COMPUTE ATLAS Supercomputer supply-chain graph
1153 systems 482 sites Sourced data

Blog/2026-10-07

· Compute Atlas

Microsoft Fairwater: two sites, few numbers

Microsoft says Fairwater links Wisconsin and Atlanta with hundreds of thousands of Blackwell GPUs. Epoch AI estimates about 480,000 across the two, at 369 MW and 636 MW of IT power, but Microsoft has published no count or megawatts.

Fairwater is Microsoft’s name for a class of purpose-built AI datacenters, and the two sites that matter so far are in Mount Pleasant, Wisconsin and in the Atlanta area. Microsoft says the design uses NVIDIA GB200 and GB300 racks of up to 72 GPUs, liquid cooling, and a dedicated wide-area network that ties the sites together. It has also said Fairwater will hold hundreds of thousands of NVIDIA GPUs on a single flat network. What it has not published is an exact GPU count, a megawatt figure for either site, or a benchmark behind its claim of ten times the performance of the world’s fastest supercomputer. Epoch AI fills the gap with modelled estimates: 176,391 Blackwell GPUs and 369 MW of IT power in Wisconsin, and 304,279 GPUs and 636 MW in Atlanta. This article separates what Microsoft states from what Epoch infers. Our rows are Fairwater (Wisconsin) and Fairwater (Atlanta), both at the “reported” confidence tier.

How Fairwater came to exist

Wisconsin came first. Microsoft’s own completion notice dates the Mount Pleasant announcement to May 2024. Epoch’s satellite record has land clearing starting in July 2023, which suggests site work began before the public announcement. Microsoft unveiled the design in a September 2025 post by Scott Guthrie: 315 acres, three buildings totalling 1.2 million square feet, and on track to come online in early 2026, per the companion announcement.

Microsoft’s 23 June 2026 release, says equipment came online and started up in April 2026, and it declared the first datacenter complete and “fully operational” on 23 June, a little later than the early-2026 target. Brad Smith said Wisconsin is now home to the world’s most powerful supercomputer, a claim with no stated metric. The release says a second datacenter next door is under construction for 2028. The September 2025 post described that second building as of similar size and scale.

Atlanta followed on 12 November 2025, when Guthrie unveiled it as the next Fairwater site, connected to Wisconsin and to earlier supercomputers to form what Microsoft called the first planet-scale AI superfactory. Epoch dates the first two Atlanta buildings to 14 October 2025, and The Register reported the first node came online in October 2025. One location detail differs by source. Microsoft says Atlanta, Georgia. Epoch places the campus in Fayetteville, Georgia, and says it was built by QTS with Microsoft as the tenant. Microsoft’s posts do not mention a developer, so we treat the QTS role as Epoch’s finding.

Who pays and who uses it

The only dollar figures Microsoft gives are for Wisconsin. The September 2025 announcement put $3.3 billion into the first datacenter and a further $4 billion into a second, for more than $7 billion in total, and our capex row records the $7 billion as a floor covering both. The June 2026 release adds a different number: Microsoft estimates it will have spent $4.7 billion between 2024 and 2028 locally on hyperscale construction in Wisconsin. The two figures do not reconcile on their face, and the release does not say how the scopes differ. Atlanta has no stated figure.

Epoch’s cost estimates are far higher because they include the computers: about $14.0 billion for Wisconsin today and $24.1 billion for Atlanta, scaled from IT power. They are not disclosed spends, and the comparison with Microsoft’s construction-oriented figures is not like for like.

On who uses the compute, Microsoft’s posts describe the capacity for AI training and inference generally. Epoch lists OpenAI and Microsoft as “likely” users of both sites and says Atlanta likely provides OpenAI training compute. That is inference from outside, and we do not record a customer in our rows.

The hardware

Microsoft’s September description of the Wisconsin racks is 72 NVIDIA Blackwell GPUs per rack, 1.8 TB/s of GPU-to-GPU bandwidth and 14 TB of pooled memory. The Microsoft post says the figure applies to GB200 servers; the trade report on the June 2026 completion, from W.Media, likewise describes GB200 racks at 72 GPUs. Microsoft’s blog says GB200 and GB300 deployments exist globally and that the Fairwater network can integrate hundreds of thousands of GB200 and GB300 GPUs. The September post is more specific than the November one: it names GB200 for Wisconsin and says the planned Norway and UK datacenters will use GB300. The November post does not say which generation is installed in Atlanta, or how many racks. We therefore list GB200 NVL72 as the accelerator for both sites and do not claim GB300 for either.

The September post adds two claims worth reading carefully. It says each Wisconsin rack can process 865,000 tokens per second, which it calls the highest throughput of any cloud platform today. The figure is per rack, names no model, precision or batch setting, and is a marketing throughput number, not a benchmark we can reproduce. The same post describes a two-story layout in which racks connect to neighbours and to racks directly above or below, to cut cable length and latency. Epoch’s Wisconsin record does not test that, and neither can we.

Atlanta’s rack specification is better documented. Microsoft gives about 140 kW per rack and 1,360 kW per row, up to 72 GPUs per rack on NVLink, and a two-tier Ethernet back-end running at 800 Gbps per GPU. The Register, reporting the same launch, put rack power at over 120 kW and listed 720 petaFLOPS of sparse FP8 and 13 TB of HBM3e per rack. The rack figures differ because they come from different descriptions of what is being counted, and we use Microsoft’s own 140 kW.

