Microsoft has approximately 2 gigawatts of AI-specific data center capacity. Bloomberg confirmed this in September 2026. The problem: Microsoft spent the better part of a year claiming it was adding a gigawatt of capacity per quarter, three quarters straight. Those two numbers do not reconcile. Ed Zitron's investigation, run jointly with The Guardian, found roughly 2.2 million GPUs in operation, representing about 1.993GW of compute backed by approximately $50 billion in hardware, predominantly NVIDIA H100 and H200 chips with some GB200 and GB300s. The remaining 10 gigawatts Microsoft cites is CPU infrastructure, not AI-specific capacity.
The core argument here is not about delayed construction timelines. It is about what Satya Nadella himself called GPUs 'sitting in inventory that we can't plug in.' Zitron has been tracking this since an earlier piece asking where the data centers physically are. The answer, across multiple investigations, is that most announced Microsoft projects had barely broken ground. The Fairwater campus, dedicated to OpenAI, is partially lit. The rest is either under construction or not yet started. NVIDIA's revenue growth, Zitron argues, is fueled almost entirely by speculative bulk purchases from hyperscalers and neoclouds, chips that take years to actually deploy.
Part 2 of Zitron's Hater's Guide to AI Debt publishes Friday, focused on the spiraling costs of standing up compute infrastructure and the existential risk this poses to counterparties including Oracle. The original piece is worth reading in full for the sourced GPU inventory breakdown, the specific wattage calculations, and the detailed anatomy of how Microsoft's public capacity claims diverge from what investigators can verify on the ground. If you have information on Anthropic, OpenAI, or related companies, Zitron is reachable on Signal at ezitron.76.
[READ ORIGINAL →]