Microsoft confirmed June 23, 2026, that its Fairwater campus in Mount Pleasant, Wisconsin, is fully operational — and the engineering behind it makes the facility something fundamentally different from every data center that came before it. Where conventional cloud infrastructure racks up general-purpose servers and parcels out workloads to each one independently, Fairwater links hundreds of thousands of NVIDIA GB200 Blackwell GPUs into a single, coherent cluster using a two-story building design, 800-gigabit-per-second Ethernet fabric, and a proprietary networking protocol co-developed with OpenAI and NVIDIA. The result, according to Microsoft, is the closest thing to a purpose-built AI supercomputer that any company has ever placed in commercial operation — and the most consequential question about it may not be how it was built, but who will pay for the power that keeps it running.

The $7.3 billion Wisconsin investment marks the fulfillment of a commitment Microsoft made in May 2024, when the company pledged $3.3 billion to build an AI campus on land in Racine County that had been set aside — and largely abandoned — for a Foxconn LCD manufacturing facility that never materialized. Microsoft Source confirmed the campus reached full operational status after equipment came online in April and startup activities followed. Microsoft broke ground in 2023 and expanded the commitment to $7.3 billion in September 2025, adding a second facility of similar scale now under active construction next door. About 10,000 construction workers built the first facility over roughly two years; approximately 550 full-time employees are currently on site.

Two Stories, 800G Ethernet, and a Protocol Called MRC

The most technically significant choice Fairwater makes is one that is invisible to anyone standing outside the building: it does not use a conventional single-story warehouse layout.

Each Fairwater building uses two stories of GPU racks with through-floor networking connecting racks above and below. The reason is physics. AI training at this scale requires thousands of GPUs to exchange data continuously — every chip needs to send its computation results to every other chip, then update the model together. Any added distance in the cables carrying those signals introduces latency; enough accumulated latency and GPUs sit idle waiting for others to finish, wasting compute. By stacking racks vertically and running cables between floors, Microsoft shortened the physical path between chips that would otherwise be hundreds of meters apart on a single warehouse floor.

Connecting the racks is a two-level networking tree built on 800G Ethernet — a standard that carries data at 800 gigabits per second — using a protocol Microsoft developed alongside OpenAI and NVIDIA called Multi-Path Reliable Connected (MRC). Microsoft’s AI superfactory technical overview describes MRC as a software-layer innovation on top of commodity Ethernet hardware that provides advanced congestion control, rapid packet detection and retransmission, and agile load balancing across multiple simultaneous network paths. The practical effect is that Fairwater can use lower-cost commodity switches rather than proprietary networking gear while still achieving the deterministic, low-latency performance that AI training at this scale requires.

Within each rack, NVIDIA’s GB200 NVL72 systems connect 72 Blackwell GPUs through NVLink 5.0, a fifth-generation interconnect fabric running at 1.8 terabytes per second of GPU-to-GPU bandwidth per chip. That rack-level fabric makes all 72 chips behave as a single accelerator with pooled memory. The 800G Ethernet MRC layer then connects thousands of those racks together across the building — and, through Microsoft’s AI Wide Area Network (AI WAN), across Fairwater campuses in different states.

The AI WAN is a dedicated optical fiber backbone that Microsoft built specifically to let geographically separated Fairwater sites participate in the same training job. The company deployed more than 120,000 new fiber miles across the United States to support this network — a 25% expansion of its overall fiber footprint in a single year. A Wisconsin-based model training job can now draw on compute at both the Wisconsin and Atlanta Fairwater sites simultaneously, completing in weeks what previously would have taken months across isolated data centers.

The Cooling Architecture That Defines the New Standard

If the networking is Fairwater’s brain, its cooling system is the design choice that most dramatically separates this generation of AI infrastructure from what preceded it.

Air cooling is effectively obsolete for GPU clusters at this density. A single GB200 NVL72 rack draws approximately 140 kilowatts of power — ten times what a standard server rack drew a decade ago — and releases all of that energy as heat. Fairwater’s response is a closed-loop water cooling system whose primary innovation is that the water in it is filled once at construction and never replenished: there is no evaporation, no cooling tower, and no ongoing water draw. The circulating water absorbs heat at the chips, travels to external cooling fins on each side of the data center building, and is chilled by 172 fans before returning to the racks. Microsoft says more than 90% of the facility’s compute capacity uses this system, with the remaining 10% — conventional servers — cooled by outdoor air except on the hottest days.

