blogs.microsoft.com web signal

Microsoft opens Fairwater, its largest AI datacenter yet

TL;DR

  • Microsoft's Fairwater datacenter in Wisconsin covers 315 acres across three buildings totalling 1.2 million square feet under roof.
  • Each rack packs 72 NVIDIA Blackwell GPUs with 1.8 TB of GPU-to-GPU bandwidth and 14 TB of pooled memory.
  • Microsoft says the site is 'purpose-built for frontier models' and will deliver '10X the performance of the world's fastest supercomputer today.'

Microsoft's newest AI datacenter, Fairwater, sits on 315 acres in Wisconsin, spans three buildings totalling 1.2 million square feet, and is what the company calls "the largest and most sophisticated AI factory we've built yet."

The write-up, from Scott Guthrie, Executive Vice President for Cloud and AI, on the official Microsoft blog, is heavy on physical numbers: 46.6 miles of deep foundation piles, 26.5 million pounds of structural steel, 120 miles of underground cable, and what Microsoft calls "the second largest water-cooled chiller plant on the planet," with 172 cooling fans standing 20 feet tall. The loop is filled once and then recirculated, with "Over 90% of our datacenter capacity" running liquid-cooled.

Inside, the racks are NVIDIA GB200 servers with 72 Blackwell GPUs each, 1.8 terabytes of GPU-to-GPU bandwidth and 14 terabytes of pooled memory per rack. Microsoft says the site is "capable of processing an astonishing 865,000 tokens per second" and pitches it as "purpose-built for frontier models." The bigger claim in the post: "It will deliver 10X the performance of the world's fastest supercomputer today."

Fairwater is not a one-off. Microsoft says there are "multiple identical Fairwater datacenters under construction in other locations across the US," alongside a hyperscale AI site in Narvik, Norway with nScale and an Aker joint venture, and what the company calls the UK's largest supercomputer in Loughton, again with nScale. All are stitched together through Microsoft's "Wide Area Network (WAN)" to enable "large-scale distributed training across multiple, geographically diverse Azure regions."

The post does not disclose the site's capex, its power draw at full utilization, the water volume consumed during the initial fill, or a date when Fairwater actually goes live.

Shared on Bluesky by 2 AI experts