. If the labs exit this year at around $200 billion of run-rate revenue, he says the figure has to travel to $450 billion, then $800 billion or a trillion, just to keep the capex flowing.

He does not think that is a demand problem. Knowledge work, by his measure, is the largest addressable market in history, and he calculates that roughly 4% of it — about $1.2 trillion — would cover the build. The supporting series he offers: 47 quadrillion tokens produced this year, Codex users up 40 times in eight months, and median enterprise spending on AI up about 17 times over the past 18 months. The margin story follows the same logic. From 2015 to 2025 the Nasdaq compounded earnings at about 10% a year — 6% revenue growth plus roughly 38 basis points of margin expansion annually — and his question is whether AI can turn those 38 basis points into 100. Uber growing 20% without growing headcount and Snowflake growing 30% without growing headcount are, in his reading, what that looks like from the inside of a company.

Here the talk stops being a defence of the trade and becomes a forecast of its physical limits. In 2026, about 19 gigawatts of compute was added, seven of it to the two leading labs. For 2027 he cites a SemiAnalysis forecast of 43 gigawatts, 14 of them to those two labs — a build-out the size of the entire installed compute base of the United States, which he puts at under 40 gigawatts today.

Gerstner thinks that forecast is too aggressive. Permitting and local opposition, grid interconnection queues, skilled-labour shortages and sold-out power equipment all argue against it, and his own estimate is nearer 25 gigawatts, about half of it for Anthropic and OpenAI. That would still be enough, on his arithmetic: Anthropic is reportedly heading for $100 billion to $110 billion of revenue this year on about 1.5 gigawatts, so another four or five gigawatts supports another $100 billion.

Rates come first. He put the odds of a rate increase at better than 90%, which matters because the build-out is financed with borrowed money and the hurdle rate on that money is rising. He invoked Buffett's line about interest rates being to stocks what gravity is to matter, and named a 10-year at 5.5% as a serious burden for equities.

Regulation comes second, with a historical analogy he returns to: the 67 fission reactors the United States shut down, which in his account was a self-inflicted wound that handed clean-energy leadership elsewhere. His stated position on AI rules is deliberately unheroic — pragmatic fixes that keep voters from being frightened, with peer review of models as one mechanism, a suggestion he credited to Elon Musk's appearance the previous day. Power availability and the timing of an Anthropic IPO round out the list.

What he does with all this is smaller than his case. His portfolio is the medium of three sizes, with the right to shrink if the next months of revenue disappoint or oil and rates push higher, and to add if the monthly numbers come in large. The reason is the difference between this year and the last three. From 2023 to 2025, he said, an investor had to get exactly one thing right — that AI would be the largest technology super-cycle in history — and then push in. That trade is over because everybody has made it. In 2026, the returns belong to facts and circumstances, which is why a talk that sounds like a bull case ends with an instruction to stay mentally flexible.