billions to billions in AI infrastructure spending is projected for 2026, creating one of the largest synchronized capital deployment cycles in tech history.1 The bet hinges on one assumption: AI workloads will generate returns before the debt compounds.
That assumption carries real risk. Major hyperscalers — the cloud operators building data centers at scale — face a structural gap between when infrastructure goes live and when it generates revenue.1 Fixed costs accumulate from day one. Revenue follows later, if it follows at scale at all.
The risk assessment is direct: this exposure rates as catastrophic in severity.1 The mechanism is straightforward. Hyperscalers commission power management chips, networking gear, and GPU clusters months before customer workloads fill that capacity. If enterprise AI adoption moves slower than buildout timelines, the gap widens.
Stranded assets represent the worst-case outcome. Data centers built for AI workloads that never materialize at projected scale become write-down candidates.1 Write-downs at this magnitude trigger investor backlash and compress capital available for subsequent spending cycles — slowing AI development momentum across the industry.
Power infrastructure amplifies the constraint. AI data centers consume roughly 5–10x more power per rack than conventional compute. Grid capacity, permitting timelines, and energy costs create physical limits on how quickly costs can be recovered. Infrastructure that sits underutilized still draws power.
Supply chain exposure runs deep. ON Semiconductor's role as a power management chip supplier to major hyperscalers illustrates how far upstream the dependencies extend. Chip orders placed today reflect capacity plans for 2027 and beyond. If those plans prove optimistic, cancellation risk propagates backward through the supply chain.1
The systemic dimension is what separates this cycle from normal corporate overbuilding. Multiple hyperscalers are deploying capital on overlapping timelines. A correction doesn't stay contained to one balance sheet. Pullbacks ripple through semiconductor suppliers, data center REITs, power equipment manufacturers, and fiber networks — all simultaneously.
billions–billions deployed in a single year must be monetized over a decade to justify the spend.1 The math works if enterprise AI adoption curves steepen through 2027 and 2028. It breaks down if adoption plateaus while fixed infrastructure costs continue accumulating. The window for the cycle to prove itself is narrower than the buildout timelines suggest.
Sources:
1 Via News Risk Assessment — Major Hyperscalers AI Infrastructure Capex Exposure, May 2026