The Capex Cliff That Wasn't
Third-quarter results from the hyperscalers came in well above the bear case. The question is what year the bill comes due.
Coming into earnings season, the bear case on the largest cloud operators was a simple one: capital spending was running ahead of revenue attached to AI workloads, and the gap would eventually have to close on the revenue side or open on the cash-flow side.
Third-quarter results, in aggregate, did not validate that thesis. Combined AI-attributable revenue across the four reporting hyperscalers rose 62 percent quarter-over-quarter, and capital intensity, while elevated, fell modestly as a share of revenue. The interesting line item is the one for purchase commitments — the off-balance-sheet obligation to take delivery of accelerator inventory through 2028.
The commitment stack
Aggregate accelerator purchase commitments across the four hyperscalers rose to $312 billion by the end of the quarter, up from $221 billion at the end of the second quarter and $118 billion a year ago. Roughly two-thirds of that stack comes due before the end of 2027.
Those commitments are, in effect, a bet on a compound-annual growth rate for AI-attributable cloud revenue somewhere in the mid-forties percent range through the end of the decade. The bet is not obviously wrong. It is also not obviously right, and it is being taken with balance-sheet capacity that used to fund a great deal of buyback.
What the CFOs are and are not saying
All four CFOs, on their respective calls, emphasized "disciplined pacing" and "customer-committed capacity." None disclosed the share of the accelerator stack that is currently unallocated to a specific customer contract. Analyst questions on that number were, in every case, deflected. The number matters more than any of the ones on the income statement.
What breaks the bull case
A single quarter of decelerating AI-attributable revenue against a continuing capex ramp. Not two. One. The stack is now large enough, and the maturities close enough, that a five-point growth-rate disappointment would meaningfully change the sector's earnings algorithm.
Tech Editor based in San Francisco. Covers AI infrastructure and the people building it.
More from Tech
The Southeast's Power Grid Is Quietly Becoming an AI Story
Georgia Power's interconnection queue has grown sixfold in eighteen months. The line is almost entirely data centers, and the math no longer works without nuclear.
Anthropic's Claude 5 Lands With a Quieter Pitch
The marketing is enterprise-first. The benchmarks tell a more interesting story.
The Quiet Consolidation
The AI map at the start of 2025 looks less like a Cambrian explosion and more like an early oil patch.