Microsoft announces 'world's most powerful' AI data center — 315-acre site to house 'hundreds of thousands' of Nvidia GPUs and enough fiber to circle the Earth 4.5 times
The Wisconsin-based data center is set to go online in 2026
Microsoft is planning to bring the "world's most powerful" AI datacenter online in early 2026, the company announced today. The Mount Pleasant, Wisconsin-based datacenter, dubbed Fairwater, is. meant specifically for training AI models as well as running large-scale models. The datacenter will be housed on 315 acres of land, with 1.2 million square feet in three buildings to house "hundreds of thousands" of Nvidia GB200 and GB300 GPUs.
If intelligence is the log of compute… it starts with a lot of compute! And that’s why we’re scaling our GPU fleet faster than anyone else.Just last year, we added over 2 gigawatts of new capacity – roughly the output of 2 nuclear power plants.And today we’re going further,… pic.twitter.com/cZJ3pdN1rXSeptember 18, 2025
On X, Microsoft CEO Satya Nadella wrote that these GPUs will be "connected by enough fiber to circle the Earth 4.5 times" and said that they will deliver ten times more performance than today's fastest supercomputer. This is likely a comparison to xAI's Colossus, which uses over 200,000 GPUs and 300 megawatts of power. Microsoft didn't specify its exact number of GPUs nor the expected power consumption.
Fairwater uses closed-loop water cooling, which the company suggests will have "zero water waste," with all of the water supplied once, at construction. In fact, Microsoft says it's the second-largest water-cooled chiller plant on Earth. Hot water will be sent out to cooling fins on each side of Fairwater, and then cooled with 172 20-foot fans before being sent back in to cool the GPUs again.
The other 10% will be traditional servers using outside air for cooling, and will move to water "only during the hottest days."
In a separate blog post, Microsoft president Brad Smith wrote that the company is working to avoid driving up electricity costs for surrounding communities.
The construction sounds immense. Executive vice president of Cloud and AI, Scott Guthrie, wrote that the new datacenter uses "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." The datacenter's storage systems alone are "five football fields" long.
Beyond the Mount Pleasant facility, Guthrie adds that several identical Fairwater data centers are under construction elsewhere in the United States.
Follow Tom's Hardware on Google News to get our up-to-date news, analysis, and reviews in your feeds. Make sure to click the Follow button.
Get Tom's Hardware's best news and in-depth reviews, straight to your inbox.

Andrew E. Freedman is a senior editor at Tom's Hardware focusing on laptops, desktops and gaming. He also keeps up with the latest news. A lover of all things gaming and tech, his previous work has shown up in Tom's Guide, Laptop Mag, Kotaku, PCMag and Complex, among others. Follow him on Threads @FreedmanAE and BlueSky @andrewfreedman.net. You can send him tips on Signal: andrewfreedman.01
-
DS426 300 MW of heat actually being dissipated up north by somewhat cooler air and non-evaporative cooling... Microsoft actually showing some sanity for once!?Reply -
hotaru251 alternative title: Wisconsin population about to see a big jump in their power bill due to mega corp setting up shopReply -
m3city I wonder if MS actually found AI stuff financially feasible. If not, this is going to sting when it comes to a simple .xls file with two columns: expense and income. And why not use the heat for citizens? I know solutions like this have been used in europe, at much smaller scale naturally.Reply -
iEatBalut Just go with the GB300s. You’ll be better off in the long run. The GB200s are the biggest pieces of crap NVidia ever made.Reply -
bill001g This smells of a AI article or someone who made no effort to research the announcements. It is even stated at the bottom "mount pleasant" which is on the other side of the state.Reply
Makes very little sense to put a data center out in the middle of nowhere where you can not get a electrical connections. Pleasantville wi is not really even a city it is part of a town called hale that has a population of less than 1000.
This is more likely related to other news article you see from more reliable publication like those in milwaukee wi talking about a second data center in "mount pleasant" which is just north of milwaukee. They were using water from lake Michigan to cool the first data center. -
jp7189 Wait a tick Nadella recently said they were slowing down new datacenter deployments because he expected the AI hype would cause others to overbuild and they expected to rent the extra capacity at a fraction of the cost.Reply -
JC5000 As a nuclear engineer that can use Chat GPT... 2 GW of gas power makes a woping 7,200,000 TONS of CO2 per year. Starting when the first power plant was built in 1882, it took human race 20 years to burn that much coal to produce what this does in 1 year... for a single data center. About 6 people die from lung disease per million tons so this single data center will cause 1000+ people to die prematurely over its lifetimeReply -
vanadiel007 Sometimes you have to truly admire nature. We consume food, turn it into energy, run our brain, and produce intelligence far beyond AI.Reply
AI that needs Gigawatts of power, hundreds of thousands of GPU's, and cannot reach intelligence.
Nature is amazing indeed. -
JRStern ReplyNadella said:
But that's kind of busted now, intelligence is NOT the log of compute, or it's an s-curve that hits an asymptote and we're already there. All the progress in LLMs in the last three years has been breaking that log scale down even further, coordinating large modules, and at least attempting to push off to inference time to avoid exponentials and logs on the entire universe.
If intelligence is the log of compute… it starts with a lot of compute! And that’s why we’re scaling our GPU fleet faster than anyone else.
It's still exponential/logarithmic even at inference time, but now you're doing it just on Taylor Swift and not also on quantum physics and the Peloponnesian wars so you save 99.999% of the work. -
Arkitekt78 This is the same MS that claims to care so much about the environment that they built this monstrosity and are forcing hundreds of millions of computers into the ewaste stream next month???Reply