AI's $3 trillion payback runs through six hyperscaler balance sheets
Sequoia partner David Cahn's updated capital spending estimate needs roughly $3T in cumulative AI revenue.
Sequoia partner David Cahn's updated capital spending estimate needs roughly $3T in cumulative AI revenue.
David Cahn's updated math says the AI infrastructure build needs about $1.5 trillion in annual capital spending, and roughly $3 trillion in cumulative AI revenue to pay it back. Cahn's 2026 update on Substack revives the framework from his 2023 "AI's $600B Question" and scales it 2.5x.
Cahn's required revenue per gigawatt of CapEx has risen sharply since 2023, because memory bottlenecks, HBM (high-bandwidth memory) supply, and a shift toward inference-specific silicon have raised the cost of delivering a unit of compute. Cahn notes that these factors mean the $1.5T and $3T figures may themselves be underestimates as the GPU-based methodology will increasingly underestimate future AI revenue requirements. The GPU that mattered in 2023 is no longer the chip that closes the loop in 2026. The wire restatements of his post this week have repeated the $1.5T and $3T figures; they have not named the mechanism.
The $1.5T in CapEx runs through a small set of counterparties. The de facto buyers are Microsoft, Google, Amazon, Meta, Oracle, and a handful of neoclouds (smaller, AI-focused cloud operators such as CoreWeave), the same group that has been buying Nvidia systems at scale since 2023. The set has not really diversified. If anything, it has tightened: the largest hyperscalers have pulled ahead on custom silicon, while smaller cloud buyers have fallen further behind on HBM allocation. Cahn's payback math does not assume demand from a diversified base of corporate buyers. It assumes that this small group keeps writing the checks, that their data-center buildouts keep clearing internal hurdle rates, and that the AI revenue their customers generate eventually catches up to the spend.
Apollo chief economist Torsten Sløk has framed the same risk from the macro side. A slower-than-expected AI payoff affects broad market conditions, not just the AI sector, because hyperscaler earnings feed the broader market, the broader economy, and the credit spreads that now help finance the build. The AI CapEx cycle is no longer funded only out of operating cash flow. It is also supported by data-center-backed debt, project finance from large equipment vendors, and private credit. Sløk does not dispute Cahn's math. He puts a name on what happens if the math slips: the drag transmits through the same small set of balance sheets, and through the financing structures built on top of them.
The revenue side is also split. Sequoia's "Tale of Two AIs" framing separates consumer AI (chat, assistants, image generation) from enterprise AI (agentic workflows, vertical software). The $3T target is unlikely to close evenly across the two. Enterprise AI is where per-seat and per-workflow pricing is supposed to scale, and where the closed deals are concentrated in a small number of Fortune 500 buyers. If enterprise AI revenue lags, consumer subscriptions do not make up the gap.
A broad market absorbs stumbles across many buyers. A concentrated one transmits them, because a capex guide-down, an earnings miss, or a debt covenant at any one of the six can reset the trajectory. Cahn's payback figure is the most quoted number in AI infrastructure this week. The number of counterparties underwriting it is the part that should travel with the quote.
The watch item into the second half of 2026 is whether any of the six hyperscalers signals a CapEx re-rate. The first to walk back its data-center commitments, or to push buildouts further out, is the one that turns Cahn's $3T question from a math problem into a market event.