Microsoft Fairwater
Last reviewed
Jun 2, 2026
Sources
9 citations
Review status
Source-backed
Revision
v1 ยท 2,219 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 2, 2026
Sources
9 citations
Review status
Source-backed
Revision
v1 ยท 2,219 words
Add missing citations, update stale details, or suggest a clearer explanation.
Fairwater is Microsoft's name for a class of large AI datacenter built to train and serve frontier artificial intelligence models. The first Fairwater site, in Mount Pleasant, Wisconsin, was unveiled in September 2025; a second, in Atlanta, Georgia, came online in October 2025 and was announced that November. [1][2] Microsoft connects the sites with a dedicated wide-area fiber backbone it calls the "AI WAN" and describes the resulting distributed system, marketed for its Microsoft Azure cloud, as an "AI superfactory." [2][3] Each Fairwater datacenter is engineered as a single, flat-networked supercomputer holding hundreds of thousands of NVIDIA Blackwell GPUs rather than as a conventional cloud facility partitioned among many tenants. [1][2]
Microsoft introduced Fairwater on September 18, 2025, presenting the Wisconsin facility as "the largest and most sophisticated AI factory we've built yet." [1] Unlike a traditional data center designed to run many independent customer workloads, a Fairwater site is built to behave as one enormous machine: hundreds of thousands of tightly coupled GPUs that act together on a single training job. [1][2] To achieve this, Microsoft abandoned the usual practice of carving a building into separate network zones and instead used a single flat network spanning the whole site, so that any GPU can communicate with any other at high bandwidth. [1]
The company frames Fairwater as a repeatable design rather than a one-off. [4] After Wisconsin, a near-identical site was built outside Atlanta and brought online in October 2025, and Microsoft has said further Fairwater-class datacenters are planned, including hyperscale sites in Narvik, Norway and Loughton in the United Kingdom. [2][3] By linking these sites over the AI WAN, Microsoft aims to operate them collectively, distributing a single model-training run across geographically separated buildings. [2][3]
Microsoft uses the term "AI superfactory" to describe several Fairwater datacenters operating as one system rather than as isolated facilities. [2] In the company's framing, a conventional cloud datacenter serves a broad mix of workloads, whereas an AI superfactory is purpose-built to run a single, very large distributed job, training a frontier model, across multiple connected sites at the same time. [2][3]
The motivation Microsoft gives is that the largest models can no longer be trained inside a single building. Mark Russinovich, chief technology officer of Azure, has said that "the amount of infrastructure required now to train these models is not just one datacenter, not two, but multiples of that." [3] Alistair Speirs, a general manager for Azure infrastructure, described the goal as "building a distributed network that can act as a virtual supercomputer for tackling the world's biggest challenges in ways that you just could not do in a single facility." [3] By spreading a run across sites that work in tandem, Microsoft says models can be trained in weeks rather than the several months a single facility would require. [2][3] The marketing label "superfactory" is Microsoft's own; the underlying engineering claim is distributed, synchronized training across buildings.
