The shift
One job, not a million apps
A traditional cloud datacenter runs millions of separate applications for many customers. A Fairwater hall runs one complex job across every chip inside it.
- Microsoft calls the result an AI superfactory: one job spread over millions of pieces of hardware.
- That inverts the design goals. Density and short wires matter more than mixing customers together.
The limit
Heat, not floor space, sets the size
Packing chips tightly is the point, and every watt leaves as heat. Air stopped being enough, so coolant now runs across the chips themselves.
| Rack power | about 140 kW |
| Row power | about 1,360 kW |
| First fill of coolant | what 20 homes use in a year |
| Coolant replaced | only if the chemistry says so, built for six-plus years |
It is a closed loop, so nothing evaporates. The water bill is a one-time cost, not a monthly one.
Geometry
Two floors to shorten the cables
Light itself is now the bottleneck. Every GPU connects to every other GPU, so a longer cable means a slower cluster.
- Two storeys let Microsoft arrange racks in three dimensions instead of one long row.
- Shorter runs buy latency, bandwidth, reliability and cost all at once.
Power
It trusts the grid and skips the diesels
Atlanta was chosen for a resilient utility grid, and that choice removed equipment other halls cannot do without.
- No on-site generation, no UPS, no dual-corded distribution.
- The target is four nines of availability at a three-nines cost.
- On-site storage and GPU power limits absorb the swings a huge training job makes.
Network
One flat network across the hall
Each rack holds up to 72 Blackwell GPUs on NVLink: 1.8 TB a second between them, and 14 TB of pooled memory for each.
- Racks join into pods and clusters over a two-tier Ethernet backend running at 800 Gbps.
- The switches run SONiC, Microsoft's own software, on commodity hardware rather than a vendor's.
Scale
Several sites behaving as one machine
One site runs out of power and space before a frontier model runs out of parameters, so the network keeps reaching outward.
- More than 120,000 new fibre miles went across the US in one year.
- Work moves between sites, so one model can train across several buildings.