Independent 2026 estimates now cluster between $600 billion and $690 billion of hyperscaler infrastructure spend. That money doesn’t just buy GPUs — it buys the transformers, the construction crews and the front of every interconnection queue.
Multiple independent analyses of 2026 hyperscaler capital expenditure now land in the same range: roughly $600 billion at the floor and as high as $690 billion, concentrated among Microsoft, Google, Amazon and Meta, with Google alone guiding around $205 billion and Amazon near $200 billion.
The second-order effects matter more than the headline. Spend at this scale absorbs the global supply of long-lead electrical equipment, locks up EPC capacity, and fills utility interconnection queues years forward. Analysts tracking the buildout describe a market where the constraint has shifted decisively from silicon to sites, substations and megawatts.
For everyone competing underneath the hyperscalers — neoclouds, enterprises, sovereign AI programs — the practical consequence is scarcity in exactly the assets that can’t be fabbed: powered land, energized substations, and clusters that already have a delivery date.
What it means if you’re buying: When four companies absorb the primary market, the secondary market is where speed lives. Capacity that is connected, contracted or carrying a firm RFS date trades at a premium for a reason — it’s the only thing the capex wave can’t take from you.
What to do about it: The Ax3 book is built for exactly this squeeze: powered sites flowing today (ALTAIR 12, SW Kansas · VEGA 60, Georgia), 200–500 MW development platforms (POLARIS 200, ANDROMEDA 500), and GPU clusters with dates, not waitlists (MAGNETAR, December 1). Scarcity favors whoever moves first.
Analyses · futurumgroup.com ↗ · introl.com ↗ · bnef.com ↗ · Summary and analysis by Ax3
