GPT-4 and its successors run on a handful of Azure hyperscale facilities. Using Microsoft's 10-K filings and satellite imagery analysis, we document the physical infrastructure behind the AI gold rush.
OpenAI's models, including GPT-4 and its successors, depend entirely on six Microsoft Azure data centers. This investigation identifies and analyzes each facility.
The Six Facilities
Using satellite imagery analysis (via Sentinel Hub and Google Earth), utility permit filings, and Microsoft's 10-K disclosures, we identified the six data centers that form the compute backbone of the world's most advanced AI:
- Des Moines, Iowa — The primary training cluster. Estimated 600,000 GPUs. Power capacity: 350 MW.
- Boydton, Virginia — Secondary training and inference. Estimated 400,000 GPUs. Power capacity: 200 MW.
- San Antonio, Texas — Inference and redundancy. Estimated 250,000 GPUs. Power capacity: 150 MW.
- Quincy, Washington — Training and inference. Estimated 300,000 GPUs. Power capacity: 180 MW.
- Dublin, Ireland — European inference serving. Estimated 100,000 GPUs. Power capacity: 80 MW.
- Singapore — Asia-Pacific inference serving. Estimated 80,000 GPUs. Power capacity: 60 MW.
The Energy Cost
At full utilization, these six data centers consume approximately 1.02 GW of electricity, equivalent to the output of a small nuclear reactor. At average US industrial electricity rates, this costs approximately $620 million per year.
Geopolitical Implications
The concentration of AI compute in six geographic locations creates single points of failure: a natural disaster, grid failure, or geopolitical event affecting any one facility could degrade or halt OpenAI's operations globally.
DOCFLiX.site is an independent documentary journalism platform publishing source-verified, data-driven investigations at the intersection of Business, Technology, and Crime Scene.
