Microsoft Fairwater Goes Live Ahead of Schedule — $3.3B Wisconsin Campus with Hundreds of Thousands of NVIDIA GB200s
Microsoft’s Fairwater AI data center in Mount Pleasant, Wisconsin is going live ahead of schedule, CEO Satya Nadella confirmed via social media. Nadella described Fairwater as “the world’s most powerful AI data center” — a campus designed to integrate hundreds of thousands of NVIDIA GB200 chips into a single tightly coupled cluster.
The site broke ground in 2023 as a 315-acre facility and has since expanded substantially. Microsoft secured planning approval for an additional 1,000 acres and received sign-off on 15 additional data center buildings in early 2026. A 160-acre adjacent parcel was acquired for $43 million in 2025. The $3.3B figure cited at the original September 2025 announcement covers Phase 1; Microsoft has committed a further $4 billion for a second data center on the same Wisconsin campus at similar scale.
GB200 at Density
The GB200 NVL72 rack — NVIDIA’s current Blackwell configuration — integrates 36 GPUs and 36 CPUs per rack with NVLink fabric, running at up to 120kW per rack. “Hundreds of thousands” of GB200 chips at Fairwater implies a cluster in the 3,000–5,000+ rack range, putting it among the largest dense GPU clusters built to date. For context, a 3,000-rack Blackwell cluster delivers roughly 3–5 exaflops of AI compute depending on precision.
Microsoft has not published exact chip counts or confirmed active workload status. The qualification “going live” likely means the facility is in final commissioning or has begun limited production inference, not full-scale training.
Hyperscaler Arms Race Context
Fairwater sits alongside comparable buildouts at AWS, Google, and Meta. Amazon’s Project Rainier cluster, built for Anthropic training, runs nearly 500,000 Trainium2 chips. Google’s TPU 8i inference chip, unveiled at Cloud Next this week, is engineered for the same dense agentic inference workloads Fairwater targets. Meta broke ground on a new $1B+ Oklahoma data center this week — its 32nd globally.
The common signal: training-to-inference ratios are shifting as frontier models stabilise and inference demand from agent deployments scales. Fairwater’s GB200 configuration reflects inference-first design — the GB200 NVL72 is purpose-built for long-context transformer serving at low latency, not pre-training runs.
Microsoft has not announced which Azure workloads or AI services will route through Fairwater first.