The Big Four hyperscalers just told markets they will spend roughly $725 billion on capex in 2026 — a 77% jump from 2025's $410B and the single largest one-year corporate infrastructure buildout in modern history. Amazon is leading at approximately $200B (a $60B jump from 2025), Microsoft ~$190B, Alphabet $175-185B, and Meta $115-135B based on the latest restated guidance.
The context: as of June 30, 2026, AI stocks entered what analyst Dan Ives called an 'air pocket,' with Meta and Microsoft trading like 'bear market names.' Public markets are finally asking whether the spending will produce enough monetizable AI revenue in 2026 and 2027 to justify the depreciation drag. Microsoft's most recent quarter posted 34% cloud growth — strong, but not the acceleration bulls needed to underwrite $190B in spend.
“Microsoft's most recent quarter posted 34% cloud growth — strong, but not the acceleration bulls needed to underwrite $190B in spend.”
The money flow is almost entirely into three buckets: Nvidia GPUs (Blackwell and Rubin generations), custom silicon (Meta's MTIA, Google TPU v7, Microsoft's Maia, Amazon's Trainium 3), and physical data-center buildout — power, cooling, land. Grid interconnect delays are pushing hyperscalers into nuclear-power PPAs (Amazon-Talen, Microsoft-Constellation, Meta-Entergy) and even geothermal (Google-Fervo).
Second-order winners are massive: TSMC's 2026 revenue is expected to top $150B on this spend, ASML's EUV order book is booked out through 2027, and private AI-infra startups like Baseten, Groq, Modal and Together AI just posted their strongest bookings quarters ever. Contrarian read: the capex intensity looks a lot like the 1999 telecom buildout — durable long-term infrastructure but potentially followed by a brutal digestion period as depreciation catches up with monetization.
What to watch: Q3 earnings will be the first test of whether AI revenue growth is accelerating in line with capex growth. If not, expect one of the four to blink and cut 2027 guidance — the moment that happens, the entire AI-infra multiple compresses.