The first Fairwater datacenter sits on a 315-acre site in Mount Pleasant, Wisconsin, and comprises three buildings with about 1.2 million square feet of combined floor area. [1][5] Microsoft has described the construction scale in concrete terms: roughly 46.6 miles of deep foundation piling, 26.5 million pounds of structural steel, 120 miles of medium-voltage underground cable, and 72.6 miles of mechanical piping. [1] The facility's storage systems alone span a length Microsoft likens to five football fields. [1]
Independent analysis indicates the Wisconsin campus will grow well beyond its initial footprint. Research group Epoch AI estimated that the multi-building site is projected to draw about 3.3 gigawatts of power once a fourth building becomes operational in late 2027, with each structure consuming on the order of one gigawatt. [6] For comparison, Epoch noted that the city of Los Angeles used roughly 2.4 gigawatts on average in 2023. [6] Microsoft's disclosed financial commitment in Wisconsin rose accordingly: an initial investment of about 3.3 billion US dollars, with a further roughly 4 billion dollars planned for a second datacenter, bringing the company's stated commitment in the state above 7 billion dollars. [5]
The second Fairwater datacenter, near Atlanta, Georgia, began operating in October 2025 and was announced on November 12, 2025. [2] Microsoft built it to the same architectural template as Wisconsin and connected it over the AI WAN to form what the company called the world's first "planet-scale AI superfactory." [2][7] The Atlanta and Wisconsin campuses sit roughly 700 miles apart. [2][4]
Press coverage highlighted that the Atlanta building omits the on-site uninterruptible power supplies and backup diesel generators common in traditional datacenters, part of a design optimized for AI training rather than for the always-on availability expected of general cloud services. [8] Microsoft characterizes the trade-off as delivering "four nines" of effective availability at "three nines" of cost. [7]
Microsoft has said the Fairwater pattern will be replicated internationally, naming planned hyperscale AI datacenters in Narvik, Norway and Loughton in the United Kingdom, alongside continued expansion of its broader Azure footprint. [3] The company tied this build-out to capital spending it put at more than 80 billion dollars for its 2025 fiscal year. [3]
Each Fairwater site is built around NVIDIA's rack-scale Blackwell systems, integrated into a single flat network so that hundreds of thousands of GPUs can be addressed as one supercomputer. [1][2] Microsoft states that the Wisconsin facility can deliver up to ten times the performance of the fastest supercomputer in operation at the time of its announcement. [1] The figures below are drawn from Microsoft's own technical descriptions.
| Attribute | Reported figure | Notes |
|---|---|---|
| GPU systems | NVIDIA GB200 NVL72 (Wisconsin); GB200 and GB300 (Atlanta) | Rack-scale Blackwell systems [1][7] |
| GPUs per rack | Up to 72 | One GB200 NVL72 rack [1][7] |
| GPU-to-GPU bandwidth | 1.8 TB per second per rack | Within-rack NVLink fabric [1][7] |
| Pooled memory | 14 TB per rack | Shared across the rack [1][7] |
| Scale-out connectivity | 800 Gbps GPU-to-GPU | Ethernet-based backend between racks [1][7] |
| Network topology | Single flat, non-blocking fat tree | Spans the whole site [1] |
| Power density | About 140 kW per rack; 1,360 kW per row | Atlanta figures [7] |
| Throughput (illustrative) | 865,000 tokens per second | Microsoft-cited figure for Wisconsin [1] |
| Relative performance | Up to 10x the fastest current supercomputer | Microsoft claim, Wisconsin [1] |
The networking is organized in two main tiers. Within a rack, GPUs are joined over NVIDIA's NVLink interconnect to act as a single large accelerator (the "scale-up" domain); across racks and pods, an Ethernet backend at 800 Gbps per GPU stitches the racks into a site-wide pool (the "scale-out" domain). [7] Microsoft runs this fabric on its open-source SONiC (Software for Open Networking in the Cloud) network operating system. [7]
To pack racks at roughly 140 kilowatts, far above air-cooled densities, Fairwater relies on direct liquid cooling. [7] Microsoft built a closed-loop system in which coolant is circulated to the GPUs, carried out of the building to be chilled, and returned, so that the same water is reused rather than continuously drawn and discharged. [1][2] The company says more than 90 percent of the Wisconsin facility's capacity uses this closed loop, with outside-air cooling employed only during the hottest periods, and that the loop is filled once and then recirculated for more than six years before the water chemistry requires replacement. [1][5][7] Microsoft characterizes the initial fill as roughly the annual water use of 20 homes and, separately, of a single restaurant, and describes the operational water waste as effectively zero. [3][5] It has also said the Wisconsin site uses the second-largest water-cooled chiller plant in the world and employs 172 large fans for heat recirculation. [1]
