Hyperscaler capex

What it is

"Hyperscaler capex" refers to the annual capital expenditure of the four US technology companies that operate the world's dominant public-cloud and AI compute platforms: Amazon Web Services, Microsoft Azure, Alphabet's Google Cloud, and Meta. The spending funds the physical layer of the AI economy, primarily GPU clusters, custom-silicon arrays, fiber networks linking sites, and power infrastructure. Nvidia captures roughly 90% of AI accelerator purchases from these four firms, making it the central vendor of the cycle. All four also develop proprietary chips as a partial hedge: Amazon's Trainium series, Google's TPU, and Meta's MTIA. Each company's annual capex crossed US$100bn by 2026, a threshold no private firm had reached before the AI buildout.

History

Public-cloud infrastructure spending grew steadily through the 2010s, funding conventional compute and storage. The release of OpenAI's ChatGPT in late 2022 triggered a step-change: all four firms began acquiring Nvidia H100 GPUs at scale in 2023. Combined hyperscaler capex rose from roughly US$150bn in 2022 to approximately US$260bn in 2023, then to US$410bn in 2025. The combined 2026 guidance of US$725bn, laid out in detail in company guidance from early 2026, is up roughly 77% in a single year. Microsoft's ~US$190bn commitment includes an explicit ~US$25bn component attributed to GPU and data-center component-price inflation, the first time cost pressure appeared as a named line item in guidance rather than a background assumption.

Current state

As of July 2026, the four firms guide to a combined US$725bn for the full year: Amazon ~US$200bn, Microsoft ~US$190bn, Alphabet US$175-185bn, Meta US$125-145bn. Q1 2026 alone saw Amazon spend ~US$43.2bn on property and equipment; Alphabet reported US$35.7bn in Q1 capex; Microsoft's fiscal-Q3 capex was ~US$30.9bn, up ~84% year-on-year. Google Cloud's contracted backlog stood above US$460bn in Q1 2026. Two projects illustrate the scale: Microsoft's Fairwater "AI superfactory" links Wisconsin and Atlanta via a dedicated AI wide-area network, while Amazon's Project Rainier in Indiana spans ~30 buildings and targets more than 2.2GW of power.

Relationships

Nvidia is the primary financial recipient of hyperscaler capex: its Q1 FY27 data-center revenue of ~US$72.8bn is the direct counterpart to the four hyperscalers' GPU orders, and the two are structurally linked. Critics flag the circular financing loop in which Nvidia invests in AI labs, labs commit to cloud capacity, and the clouds buy Nvidia chips, a self-referential cycle that analysts totalled above US$800bn by mid-2026. Power grids are a second dependency that binds all four: the Fairwater superfactory and Project Rainier each name power sourcing and permitting as the primary constraint on expansion, not chip supply. Oracle, which carries reported debt-to-equity of roughly 6x against its OpenAI cloud commitments, is a fifth player exposed to hyperscaler spending cycles through its own buildout.

What to watch

A June 2026 analysis placed the capex-to-revenue divergence at roughly 46% of hyperscaler revenue, exceeding the 32% peak observed during the 2001 US telecom excess cycle, and estimated a US$600bn gap between AI infrastructure spend and actual AI ecosystem sales. Amazon was projected to turn free-cash-flow negative in 2026 on the weight of its US$200bn program. The central questions: whether any of the four revises 2026 guidance downward on a demand miss; whether enterprise adoption converts AI pilots into measurable P&L impact (an MIT study found 95% of enterprise generative-AI pilots produced no measurable financial result on some US$30-40bn in corporate spending); and whether Amazon Trainium and Google TPU can reduce Nvidia's ~90% hold on AI accelerator revenue enough to meaningfully shift the cost structure of the buildout.