AI’s Next Test Is Turning Data-Center CapEx Into Revenue

Brad Gerstner rejects a dot-com-style reading of the market’s AI-driven rise. His case is that earnings, rather than expanding valuation multiples, have carried the Nasdaq: Altimeter’s presentation showed Nasdaq-100 forward P/E falling from 26x to 24x year to date while next-twelve-month earnings per share rose 26%. Nvidia, Gerstner said, traded at 14 times the following year’s fully taxed GAAP earnings, versus Cisco at more than 100 times earnings in March 2000.

That does not make the market broadly healthy in his account. The gains are concentrated in the infrastructure layer supplying the AI build-out. Altimeter’s sector chart put energy up 42%, technology hardware up 39%, and semiconductors up 30%, while software was down 2% and consumer discretionary down 10%. Its Nasdaq-return chart attributed 68% of the index’s year-to-date return to semiconductors, up from 25% in 2023, 38% in 2024, and 44% in 2025.

Gerstner’s shorthand is that the “makers of tokens” are taking the gains while the buyers are “basically going along for the ride.” He pointed to infrastructure tightness as the reason public companies can produce what resemble venture-style returns: Dell was up fivefold and SK hynix ninefold in 18 months, he said.

The distribution of cash flows is central to the distinction. Gerstner described hyperscaler capital expenditure as flowing “almost dollar for dollar” into semiconductor-company free cash flow. Altimeter’s chart showed hyperscaler capex of $279 billion in 2025, $788 billion in 2026, and $1.36 trillion in 2027, alongside semiconductor free cash flow of $422 billion, $772 billion, and $1.288 trillion, respectively.

For Gerstner, the broad AI call was easier earlier in the cycle: from 2023 through 2025, an investor mainly needed to recognize AI as a historic technology supercycle and own the trade. That no longer settles the investment question. “Everybody knows about AI. It’s all priced,” he said. “Now it’s about facts and circumstances.”

The mismatch Gerstner is watching is straightforward: the companies building data centers are spending to rent capacity, and the rent must ultimately be paid by AI-lab and customer revenue. Microsoft, Google, and Amazon are not, in his description, financing this compute simply for internal use. The required offtake has to rise quickly enough to justify the infrastructure being financed.

Altimeter’s scenario begins with roughly $200 billion of lab AI revenue in 2025 and requires a steep progression: $450 billion in 2026, $800 billion in 2027, and $1.2 trillion in 2028. The corresponding projected hyperscaler-capex path rises from $422 billion to $663 billion, $956 billion, and $1.309 trillion.

The numerical hinge, Gerstner said, is monthly revenue at Anthropic and OpenAI.

Anthropic’s reported revenue progression is important to him chiefly because it changed the market’s operating test. After Opus 4.5 and Claude Code arrived in December, he said Anthropic revenue moved from $2 billion in January to $4 billion in February and $11 billion in March. That appeared to answer the question of whether AI revenue would materialize and, in his telling, helped drive the April and May rally. But when Anthropic later reported a $65 billion annualized run rate rather than the $75 billion some investors expected, estimates were revised and concerns about open-source competition returned. The issue is no longer simply whether labs can produce revenue; it is whether each month’s revenue supports the much larger spending path now assumed.

He estimated that Anthropic, OpenAI, and SpaceX had about $100 billion of collective run-rate revenue, explicitly basing that figure on rumors circulating in July. His threshold for keeping the AI trade intact was at least $180 billion by year-end, requiring another $80 billion across those three companies. The gap between $4 billion and $8 billion in monthly lab revenue, on this view, changes whether expected offtake can plausibly support the infrastructure being financed.

Gerstner does not see the addressable market as the weak link in the thesis. Altimeter’s presentation put total knowledge work at $30 trillion and argued that capturing 4% of it—$1.2 trillion of revenue—would be enough to support the projected capital expenditure. Gerstner included consumer services, advertising, coding, and white-collar enterprise workflows in that opportunity.

The usage evidence, he argued, is consistent with a demand surge. The presentation projected 47 quadrillion annual tokens processed in 2026, versus 0.6 quadrillion in 2023. Its “age of agents” slide represented the progression from chat consuming roughly 10³ tokens per unit of work in 2023, to reasoning at 10⁶ in 2025, and agents at 10¹² from 2026 onward.

