Microsoft Fairwater: A Full Guide to the Wisconsin and Atlanta AI Superfactory, AI WAN, Power and Limits

Microsoft calls Fairwater the world's most powerful AI datacenter. It has never said how many GPUs are inside, or how many megawatts it draws.

That gap defines Microsoft Fairwater at the end of 2026. The first Wisconsin building has been fully operational since June 23, 2026, Atlanta has been running since October 2025, and the two are tied together by a dedicated fiber network Microsoft calls the AI WAN. The hard numbers people quote, from GPU counts to gigawatts, come from outside estimates.

This full guide covers what Fairwater is, the Wisconsin and Atlanta sites, the racks and network inside, the AI WAN, power and water, the money, who likely uses it, how it compares with Stargate, Hyperion and Colossus 2, and the honest limits. Last updated: November 12, 2026.

Key Takeaways

• Fairwater Wisconsin's first facility has been fully operational since June 23, 2026; a second is due in 2028.
• Fairwater Atlanta began operation in October 2025 and links to Wisconsin over Microsoft's AI WAN.
• Microsoft says racks run at about 140 kW each and 1,360 kW per row with closed-loop liquid cooling.
• Microsoft discloses no GPU count or megawatts; Epoch AI estimates 369 MW of IT power in Wisconsin.
• Wisconsin's stated budget is more than $7 billion across two datacenters, before the chips inside.

Microsoft Fairwater at a glance

Let's start with what's on the record. The table separates Microsoft's own statements from third-party estimates, because the two get mixed up constantly.

Source: Microsoft's September 2025 Fairwater post, Made in Wisconsin post, November 2025 Atlanta post, Atlanta feature and June 2026 completion release; Epoch AI's Wisconsin and Atlanta records.

Notice the right-hand column. Every capacity number that matters most comes from someone other than Microsoft.

What Microsoft Fairwater actually is

Fairwater isn't a single datacenter. It's Microsoft's name for a design family of AI sites built to act as one supercomputer rather than a pool of separate cloud servers.

Scott Guthrie, who runs Microsoft's cloud and AI group, introduced it in September 2025 as the largest and most sophisticated AI factory Microsoft had built. The same post said multiple identical Fairwater datacenters were under construction in other U.S. locations.

Two things separate Fairwater from a normal Azure region. First, every GPU sits on a single flat network so a training job can span the whole site. Second, each site plugs into a dedicated long-haul network so several sites can work as one larger machine.

Microsoft calls that combination an AI superfactory. Here's why that matters: it's a bet that no single campus, however big, will have enough power for the largest training runs.

Fairwater Wisconsin: from the Foxconn site to fully operational

The first Fairwater sits in Mount Pleasant, Racine County, on land once earmarked for a Foxconn factory. Microsoft announced the facility in May 2024, and the June 2026 release says equipment came online in April before the site was declared fully operational.

Source: Microsoft's Made in Wisconsin post and June 2026 release; WPR on the January 2026 village approval; DCD on Nadella's April 2026 post.

The construction numbers are huge. Microsoft lists 46.6 miles of deep foundation piles, 26.5 million pounds of structural steel, 120 miles of medium-voltage cable and 72.6 miles of mechanical piping for the first site.

The local spending figure is smaller than the headline. Microsoft's June 2026 release estimates $4.7 billion of local hyperscale construction spend in Wisconsin from 2024 to 2028, about 10,000 construction workers, and purchases from 29 Wisconsin businesses. That's construction, not chips, so don't read it as the cost of the supercomputer.

The 15-building approval is the number to watch. WPR reported in June 2026 that the village board had approved site plans for up to 15 more Microsoft datacenters that could be built over the next decade. Approval isn't a commitment to build.

Fairwater Atlanta and the AI superfactory claim

Atlanta is the second Fairwater site. Microsoft unveiled it on November 12, 2025, and its newsroom feature says the site began operation in October that year.

Microsoft says Atlanta shares Wisconsin's architecture: GB200 NVL72 racks, a two-story building and closed-loop liquid cooling. The difference is the network role. Microsoft describes Atlanta, Wisconsin and earlier AI supercomputers together as the world's first planet-scale AI superfactory.

The exact location is fuzzier than Microsoft's wording. Epoch AI places the campus in Fayetteville, Georgia, south of Atlanta, and says it was built by QTS with Microsoft as the tenant. Microsoft's own posts name no developer.

Inside the rack: GB200 NVL72, 140 kW and a two-story hall

The racks

Each Fairwater rack holds up to 72 NVIDIA Blackwell GPUs joined by NVLink. Microsoft says each rack provides 1.8 TB of GPU-to-GPU bandwidth and more than 14 TB of pooled memory available to each GPU.

