Microsoft Fairwater Live

Microsoft says its Fairwater artificial intelligence datacenter in Mount Pleasant, Wisconsin, is going live ahead of schedule. (datacenterdynamics.com) Chief executive Satya Nadella said on April 16 that the Wisconsin site will bring together “hundreds of thousands of GB200s into a single seamless cluster.” Microsoft introduced Fairwater in September 2025 as a 315-acre campus with three buildings totaling 1.2 million square feet. (datacenterdynamics.com) (blogs.microsoft.com) A graphics processing unit, or GPU, is the chip that handles the parallel math used to train and run artificial intelligence models. Microsoft said Fairwater is built as one giant AI supercomputer, using a single flat network to connect those chips instead of splitting them into smaller cloud clusters. (blogs.microsoft.com) Microsoft said the campus is designed for both training and inference, the step where a model answers prompts after training is done. The company said Fairwater will deliver 10 times the performance of today’s fastest supercomputer for those workloads. (blogs.microsoft.com) The project is also a power-and-cooling story. Microsoft said Fairwater uses facility-wide liquid cooling, with racks drawing about 140 kilowatts each and rows drawing about 1,360 kilowatts, levels that are far above a conventional enterprise server room. (blogs.microsoft.com) Nvidia said in November 2025 that Microsoft’s Fairwater system in Wisconsin and a second site in Atlanta would integrate hundreds of thousands of Blackwell GPUs for large-scale training. Nvidia also said Microsoft planned to deploy more than 100,000 Blackwell Ultra GPUs globally for inference. (blogs.nvidia.com) Microsoft has tied Fairwater to a broader buildout. Brad Smith, the company’s vice chair and president, said on January 3, 2025 that Microsoft was on track to invest about $80 billion in fiscal 2025 on AI-enabled datacenters, with more than half of that spending in the United States. (blogs.microsoft.com) That spending has not been linear. In April 2025, Microsoft said it was “slowing or pausing” some early-stage AI datacenter projects, including work in Ohio, while continuing to align construction with customer demand. (thehill.com) (cfodive.com) Fairwater shows where Microsoft is still pushing hardest: sites with enough land, power, networking, and supply-chain coordination to pack dense Blackwell clusters into one system. Going live early does not answer how quickly Microsoft can fill every rack, but it does move one of its biggest AI capacity bets from construction into operation. (blogs.microsoft.com) (datacenterdynamics.com)