Gerstner connected that increase to Jensen Huang’s earlier claim that inference demand could rise a billion-fold—an idea, he said, that many had dismissed at the time. Altimeter also showed Codex users growing 40-fold in eight months and median enterprise knowledge-work spending rising 17-fold over 18 months.

He believes the economic effect will show up in margins as much as in direct AI revenue. From 2015 through 2025, he said, Nasdaq earnings per share grew about 10% annually, composed of 6% revenue growth and roughly 38 basis points of yearly margin expansion. The question is whether AI can lift the margin contribution to 100 basis points. Gerstner’s answer is yes: he cited Uber’s intention to grow 20% without growing headcount and Snowflake’s intention to grow 30% on the same basis. The mechanism is not necessarily widespread firing, he said, but reducing the rate at which businesses hire people and engineers as revenue grows.

The tension is that demand may be sufficient even as construction targets prove unattainable. Gerstner described a SemiAnalysis forecast for 43 gigawatts of compute additions in the following year, including roughly 14 GW for the leading labs. He emphasized that this prospective annual addition is roughly as large as the country’s cumulative compute capacity at present.

He does not believe it will happen. His own estimate is closer to 25 GW delivered, with about half going to Anthropic and OpenAI. Permitting disputes, local opposition, grid-interconnection delays, skilled-labor shortages, and sold-out power equipment all stand between announced data-center plans and operating capacity.

A lower construction outcome need not invalidate the revenue case, in Gerstner’s view. He said Nvidia was reportedly generating $100 billion to $110 billion of revenue with around 1.5 GW of compute. If leading labs add another four or five gigawatts, he believes that could be enough to produce an additional $100 billion in revenue. The relevant question is not whether every announced gigawatt comes online, but whether the capacity actually delivered can support the revenue growth the market expects.

Gerstner identifies regulation, power, and higher-rate debt financing as the principal threats to the build-out, but he treats regulatory politics and physical delivery as distinct problems.

On regulation, his position is neither that AI should be left ungoverned nor that safety concerns should determine the pace of development. He said policymakers need pragmatic measures that give voters confidence that AI is safe while allowing work at the frontier to continue. He cited Elon Musk’s suggestion of peer review as an example of the kind of approach he thinks could help, while acknowledging that the political process will be messy.

His political concern is that fear produces an overcorrection. Gerstner invoked nuclear power as a precedent: he said activist pressure helped lead to the cancellation of 67 planned U.S. fission reactors, depriving the country of clean-energy capacity. AI policy should not repeat that history, he argued. Altimeter’s presentation cited more than 300 AI-related bills across more than 30 states and described New York as having imposed the first statewide data-center moratorium.

The execution risk is more basic: even if the regulatory environment remains permissive, data centers still require land, permits, grid interconnections, transformers and other power equipment, and skilled construction labor. Those bottlenecks are why Gerstner expects 25 GW of added compute rather than the 43 GW forecast he attributed to SemiAnalysis. The constraint is not a lack of prospective AI demand, in his telling, but the difficulty of turning announced projects into operating power and compute.

Rates introduce a separate financing constraint. Altimeter’s slide projected debt-funded capex rising from $28 billion in 2026 to $95 billion in 2027. Gerstner expected near-term rate cuts, but stressed that borrowing costs remain a hurdle for data-center investment because the projects are being financed with borrowed money. Higher rates raise the return projects must generate to justify that financing.

He applied the same logic to equities, invoking Warren Buffett’s phrase that interest rates are to stocks what gravity is to matter. If investors can earn 5.5% or 6% without taking equity risk, Gerstner said, stocks face a more demanding comparison. A 5.5% 10-year Treasury yield would be a substantial burden on the equity market in his view.

His portfolio stance is therefore a medium position rather than an unconditional bet. He would increase exposure if lab revenues come in strongly over the next several months and oil prices retreat; he would reduce it if the evidence deteriorates. Extreme leverage is especially dangerous, he said, in a market where the thesis now depends on operating delivery rather than a general belief in AI’s importance.