Microsoft also claims a single Wisconsin rack can process 865,000 tokens per second, the highest throughput of any cloud platform. It names no model, precision or batch size, so treat that as a marketing figure, not a benchmark. Inside Deep Tech's GB200 NVL72 full guide explains what the 72-GPU NVLink domain does and doesn't buy you.

Density and power per rack

Liquid cooling lets Microsoft push rack and row power to about 140 kW per rack and 1,360 kW per row. That's far beyond what air cooling can handle in a standard hall.

Not every source agrees on the rack number. The Register described Atlanta's GB200 NVL72 racks as rated for over 120 kW, with 720 petaFLOPS of sparse FP8 and 13 TB of HBM3e each. The gap likely comes from what each source counts as rack load.

Why two stories

Fairwater buildings stack racks on two floors. Microsoft's Atlanta post says this lets it place racks in three dimensions so cable runs stay short, which improves latency, bandwidth, reliability and cost.

It sounds like a small layout choice. It isn't. When every GPU must talk to every other GPU, cable length becomes a performance limit.

The network: one flat fabric on Ethernet and SONiC

Fairwater's in-building network is the part Microsoft is proudest of. It uses a two-tier, Ethernet-based back-end with 800 Gbps GPU-to-GPU connectivity, running on SONiC, Microsoft's own switch operating system, on commodity hardware.

Microsoft says that design lets it scale past traditional Clos network limits to hundreds of thousands of GPUs on one flat network, while avoiding vendor lock-in. Its September post also said Azure uses both InfiniBand and Ethernet fabrics at 800 Gbps, so the exact mix per site isn't public.

There's a footnote here. SDxCentral reported that Microsoft first described a protocol called Multi-Path Reliable Connected (MRC), built with OpenAI and NVIDIA, then removed that passage. The live post now carries an editor's note saying it was updated to explain the network more clearly.

The AI WAN: how Wisconsin and Atlanta work as one

The AI WAN is what turns two big datacenters into a superfactory. Microsoft says it delivered over 120,000 new fiber miles across the U.S. in a year to build a dedicated optical network for AI traffic.

Microsoft's newsroom adds that this raised its overall fiber mileage by more than 25% in one year, mixing new cable with fiber it acquired years earlier. The goal is to let sites in different states cooperate on near-real-time training, so jobs finish in weeks instead of months.

Distance is the physics problem. Data Center Frontier puts the Atlanta to Wisconsin gap at about 700 miles, which sets a hard latency floor no switch can remove.

That means the AI WAN likely suits work that tolerates delay, such as splitting pre-training, fine-tuning, reinforcement learning and synthetic data jobs across sites. Microsoft hasn't published inter-site bandwidth, or shown a single training run that spans both sites at once.

Power: grid-first in Atlanta, solar matching in Wisconsin

Power is where Fairwater makes its most unusual design call. Microsoft says the Atlanta site was chosen for resilient utility power capable of four nines of availability at three nines of cost.

Because the grid is that reliable, Microsoft says it can skip on-site generation, UPS systems and dual-corded distribution for the GPU fleet. That cuts cost and build time, but it also means the GPUs depend heavily on one utility connection. Inside Deep Tech's Colossus 2 guide shows the opposite approach, with on-site turbines and batteries.

Microsoft also says it co-developed three fixes for the power swings big training jobs cause: software that adds filler work during lulls, GPUs that enforce their own power caps, and on-site energy storage.

Wisconsin's story is about matching, not supply. Microsoft says it will match every fossil-fuel kilowatt hour one for one with carbon-free energy, including a new 250 MW solar project in Portage County, and that it is pre-paying for the energy and grid infrastructure it uses.

Cooling and water: closed loops and 172 fans

Fairwater's cooling claim is that it barely uses water once running. In Wisconsin, Microsoft says over 90% of capacity uses a closed-loop liquid system filled once during construction, with heat rejected by 172 twenty-foot fans on the cooling fins beside the halls.

The rest of the Wisconsin site uses outside air and switches to water only on the hottest days. Microsoft compares the site's annual water use to what a typical restaurant uses in a year.

Atlanta's initial fill equals what 20 homes consume in a year, Microsoft says, and the liquid is designed to last six-plus years. Epoch AI counts 672 air-cooled chillers at Atlanta, a sign of how much heat-rejection hardware a near-zero-water design needs.

On the flip side, dry cooling trades water for electricity. Chillers and fans draw power, which is one more reason the missing megawatt figure matters.

