Home โ€บ Blog โ€บ Microsoft Fairwater
AI Factory Case Studies ยท Part 3 ยท Microsoft Fairwater

Fairwater: How Microsoft Wired Two AI Factories Into One Supercomputer

Colossus bet on speed and Abilene on financing. Microsoft bet on the network. Its Fairwater sites in Mount Pleasant, Wisconsin and Atlanta, Georgia sit hundreds of miles apart but are joined by a dedicated optical "AI WAN" so they can work as one system, which Microsoft calls an AI superfactory. This case study looks at why a hyperscaler would split its biggest AI computer across states, what the buildings look like inside, and who pays for the power.

FL
FrontierAGI Team
โš ๏ธ How this was researched (as of 26 Oct 2026) Most technical detail comes from Microsoft itself: its September 2025 post on Fairwater Wisconsin, its November 2025 post on the architecture of the "AI superfactory", and its June 2026 announcement that the first Wisconsin facility is complete. Site history, jobs and power come from Wisconsin Public Radio, Civic Media, the Daily Reporter, DCD and Data Center Frontier. Size estimates come from Epoch AI. Microsoft has not disclosed total GPU counts or megawatts for either site, so we use Epoch's estimates and mark them. Performance claims are Microsoft's own. See Reading the numbers.

1. Case file

๐Ÿ”—
Fairwater Wisconsin + Fairwater AtlantaMount Pleasant, Wisconsin ยท Atlanta, Georgia ยท joined by the AI WAN
OwnerMicrosoft (Azure)
Wisconsin site315 acres ยท 3 buildings ยท 1.2M sq ft
Wisconsin investment$3.3B + ~$4B for a second facility
ChipsNVIDIA GB200, racks of 72 GPUs
Rack density (Atlanta)~140 kW per rack ยท 1,360 kW per row
Power now (est.)Atlanta ~636 MW ยท Epoch AI
Wisconsin end state (est.)~2.26 GW by ~2028 ยท Epoch AI
CoolingClosed-loop liquid, filled once

Microsoft says Fairwater Wisconsin holds "hundreds of thousands" of GB200 GPUs, connected by enough fiber to circle the Earth 4.5 times, and claims ten times the performance of today's fastest supercomputer. These are company claims; Microsoft has not published a benchmark result to back them up.

2. Build timeline

Fairwater, 2023 โ†’ 2028

Solid dots are confirmed by Microsoft or a public body, hollow dots come from trackers or the press, โš ๏ธ marks claims we could not confirm and dashed dots are planned. Click a dot for its source.

Click a milestone to see where it comes from.

The Wisconsin site has a history. The land in Mount Pleasant was part of the area set aside for Foxconn, which in 2017 promised up to 13,000 jobs and as much as $10 billion of investment, then built far less. Microsoft bought 315 acres there for $50 million in 2023. The first facility started up equipment in April 2026 and was declared fully operational on 23 June 2026, a little later than the "early 2026" first planned. A second facility next door is due by 2028, and in January 2026 the village board approved site plans for up to 15 more Microsoft data centers over the next decade.

3. The bet: one computer, many sites

The first two case studies both tried to win at a single site. That runs into a hard limit: no one grid connection can easily supply several gigawatts, and building generation on site brings the fights seen at Colossus. Microsoft's answer is to build several large sites where power is available and link them so a single training job can use more than one.

๐Ÿ’ก Why it mattersIf training can span sites, the size of a frontier model is no longer limited by the largest single grid connection, but by the total power a company can find across a region, plus the bandwidth between sites. It also gives flexibility: capacity can be moved between training, fine-tuning and serving as demand shifts.

4. The AI WAN

Microsoft says it has deployed a dedicated AI WAN backbone to integrate each Fairwater site into "a broader elastic system", allocating workloads dynamically and maximising the use of GPUs across the combined system. It says it laid more than 120,000 new fiber miles across the US in a year. Microsoft has said its planned UK and Norway sites, using GB300 systems, will join the same network.

The superfactory map

A schematic, not to scale. Tap a site. Dashed links are planned.

United States Europe Wisconsin Atlanta UK Norway ~670 mi straight line
Tap a site to learn more.

Distance has a cost. Light in fiber covers about 200 km per millisecond, so a 670-mile straight line (about 1,080 km) means at least 5 ms one way, and real fiber routes are longer. Inside a building, GPUs talk in microseconds. That is why cross-site training does not stretch one tightly coupled job across both sites naively; it splits the job so that the chatty parts stay inside a site and only less frequent exchanges, such as gradient synchronisation between data-parallel groups, cross the WAN. Microsoft has not published how it partitions jobs; the next section is an illustration of the trade-off.

