Microsoft links Atlanta and Wisconsin into first 'AI superfactory'
Two Fairwater sites stitched by a dedicated AI wide-area network train one model in tandem, hundreds of thousands of Nvidia GB200/GB300 GPUs, 140kW racks, no UPS or gen-sets
Summary
Microsoft has wired its Fairwater sites in Wisconsin and Atlanta into what it calls its first "AI superfactory", two facilities joined by a dedicated AI wide-area network of optical fibre so multiple buildings train one model in tandem rather than each running separate jobs. The Atlanta Fairwater, online since October 2025, is a two-story design carrying hundreds of thousands of Nvidia GB200 and GB300 GPUs, ~140kW per rack and ~1,360kW per row, and, unusually, no UPS or backup gen-sets. Microsoft frames the WAN-linked clusters as cutting model-training time "from months to weeks." The build sits inside Microsoft's ~$190bn 2026 capex program.
By the numbers
- 2, Fairwater sites (Wisconsin + Atlanta) linked into one training fabric.
- ~140kW / ~1,360kW, power per rack / per row at Atlanta.
- Hundreds of thousands, Nvidia GB200/GB300 GPUs across the network.
- Oct 2025, Atlanta Fairwater began operating.
Why it matters
The "superfactory" is a structural shift: AI training is becoming multi-site, with grid and fibre, not a single building, as the binding constraint. Dropping UPS and gen-sets trades conventional reliability for density and speed, a template others will copy.
What to watch
- Whether more Fairwater sites join the WAN fabric and where they site (power-led).
- Reliability outcomes from the no-UPS/no-gen-set design under real load.
- Local grid and water permitting at each Fairwater campus.