Fairwater, Microsoft's largest AI datacenter campus, links hundreds of thousands of NVIDIA's newest AI chips into a single continuous machine. The harder question is who pays for the electricity it draws.
Microsoft confirmed Tuesday that Fairwater, its largest AI datacenter campus, is fully operational in Mount Pleasant, Wisconsin. The wire framing is a GPU tally: hundreds of thousands of NVIDIA GB200 Blackwell chips wired into one continuous cluster. The harder question is what it costs to keep that cluster running, and who picks up that bill.
Fairwater sits on land in Racine County that was originally set aside for a Foxconn liquid-crystal-display factory that never got built. Microsoft broke ground in 2023 and finished construction roughly two years later, with about 10,000 construction workers cycling through the site. The company has around 550 full-time employees on the ground today, and a second facility of similar scale is under construction next door. Total Wisconsin commitment: $7.3 billion, a figure Microsoft expanded from an initial $3.3 billion pledge in May 2024 to its current size in September 2025.
The engineering is the point. Microsoft describes Fairwater as the closest commercial operation to a purpose-built AI supercomputer. The physical layout is built around a two-story rack design with networking running through the floor between levels, an 800-gigabit-per-second Ethernet fabric across the campus, and a proprietary networking protocol co-developed with OpenAI and NVIDIA to keep every GPU close to every other GPU. The result is a topology where a single training run can move across the full GPU inventory with low latency, which is what the company points to when it calls the site a coherent supercomputer rather than a cluster of independent servers.
That architecture has a direct consequence. Hundreds of thousands of top-end AI chips running in lockstep do not behave like a cloud region full of independent servers. They behave like one machine that draws power continuously. Microsoft has not disclosed the campus's peak load or its contracted capacity with We Energies, the local electric utility, and the company's own framing leaves "who will pay for the power" as an open question rather than a settled one. That gap matters. Local utilities do not absorb gigawatt-scale new loads in step changes; they absorb them through rate cases, transmission upgrades, and contested proceedings in front of state regulators.
The pattern here extends well beyond Fairwater. The next binding constraint on AI scale is no longer chip supply or rack density. It is power delivery and site selection, which is also grid permitting, water for cooling, and the willingness of a state public service commission to approve the build-out. Microsoft, Google, Amazon, and Meta are now competing on grid access and site selection as much as on chip counts, and Wisconsin ratepayers are exposed to that race whether they signed up for it or not. The next test will be how the second Fairwater facility comes online, how We Energies handles the load growth, and whether the Public Service Commission of Wisconsin treats the campus as a routine data-center tariff customer or as the leading edge of a new industrial-scale demand profile.