Microsoft Fairwater is the name Microsoft gives to its purpose-built AI datacenters, each running a single cluster of hundreds of thousands of Nvidia Blackwell GPUs and linked to the others over a dedicated network. Two Fairwater sites are named as of October 2026: Mount Pleasant, Wisconsin, which Microsoft said was fully operational on June 23, 2026, and Atlanta, Georgia, which began operating in October 2025. Microsoft calls the connected sites an “AI superfactory.”

Everything below comes from Microsoft’s own blog posts and news releases.

On this page

What makes a datacenter “Fairwater”

Microsoft draws a line between Fairwater and its ordinary cloud datacenters. A typical cloud facility runs many small, independent workloads for many customers. A Fairwater site, in Microsoft’s description, is built to work as one AI supercomputer, using “a single flat network” that can tie together hundreds of thousands of GPUs for one training job.

Scott Guthrie, Microsoft’s executive vice president for Cloud + AI, set out the design in two posts: Inside the world’s most powerful AI datacenter on September 18, 2025, and Infinite scale: The architecture behind the Azure AI superfactory on November 12, 2025. The common elements are:

  • Rack-scale Nvidia systems. Each rack holds up to 72 Nvidia Blackwell GPUs in one NVLink domain, with 1.8 TB of GPU-to-GPU bandwidth and more than 14 TB of pooled memory available to each GPU. Microsoft says the Wisconsin site runs Nvidia GB200 servers, and that the Fairwater design can integrate “hundreds of thousands of the latest NVIDIA GB200 and GB300 GPUs.”
  • Two-story buildings. Racks are placed above and below each other as well as side by side, to shorten cable runs between GPUs and cut latency.
  • A flat scale-out network. Racks connect into pods and clusters through 800 Gbps links. The Atlanta post describes a two-tier, Ethernet-based backend network running SONiC, Microsoft’s own switch operating system.
  • Closed-loop liquid cooling. Liquid is filled once and recirculated, with heat removed by large chiller plants.

Fairwater Wisconsin (Mount Pleasant)

Microsoft introduced the Wisconsin site on September 18, 2025, calling it “the world’s most powerful AI datacenter” and saying it would deliver 10 times the performance of the world’s fastest supercomputer at that time. Guthrie’s post gave its physical scale:

  • 315 acres and three buildings with 1.2 million square feet under roof
  • 46.6 miles of deep foundation piles, 26.5 million pounds of structural steel, 120 miles of medium-voltage underground cable and 72.6 miles of mechanical piping
  • storage systems “five football fields in length”

Microsoft said each GB200 rack can process 865,000 tokens per second. Cooling runs through “the second largest water-cooled chiller plant on the planet,” with hot water piped to cooling fins on each side of the building, where 172 20-foot fans chill it before it returns. Microsoft says more than 90% of the site’s capacity uses this closed loop; the other 10%, traditional servers, uses outdoor air and switches to water only on the hottest days.

Investment and timeline. In a post the same day, Microsoft President Brad Smith said bringing the first datacenter online would fulfill an initial $3.3 billion pledge, and Microsoft committed another $4 billion over three years to build a second datacenter of similar size, bringing its total Wisconsin investment to more than $7 billion.

On June 23, 2026, Microsoft announced that it had completed construction of the first facility “after bringing equipment online and conducting startup activities in April,” and that the facility was fully operational. Microsoft said nearly 550 full-time employees work on site, the second facility next door is scheduled for completion in 2028, and it expects to spend $4.7 billion locally on hyperscale construction in Wisconsin between 2024 and 2028.

Fairwater Atlanta

The Atlanta site, according to Microsoft’s Source feature of November 12, 2025, began operation in October 2025 as the second Fairwater datacenter. It shares Wisconsin’s architecture and runs Nvidia GB200 NVL72 rack-scale systems.

Microsoft disclosed more engineering detail for Atlanta:

  • Power density: about 140 kW per rack and 1,360 kW per row.
  • Grid power: the site was chosen for resilient utility power and is “capable of achieving 4×9 availability at 3×9 cost.” That allowed Microsoft to drop on-site generation, UPS systems and dual-corded distribution for the GPU fleet.
  • Power smoothing: Microsoft says it co-developed software, GPU-level and on-site energy storage measures to damp the power swings that large training jobs create on the grid.
  • Water: the closed loop’s initial fill equals what 20 homes use in a year, is designed to last six years or more, and is replaced only if water chemistry requires it.

The AI WAN: linking sites into one machine

The feature that ties Fairwater together is the AI WAN, a dedicated optical network between Microsoft’s AI datacenters. In describing it, Microsoft says it delivered more than 120,000 new fiber miles across the US in 2024. The network lets Fairwater sites in different states, and earlier generations of Azure supercomputers, share training work such as pre-training, fine-tuning, reinforcement learning and synthetic data generation.

Microsoft explains the choice of distant sites in plain terms: land and power availability make it “more attractive – and at this point necessary – to spread the work across different physical locations.”

On May 6, 2026, Microsoft described the network protocol behind Fairwater, Multipath Reliable Connection (MRC), designed to keep large synchronous training jobs running through link failures. It said it was releasing the MRC specification to the Open Compute Project with partners and open-sourcing related libraries.

Chips: GB200 now, Vera Rubin next

Fairwater’s first two sites run Nvidia’s Blackwell generation. On January 5, 2026, Microsoft’s Azure hardware team wrote that its datacenters were designed for Nvidia Vera Rubin NVL72 racks “from current Fairwater sites in Wisconsin and Atlanta to future locations,” citing power, cooling and networking headroom built in ahead of time. Separately, Microsoft said its AI datacenters in Narvik, Norway and Loughton, UK, built with partners, will use similar clusters with Nvidia GB300.

Fairwater at a glance

Fairwater Wisconsin Fairwater Atlanta
Location Mount Pleasant, Wisconsin Atlanta, Georgia
Status (Microsoft) Fully operational, June 23, 2026 Operating since October 2025
GPUs Nvidia GB200, hundreds of thousands in one cluster Nvidia GB200 NVL72, scalable to hundreds of thousands
Building Three buildings, 1.2 million sq ft, 315 acres Two-story design
Rack power About 140 kW per rack, 1,360 kW per row
Cooling Closed-loop liquid; 172 20-foot fans Closed-loop liquid; initial fill equal to 20 homes’ annual use
Power resilience Grid power only for GPUs; no on-site generation or UPS
Investment More than $7 billion committed in Wisconsin

Specs for the GB200 and GB300 systems inside Fairwater are in our AI chips catalog, and both sites sit among other projects in the AI data centers tracker.

Sources