The scale of the cooling infrastructure matches the scale of the compute: Microsoft’s technical documentation describes Fairwater as supported by the second-largest water-cooled chiller plant on the planet. The environmental case for this design is direct. A traditional evaporative cooling data center of this power density would consume millions of gallons of water annually. Microsoft CEO Satya Nadella has said Fairwater’s closed-loop system uses the equivalent of what a single restaurant consumes per year in ongoing operations — a company claim consistent with what closed-loop cooling engineering achieves but not yet independently audited.

What Microsoft Claims About Performance — and What the Evidence Actually Shows

Brad Smith, Microsoft’s vice chair and president, declared at the June 23 announcement that “Wisconsin is now home to the world’s most powerful supercomputer.” That claim requires careful context.

Microsoft’s assertion is based on AI training throughput — specifically, the rate at which Fairwater can process the massive matrix computations involved in training large language models. On that metric, the company says Fairwater delivers roughly ten times the performance of any existing system. The claim has not been independently verified by a third party, and no benchmark submission to the TOP500 — the authoritative twice-yearly list of the world’s most powerful supercomputers — has been filed for Fairwater.

The timing of the announcement gives the “most powerful” claim an additional complication: the June 2026 TOP500 list, published the day after Microsoft’s announcement, placed China’s LineShine system at number one with 2.198 exaflops on the High Performance Linpack (HPL) benchmark. Racine County Eye, which is tracking the Fairwater campus independently, has asked Microsoft for a response to the supercomputer claim and has not yet received one.

The discrepancy is not a contradiction so much as a measurement problem. HPL measures dense linear algebra floating-point operations and was designed for scientific high-performance computing — weather modeling, nuclear simulations, protein folding. AI training throughput is a different workload with different characteristics, and the two metrics are not directly commensurable. Jack Dongarra, an emeritus professor at the University of Tennessee and one of the organizers of the TOP500 list, said of the rankings that they assess “one benchmark” and should not be viewed as a “complete measure of technological leadership.” Microsoft’s claim belongs in that same category: it is a company-reported figure for a specific AI workload class, not a broadly verified system-level ranking.

What is verifiable is that Fairwater represents a genuinely new category of AI compute infrastructure — a system designed from the building up for the specific demands of frontier model training, with every architectural choice from floor plan to network protocol optimized for that single purpose.

Wisconsin as a Hub — and a Template

The completion of Fairwater anchors a much larger investment thesis in Wisconsin and signals a geographic shift in where AI compute gets built.

Alongside the second Fairwater facility now under construction in Mount Pleasant — a $4 billion project expected to be completed in 2028 — Meta Platforms is building a $1 billion data center in Beaver Dam, and Vantage Data Centers is constructing a $15 billion campus in Port Washington. The combination of available land, moderate climate, and proximity to Chicago’s fiber infrastructure has made Wisconsin’s I-94 corridor an emerging rival to Northern Virginia and Texas as an AI infrastructure hub.

Microsoft has made accompanying community investments alongside the hardware: a Datacenter Academy at Gateway Technical College to train workers for data center operations roles, an AI Co-Innovation Lab at the University of Wisconsin-Milwaukee, and ecological restoration projects in Racine and Kenosha Counties through the Root-Pike Watershed Initiative Network. The company also deployed 120,000 miles of dedicated fiber for its AI WAN network, increasing its overall fiber footprint by more than 25% in a single year.

The Fairwater design is also being replicated internationally, with Fairwater-class campuses under development in Loughton, UK, and Narvik, Norway — the latter backed by hydroelectric energy.

👉 Read more: 318651 NVIDIA Blackwell deployments at Fairwater

What Wisconsin Ratepayers Should Know About the Energy Cost

Fairwater’s engineering achievements do not insulate Wisconsin residents from its energy consequences. According to a Microsoft consultant’s testimony filed with the Wisconsin Public Service Commission (PSC), Fairwater is expected to become the largest single electric load ever served by Wisconsin Electric, the We Energies subsidiary that serves 1.1 million customers. WEC Energy Group, We Energies’ parent company, projects regional electric demand will rise approximately 45% over five years as a direct result of the Microsoft campus and the Vantage facility in Port Washington.

We Energies customers have already absorbed a cumulative 12.38% electric rate increase across 2025 and 2026, driven primarily by grid investments and renewable energy construction. A further increase of approximately $13 per month is projected for 2027. The rate increases are not yet directly attributed to data center demand — the PSC has noted the increases are driven by broader grid and generation investments — but the data center buildout is happening against a backdrop of already-climbing bills and deepening public concern.

