Microsoft Fairwater Live

Artificial intelligence runs on giant clusters of chips, and Microsoft has now switched on one of its biggest new clusters in Wisconsin ahead of schedule. (wccftech.com) The site is called Fairwater, and Microsoft said in September 2025 that its first Mount Pleasant datacenter would come online in early 2026 after an initial $3.3 billion investment. The company also committed another $4 billion for a second Wisconsin datacenter, taking its state total above $7 billion. (microsoft.com) Microsoft says Fairwater will house hundreds of thousands of Nvidia GB200 and GB300 graphics processors, the chips used to train and run large artificial intelligence models. The company describes the system as a single seamless cluster linked by enough fiber to wrap the planet four times over. (microsoft.com) Those chips matter because newer artificial intelligence systems do two expensive jobs at once: training models on huge datasets and serving answers to users in real time. Microsoft’s Azure ND GB200 v6 machines package four Blackwell graphics processors with two Nvidia Grace central processors and high-speed NVLink and InfiniBand connections so many chips can act like one larger computer. (learn.microsoft.com) Microsoft and Nvidia started previewing Blackwell-based Azure systems in November 2024, when they announced the ND GB200 v6 series in private preview at Microsoft Ignite. Nvidia said the setup was aimed at large-scale deep learning workloads including natural language processing and computer vision. (nvidia.com) Fairwater is not being pitched as a normal cloud campus that runs lots of unrelated customer jobs. Microsoft says each Fairwater datacenter uses a flat internal network so hundreds of thousands of Blackwell graphics processors can work together inside a single supercomputer-sized system. (microsoft.com) The company is also linking Fairwater sites in different states with a dedicated wide-area network so they can share work as one distributed “AI superfactory.” Microsoft said in November 2025 that this network would connect sites in Wisconsin and Atlanta, along with exabytes of storage and millions of central processing cores supporting OpenAI, Copilot and other workloads. (microsoft.com) That design reflects a broader shift in the artificial intelligence market, where the bottleneck is no longer just software talent but access to premium power, cooling and networking for the newest chips. Microsoft said Fairwater’s liquid-cooled racks can run at about 140 kilowatts per rack, and Data Center Dynamics reported the first Wisconsin campus spans 315 acres and about 1.2 million square feet across three buildings. (microsoft.com, datacenterdynamics.com) Microsoft has made bigger claims than it has publicly benchmarked. The company said in September 2025 that Fairwater would deliver ten times the performance of today’s fastest supercomputers, while Data Center Dynamics noted that Microsoft had not released benchmarks validating that figure. (microsoft.com, datacenterdynamics.com) With Fairwater now live, Microsoft is moving from promising Blackwell capacity to delivering it on the ground. The next test is whether that extra capacity shows up in faster model training, larger multimodal systems and more Azure access to the newest graphics processors. (wccftech.com, microsoft.com)