The network vendor is disputed. Microsoft says its fabric runs on SONiC, its own switch operating system, on commodity Ethernet hardware, and its September post says Azure uses both InfiniBand and Ethernet fabrics at 800 Gbps in a fat tree. Data Center Frontier says Spectrum-X handles scale-out in Atlanta, while The Register notes Microsoft has not said what links the sites and lists candidates. We hold this as an open question and record no interconnect for Atlanta.

The wide-area piece is the distinctive part. Microsoft says it added more than 120,000 fiber miles across the US in the prior year for its AI WAN. Data Center Frontier gives the Atlanta to Wisconsin distance as about 700 miles and The Register as about 1,000 km, which agree. Whether jobs actually span both sites at once, and at what bandwidth, is not stated.

Power and cooling

Microsoft’s cooling story is about water use. In Wisconsin, over 90 percent of capacity uses closed-loop liquid cooling filled once and reused with no evaporation, with the remainder using outdoor air, and 172 twenty-foot fans rejecting heat. Microsoft says the site’s annual water use is roughly what a typical restaurant uses. In Atlanta, the initial fill is described as equal to what 20 homes use in a year, with the liquid replaced only if its chemistry requires it, and a design life of six-plus years. Epoch’s equipment counts line up: 172 air-cooled condenser units in Wisconsin, matching Microsoft’s 172 fans, and 672 air-cooled chillers at Atlanta with 794.9 MW of total capacity.

Power supply is where Atlanta is unusual. Microsoft says the site was chosen for resilient utility power capable of “4x9 availability at 3x9 cost” and that it can therefore forgo on-site generation, UPS systems and dual-corded distribution for the GPU fleet. That is a stated design choice, not an independent measurement. For Wisconsin, Microsoft says it is funding a 250 MW solar project in Portage County and has pledged to match every kilowatt hour from a fossil source one for one with carbon-free energy put back on the grid. That is a matching commitment, not a description of how the site is physically supplied, and Microsoft gives no load figure. It also says it is pre-paying for the energy and electrical infrastructure it will use.

The megawatt numbers are Epoch’s. Our dataset holds 369 MW IT power and 442 MW facility power for Wisconsin as of May 2026, and 636 MW IT and 859 MW facility for Atlanta as of June 2026, each marked third-party estimate. As a plausibility check, Microsoft’s 140 kW per rack and Epoch’s 636 MW imply roughly 4,500 racks, or about 327,000 GPUs at 72 per rack, against Epoch’s 304,279. That is our arithmetic, and it ignores networking and storage load, but the agreement at the order-of-magnitude level is reassuring about Epoch’s model and says nothing about Microsoft’s actual install.

Estimates against statements

ItemMicrosoft statesEpoch AI estimates
GPUs“Hundreds of thousands” across Fairwater176,391 (Wisconsin), 304,279 (Atlanta)
IT powerNot given369 MW (Wisconsin), 636 MW (Atlanta)
Wisconsin statusFully operational, 23 June 2026Building 1 operational April 2026; 2,263 MW projected by Q2 2028
Performance10x today’s fastest supercomputer8.82e20 8-bit ops/s at Wisconsin now
Cost$3.3 bn first datacenter, over $7 bn total$14.0 bn now, $85.7 bn projected

Adding Epoch’s two GPU counts gives about 480,000, which sits inside “hundreds of thousands”, but the counts are dated April and June 2026 and Epoch labels all of them B200-class. The ten-times claim cannot be tested. Microsoft gives no benchmark, and Epoch’s compute figure uses a different unit; our row for Wisconsin deliberately records no FLOPS. Epoch’s projection for Wisconsin is also much larger than anything Microsoft has committed to: it assumes buildings 3 to 11 operational by May 2028, against Microsoft’s two datacenters.

What to watch

Three things would sharpen the picture. First, the second Wisconsin datacenter, due in 2028 per Microsoft, and whether the other buildings Epoch projects ever get announced. Second, any Microsoft disclosure of GB300 at Fairwater, since the blog’s wording covers both generations. Third, the AI WAN: a statement of the inter-site bandwidth, or evidence of a training run that uses both sites, would show whether the “superfactory” is one machine or a pair.

What we could not confirm

  • An exact GPU count or power figure from Microsoft for either site. All counts and megawatts here are Epoch’s models.
  • Whether GB200 or GB300 is installed at each site, and in what proportion.
  • Which vendor supplies Atlanta’s scale-out network (SONiC-based Ethernet per Microsoft, Spectrum-X per Data Center Frontier).
  • The cause of the gap between 140 kW per rack (Microsoft) and over 120 kW (The Register).
  • Why Microsoft’s $4.7 billion local-spend figure sits below its more than $7 billion Wisconsin commitment.
  • The customer. Epoch says OpenAI and Microsoft are likely users; Microsoft’s pages do not name one.
  • Anything behind “ten times the performance of today’s fastest supercomputers”. It is a qualitative claim, not a FLOPS measurement.

Sources

deep-diveai-clusterspower-and-coolingaccelerators