Microsoft opens AI Superfactory in Atlanta

Microsoft has introduced its next Fairwater AI datacenter site in Atlanta. The site connects with the original Fairwater facility in Wisconsin and other Azure AI infrastructure to form what Microsoft calls a planet scale AI superfactory. The design moves away from traditional cloud layouts and focuses on dense compute, low latency networking and flexible workload management.

Fairwater uses a single flat network that links hundreds of thousands of NVIDIA GB200 and GB300 GPUs so they operate as one large system. A dedicated AI WAN backbone connects multiple sites which allows training, fine tuning and other workloads to be distributed efficiently while increasing GPU utilization.
To support high density compute, the datacenters use direct liquid cooling with a closed loop system and a two story layout that shortens cable lengths. Each rack can support up to 72 GPUs, 1.8 TB of GPU to GPU bandwidth and access to large pooled memory. The design supports large training jobs and smaller tasks with different needs.

Microsoft also optimized power use, introduced software and hardware solutions to stabilize large jobs and built an optical backbone with more than 120,000 new fiber miles. These steps combine to deliver a flexible system built for the next wave of AI development.