A second design choice is the two-story building layout used at the Fairwater sites. By stacking racks vertically as well as horizontally, Microsoft can place GPUs physically closer together in three dimensions, which shortens cabling and reduces signal latency between racks. [2][7] In the Atlanta build, Microsoft pushed this efficiency further by dropping on-site uninterruptible power supplies and backup generators, reflecting a facility tuned for training throughput rather than the continuous uptime guarantees of a general-purpose cloud region. [8]
The element that ties separate Fairwater sites into a single training system is the AI WAN, a dedicated long-haul optical network built specifically to carry AI training traffic between datacenters. [2][3] Microsoft describes it as private fiber that lets data move between sites "congestion-free, at nearly the speed of light," so that GPUs in different states can participate in the same synchronized job. [2][3] To support it, the company said it had deployed about 120,000 miles of new dedicated fiber across the United States over the prior year, increasing its total mileage by more than 25 percent in a single year. [2][3]
Microsoft's stated rationale is that modern AI infrastructure is increasingly limited by physics: the speed of light sets a floor on how quickly data can move between distant buildings, so the network is engineered to keep that latency low enough for distributed training to remain efficient. [2] Scott Guthrie, the executive vice president responsible for Microsoft's Cloud and AI group, framed the broader effort by saying that "leading in AI isn't just about adding more GPUs, it's about building the infrastructure that makes them work together as one system." [3]
Fairwater is built to train and run frontier AI models at a scale Microsoft says exceeds what any single building can support. [2][3] The capacity underpins Microsoft's Azure cloud and its AI products, and supports its long-standing partnership with OpenAI, whose models have been trained and served on Microsoft infrastructure. [1][2] Microsoft positions the superfactory as the way to meet what it calls unprecedented demand for AI compute and to push the frontier of model capability, by distributing a single large training run across multiple connected sites. [2]
Beyond training, the sites are provisioned for the full lifecycle of large-model work, including data processing, fine-tuning, evaluation, and serving. Microsoft cites supporting capacity such as millions of CPU cores for operational tasks, exabytes of storage, and an Azure Blob Storage layer sustaining more than two million read and write transactions per second at the Wisconsin site. [1][2]
Fairwater marks a shift in how the largest AI systems are built. Rather than scaling a single datacenter, Microsoft is coupling multiple purpose-built facilities into one distributed supercomputer, an approach the company presents as a repeatable blueprint for future sites. [2][4] The combination of site-wide flat networking, dense liquid-cooled Blackwell racks, and a dedicated inter-site optical backbone reflects a broader industry move toward treating geographically separated buildings as a single training fabric, paralleling comparably large efforts such as the Stargate datacenter program. [4] Microsoft's claim that distributed training can compress model build times from months to weeks, if borne out in practice, points toward multi-site clusters becoming the default substrate for frontier-scale training. [2][3]
The scale that makes Fairwater notable also draws scrutiny over energy and water. Epoch AI's estimate that the Wisconsin campus could reach roughly 3.3 gigawatts of demand by late 2027, more than the average power draw of Los Angeles and equivalent to several large nuclear reactors, illustrates the strain such facilities can place on regional grids. [6] Microsoft has said it is pre-paying for the energy and electrical infrastructure it will use in Wisconsin, which it argues will keep prices stable and shield local consumers from cost increases, but observers have noted the company has disclosed little about how the required power will be generated. [5]
Water has been a particular point of contention in Wisconsin. Microsoft emphasizes that Fairwater's closed-loop cooling consumes almost no water during operation after the initial fill. [3][5] Local advocacy group Clean Wisconsin has nonetheless argued that data center water consumption in the state will be considerably higher than companies' headline figures suggest, and that large AI facilities pose risks to surface and groundwater resources because of their substantial energy and cooling demands. [9] The tension between Microsoft's efficiency claims and independent estimates of resource use is likely to remain a live issue as additional Fairwater sites are built.