In April 2026, the Wisconsin PSC unanimously approved a first-of-its-kind rate structure in the Midwest requiring large data center customers — any customer drawing more than 100 megawatts at peak demand — to cover the full cost of new power generation and transmission infrastructure built specifically to serve them. The rule was described by PSC Commissioner Kristy Nieto as having “the potential to fundamentally reshape the utility system.” Tom Content, executive director of the Citizens Utility Board of Wisconsin, welcomed the decision as a signal that commissioners “heard loud and clear that Wisconsinites have significant concerns about energy affordability and AI data centers” — while noting that implementation details in future rate cases remain unresolved.

The PSC decision was itself partly a response to community friction that accompanied the Fairwater buildout. When Fairwater’s cooling fans came online in April 2026, residents of Sturtevant — a village directly north of the campus that was not consulted before the project was approved and had no vote on it — reported a persistent humming audible inside their homes with windows closed. Mike Rosenbaum, Sturtevant’s village president, told Racine County Eye it was something that “once you hear it, you can’t unhear it.” Microsoft acknowledged the noise was unanticipated, adjusted fan speeds, and installed sound-reduction components. The company says the issue is resolved. The governance gap Sturtevant represents — a neighboring community with no formal voice in a project that directly affects it — remains unaddressed as 15 more Microsoft buildings move forward on two new campuses.

Why Every State Watching the AI Buildout Is Watching Wisconsin

Wisconsin’s experience with Fairwater has become a policy case study precisely because the state had to develop its regulatory response in real time, without a prior model to follow.

The PSC’s April 2026 full-cost-recovery decision — requiring any data center customer above 100 megawatts to subscribe to enough generation capacity to cover peak demand, and to pay 100% of infrastructure costs even if they exit early — was among the first such rules adopted anywhere in the Midwest. It means that Microsoft, not Wisconsin’s existing We Energies customers, is required to pay for the new power plants and solar farms needed to serve Fairwater. Whether that protection extends fully to future transmission infrastructure upgrades remains, as Tom Content noted, an open question.

The Wisconsin precedent is now being watched in states across the country where hyperscale campuses are either under construction or in proposal stages. The question of who pays for the grid upgrade that an AI factory demands — the company that needs it, or the ratepayers who already live there — is rapidly becoming one of the most consequential infrastructure policy debates of the decade.

Frequently Asked Questions

What makes the Fairwater data center different from a conventional cloud data center?

Conventional cloud data centers are designed to run many smaller, independent workloads at once — websites, email, business applications. Fairwater is purpose-built as a single AI supercomputer: every GPU in the facility is connected to every other GPU through a two-story rack layout and an 800G Ethernet network running a custom congestion-control protocol called MRC, so the entire facility functions as one unified training machine. General-purpose cloud infrastructure cannot support the synchronous, high-bandwidth GPU communication that training frontier AI models at this scale requires.

How does Fairwater’s cooling system use near-zero water?

Fairwater uses a closed-loop liquid cooling system that circulates water through sealed piping from the server racks to external cooling fins on each side of the building. The water absorbs heat at the chips, travels to the fins, is chilled by 172 fans, and recirculates. Because the loop is sealed, there is no evaporation and no ongoing water draw — the system is filled once at construction and reused indefinitely. Over 90% of Fairwater’s compute capacity uses this system. The remaining 10% uses outdoor air, switching to water only during peak summer heat. This contrasts with traditional evaporative cooling towers, which can consume millions of gallons of water per year for a facility at this scale.

Will Wisconsin electricity bills increase because of the Fairwater data center?

We Energies customers have already seen a cumulative 12.38% electric rate increase across 2025 and 2026, and the utility projects a further $13-per-month increase in 2027. Those increases are driven primarily by grid and renewable energy investments, not yet by direct data center demand. However, Wisconsin’s Public Service Commission voted in April 2026 to require large data center customers drawing more than 100 megawatts to pay the full cost of any new generation infrastructure built to serve them — a rule designed to prevent future rate increases from being shifted onto existing customers. How effectively that protection covers all future grid investments remains a subject of active regulatory scrutiny.

Is Fairwater really the world’s most powerful supercomputer?

Microsoft’s Brad Smith made that claim at the June 23 announcement, and Microsoft’s published materials say Fairwater delivers roughly 10 times the performance of existing supercomputers for AI training and inference workloads. The claim is based on AI training throughput, not the High Performance Linpack benchmark used by the TOP500 list, and has not been independently verified. The June 2026 TOP500 list, published one day after Microsoft’s announcement, placed China’s LineShine system at number one with 2.198 exaflops. The two rankings measure fundamentally different things: HPL tests dense scientific computing; Microsoft’s claim targets AI training throughput. Racine County Eye has asked Microsoft for a response and has not yet received one.