5. Splitting a training run

One site or several?

An illustration, not Microsoft data. Each site is capped at the power you choose; adding sites adds compute but costs some efficiency for cross-site synchronisation.

Share of time lost waiting on the WAN for each extra site (assumed).

6. Inside the building

Microsoft's November 2025 post describes several design choices in Fairwater Atlanta:

Why two storeys shorten cables

A simplified cross-section: the same racks arranged on one floor or two. Toggle to compare the longest cable run.

Related reading: the engineers who run an AI factory's network and inside the AI factory.

7. Power and cooling

Unlike Colossus and Abilene, Fairwater does not rely on gas turbines inside the fence. It draws from the utility grid: We Energies in Wisconsin (part of the MISO grid) and the Georgia grid in Atlanta. That moves the power question from the site to the utility. WEC Energy Group, We Energies' parent, projects regional electric demand will rise about 45% over five years, driven mainly by Microsoft and another large data center campus at Port Washington, and advocates say We Energies plans to meet it with billions of dollars of new gas plants and pipelines. Wisconsin regulators require data centers to pay the full cost of new plants built to serve them. Microsoft says it is adding carbon-free electricity equal to its use, including a 250 MW solar project in Portage County.

Cooling across the three case studies

Fairwater's cooling is a closed loop. Microsoft says water is needed only once, to fill the system during construction, and is then recirculated without evaporation, supported by what it calls the second-largest water-cooled chiller plant in the world. Over 90% of its data center capacity uses this kind of system, according to Microsoft.

8. Community

Microsoft says about 550 full-time employees support the first Wisconsin facility, potentially rising to 800 once it is fully online. That is far below Foxconn's 13,000-job promise for the same area, but the investment is real: more than $7 billion announced so far. Local debate focuses on power bills and gas plants rather than on-site pollution: We Energies customers have already seen a cumulative 12.38% electric rate increase across 2025 and 2026, though not all of it is due to data centers.

โš–๏ธ Who pays for the power?Grid-connected AI factories avoid local turbine fights but push costs and emissions into the utility's plans. Wisconsin's rule that data centers pay for new plants built for them is one answer; whether it fully protects other customers is still debated.

9. Three bets compared

ColossusStargate AbileneFairwater
The betSpeedOutside capitalA network of sites
OwnerxAI / SpaceXJV, leased to Oracle, used by OpenAIMicrosoft
Size (Epoch est.)~946 MW~421 MWAtlanta ~636 MW; Wisconsin to ~2.26 GW
PowerGrid + turbinesGrid + gas microgridUtility grid only; no on-site generation in Atlanta
CoolingLiquid + chillersClosed loop, dry coolersClosed loop, large chiller plant
Main controversyTurbine permits, airTax breaks, delaysUtility gas plants, rates
The network is the computerSites linked by a private AI WAN
Density over area140 kW racks, two storeys
Grid, not turbinesCost moves to the utility
Slower, steadierOwned land, owned capital
VerdictFairwater is the hyperscaler's answer to the power ceiling: build big where the grid can cope, then wire the sites together. It is slower than Colossus and less leveraged than Abilene, but it is the only model so far that can, in principle, grow past what one site can power. The open question is how efficiently real training runs use two sites at once, and Microsoft has not published that. Next, Amazon's Project Rainier asks a different question: what if the chips are not NVIDIA's?

10. Reading the numbers

From Microsoft
315 acres, 3 buildings, 1.2M sq ft; $3.3B + ~$4B; "hundreds of thousands" of GB200; fiber to circle the Earth 4.5 times; 10ร— the fastest supercomputer (claim, no benchmark); ~140 kW/rack and 1,360 kW/row; 120,000+ new fiber miles; operational 23 Jun 2026; ~550 jobs rising to ~800.
Estimates
Atlanta ~636 MW and Wisconsin ~2.26 GW end state are Epoch AI estimates; Microsoft has not given megawatts.
Press / public bodies
$50M land purchase from Foxconn's area (2023); 15 more data centers approved (Jan 2026); WEC's ~45% demand growth; 12.38% rate increase; We Energies gas plans (advocates' description).
Illustrations
The training-split model, the two-storey cross-section and the map are simplified illustrations, not Microsoft data. The 670-mile distance is a straight-line approximation.