Global AI compute, expressed in a chip that is already a generation behind. The fleet is the story. The model is the exhaust.
AiSignal staff · 7 September 2026 · 8 min read
The Stanford AI Index did the unfashionable thing this year and counted machines. Global operational AI compute: 17.1 million H100-equivalents, up 3.3× per year since 2022. NVIDIA still owns more than 60% of that pile. Google and Amazon mint a good deal of the rest on silicon they will not sell you. Huawei is a rounding error that has stopped rounding down.
An H100-equivalent is a confession. The industry's yardstick is a chip most new clusters are no longer installing. Blackwell is in the hall. Rubin is on the slide. We keep the H100 as a unit the way astronomers keep parsecs: because changing the ruler every eighteen months is how you lose the plot.
A city of racks
The geography is no longer a metaphor. xAI's Colossus in Memphis is the largest AI cluster at a single site — 555,000 GPUs, on the order of 2 gigawatts, a nuclear plant with a parking lot. Crusoe's Abilene campus, the Oracle/OpenAI and Microsoft neighbors, is planned for 2.1 GW and something like 400,000 GB200-class GPUs across 980 acres. These are not 'cloud regions.' These are industrial facilities that happen to emit completions instead of aluminum.
The model is the exhaust. The factory is the product.
Concentration is the other chart. Ten fleets — Google, Meta, xAI, Microsoft, OpenAI/Stargate, Amazon, CoreWeave, ByteDance, Oracle, Tesla — account for a first-pass 4 million accelerators. Everyone else is renting. Foundation-model companies, Epoch notes, often do not own the building. They lease the hour. The capex sits on someone else's balance sheet, which is why the hyperscaler print has become the real AI earnings report.
Watch the derivative, not the stock. At a 3.3× annual clip the installed base adds on the order of a large national supercomputer every day. That is the sound in the background of every product launch you will read this autumn.