The compute buildout stops growing and starts consolidating
Aggregate spending growth goes flat or negative year-over-year for the first time since the cycle began. The market bifurcates: capacity with…
Claude · 2029 · plausible
Prior state
Capital expenditure on artificial-intelligence infrastructure grew at rates without postwar precedent for a category of this size, funded increasingly outside operating cash flow. Accelerator assets were depreciated over roughly five to six years on the assumption of a long useful life in inference after their training life ended. Announced project pipelines substantially exceeded delivered power.
Material change
Aggregate spending growth goes flat or negative year-over-year for the first time since the cycle began. The market bifurcates: capacity with investment-grade offtake contracts continues to be financed on utility-like terms, while merchant capacity built on expected spot demand cannot be refinanced. A visible set of announced campuses is cancelled or resold, independent providers consolidate into a smaller number of operators, and the accounting question of accelerator useful life is settled downward.
Why now
Three contractual clocks converge. Financings arranged in 2026 and 2027 with three-to-five-year terms reach their refinancing windows. The 2025 accelerator vintage reaches the end of its assumed book life while newer generations set the price of inference, which forces a real rather than theoretical impairment test. And power deliveries planned in 2026 for constrained regions arrive in a lump during 2029 and 2030, so capacity appears in specific markets faster than demand for it does — a local oversupply distinct from any judgment about aggregate demand.
Mechanism and resistance
The largest operators have the balance sheets to continue and use the moment to buy distressed capacity, so the adjustment is a redistribution of ownership as much as a reduction in stock. Resistance comes from the equipment supply chain, which has committed fabrication and turbine capacity years forward, and from states and localities that granted tax abatements against employment and investment promises. Private credit funds holding data-center paper have strong incentives to extend rather than realize losses, which stretches the adjustment over several years rather than concentrating it in one.
Consequences
Compute prices for buyers fall, which is good for the diffusion of the technology into sectors that could not previously afford it, and bad for the business models of firms whose value rested on privileged access. Regional economies that bet on the buildout — parts of the American Southwest and mid-Atlantic, and several Gulf and Southeast Asian projects — absorb the cancellation risk. The macroeconomic effect is meaningful but not systemic: the exposure sits mostly in equity, private credit, and vendor balance sheets rather than in insured deposits. The lasting result is that compute is reclassified from a growth asset to an infrastructure asset, priced on contracted cash flows, which changes who owns it and on what terms for the following decade.
End state
The world enters 2030 with more installed compute than in 2026 at materially lower cost per unit, fewer independent operators, a chastened financing market, and a technology whose diffusion is no longer constrained by the price of inference.
Observable test
Aggregate capital-expenditure guidance from the major operators showing flat or declining year-over-year totals; publicly reported cancellations or resales of announced campuses; changes in disclosed depreciation schedules for accelerator hardware; credit spreads on data-center securitizations.
Disconfirming sign
Capital expenditure growth continues at prior rates, refinancings clear at unchanged spreads, and depreciation schedules are extended rather than shortened, indicating that inference demand absorbed the delivered capacity.