How big is Fairwater? Microsoft's claims vs Epoch AI's estimates

Microsoft says Fairwater holds hundreds of thousands of GPUs and will deliver 10 times the performance of today's fastest supercomputer. It gives no benchmark behind that claim. Epoch AI, which models AI datacenters from satellite imagery and filings, fills in the numbers.

Source: Epoch AI's Wisconsin and Atlanta records; GPU counts as summarized by Compute Atlas from Epoch's model. All are third-party estimates, not Microsoft figures.

Add the two GPU estimates and you get about 480,000, which fits Microsoft's hundreds of thousands. But Compute Atlas points out that Epoch's Wisconsin projection assumes many more buildings by 2028 than the two datacenters Microsoft has committed to.

What Fairwater costs, and how it fits Microsoft's capex

Microsoft's only Fairwater dollar figures are for Wisconsin: $3.3 billion for the first datacenter and $4 billion for the second, more than $7 billion in total. Atlanta has no stated figure.

Those numbers mostly cover buildings. Epoch AI's $14.0 billion Wisconsin estimate includes the computers, which is why it's so much higher.

The wider budget is enormous. On its October 28, 2026 earnings call, Microsoft reported $45.2 billion of capex for the September quarter and guided to about $170 billion for calendar 2026 because some third-party capacity deliveries slipped. DCD notes that's down from a previous $175 billion estimate.

Nadella also said Microsoft added 24 new datacenters across five continents in the quarter and has docked its first NVIDIA Vera Rubin GPUs. DCD noted that Microsoft now expects its datacenters to last 25 years for accounting purposes, up from 15.

The point is simpler. Fairwater is a showcase inside a capex program several times its size, and Microsoft doesn't break out its share.

Who uses Fairwater

Microsoft says its AI datacenters power OpenAI, Microsoft AI and Copilot, among other workloads. It hasn't said which customer runs on which Fairwater site.

Epoch AI lists OpenAI and Microsoft as likely users of both sites and says Wisconsin likely provides training compute for OpenAI. That's an outside inference, not a disclosure.

Microsoft is also training its own models. On the October 2026 call, Nadella said Microsoft had finished pre-training MAI-Base-2 and that Maia 200 delivers up to 1.7 times more tokens per watt in Teams and Copilot. Inside Deep Tech's Microsoft Maia guide covers that chip line; Fairwater itself is described as an NVIDIA site.

How Fairwater compares with Stargate, Hyperion and Colossus 2

Every frontier lab now has a flagship campus. Fairwater's distinct bet is the network between sites rather than one giant building.

Source: Inside Deep Tech's OpenAI Stargate, Meta Hyperion and xAI Colossus 2 full guides; Fairwater row from Microsoft's Atlanta post.

OpenAI Stargate: A Full Guide to Abilene, Every Site, Power and Honest Limits

Inside Deep Tech's full guide to OpenAI Stargate: the Abilene flagship, every announced site, financing, chips, power and what is actually built.

Inside Deep Tech

Inside Deep Tech's take

Honest limits of what's known about Fairwater

• No GPU count. Microsoft says hundreds of thousands. Every exact number, including Epoch AI's, is an estimate.
• No power figure. Microsoft hasn't published megawatts for either site; the 369 MW and 636 MW figures are modeled.
• No benchmark. The 10 times faster claim has no named test, and the per-rack token figure names no model.
• Unclear chip mix. Microsoft's wording covers GB200 and GB300, but it hasn't said which is installed where, or in what share.
• No proof of cross-site training. Microsoft hasn't shown a single job running across Wisconsin and Atlanta at once.

FAQ

What to watch next

• Microsoft's next earnings call in January 2027: any Fairwater capacity figure, and whether capex tops the over $50 billion guided for the December quarter.
• Progress on the second Wisconsin facility due in 2028, and any filings for the 15 additional buildings WPR reported.
• Any statement of AI WAN bandwidth, or a named training run that uses both sites at once.
• A third Fairwater site announcement, or GB300 and Vera Rubin racks confirmed inside Fairwater halls.
• Updated Epoch AI satellite readings for Atlanta's commissioning buildings.

Fairwater has already proved Microsoft can build and light up frontier-scale AI halls fast. The next test is whether the network between them makes several sites behave like one. Watch the AI WAN, not the slogans: the first disclosed cross-site training run is the milestone that counts.

Cover background: aerial photo of a Microsoft Azure datacenter near Wenatchee, Washington (not a Fairwater site) by Tedder, CC BY 4.0 via Wikimedia Commons, darkened and blurred.