The AI Boom Needs Power: Why Microsoft Is Helping Restart Three Mile Island
Microsoft is pouring billions into AI data centers packed with GPUs, high-speed networks and new cooling systems. Here in Pennsylvania, the scale of that expansion is helping give a retired nuclear reactor a second life.
Artificial intelligence can feel strangely detached from the physical world. We open Copilot, ChatGPT or another AI service, type a question and get an answer a few seconds later. There is no smoke stack outside the browser window and no obvious machinery turning in the background.
Behind that response, however, is a growing industrial operation.
Modern AI systems run inside enormous data centers containing thousands, and increasingly hundreds of thousands, of specialized processors. Those chips have to communicate at extremely high speeds, remain cool enough to operate and have electricity available every hour of every day. As companies race to build larger models and serve millions of users, the infrastructure supporting AI is becoming as important as the software itself.
Few companies illustrate that shift better than Microsoft.
The company is spending heavily to expand Azure and build a new generation of facilities designed specifically for artificial intelligence. Microsoft expects roughly $190 billion in capital expenditures during calendar year 2026, and during one recent quarter it added about one gigawatt of new computing capacity. Even at that pace, Microsoft told investors it expects computing capacity to remain constrained through at least 2026.
In Pennsylvania, one consequence of that demand can be seen along the Susquehanna River.
A nuclear reactor that shut down in 2019 is being prepared to generate electricity again, backed by a 20-year power purchase agreement with Microsoft.
A Different Kind of Data Center
Most traditional data centers are built to handle thousands or millions of separate jobs at once. One server might host a website. Another could run a database. Others might process email, store files or support a business application.
Large AI workloads work differently.
Training an advanced AI model can require huge numbers of processors to work together on the same problem. The processors must constantly exchange information, which means the network connecting them is almost as important as the processors themselves. A slow connection between thousands of fast GPUs would waste much of the computing power Microsoft is spending billions of dollars to install.
Microsoft's new Fairwater facilities show how radically the design is changing.
The company's Fairwater AI data center in Mount Pleasant, Wisconsin covers about 315 acres and contains roughly 1.2 million square feet of building space. Microsoft says the facility connects hundreds of thousands of NVIDIA GPUs through an extremely high-speed network so they can operate together as an enormous AI supercomputer.
Microsoft is building similar Fairwater facilities elsewhere and connecting them through Azure. A site in Wisconsin and another in Georgia, for example, can participate in the same distributed AI workload rather than functioning as completely isolated data centers.
Microsoft calls the larger system an AI superfactory.
The name may sound dramatic, but the architecture is easier to understand if you picture the difference between a normal office and a factory floor. Traditional cloud computing handles many unrelated jobs. These newer facilities can coordinate massive amounts of hardware around a common AI workload.
Inside them, AI models are trained, refined and tested. Once those models are deployed, the same infrastructure can help handle inference, which is the work performed every time a user or application asks the model to do something.
Microsoft says these systems support OpenAI, Microsoft AI, Copilot and other major AI workloads.
And people are asking AI systems to do a lot.
Microsoft Is Building for Demand It Still Cannot Fully Meet
Microsoft's spending helps put the scale into perspective.
In its fiscal 2026 third quarter alone, the company reported $31.9 billion in capital expenditures. Roughly two-thirds of that spending went toward shorter-lived assets, primarily GPUs and CPUs. The rest included longer-term infrastructure such as large data center sites.
For calendar year 2026, Microsoft now expects capital expenditures of roughly $190 billion.
That kind of spending would be extraordinary even if Microsoft were getting ahead of expected demand. It isn't.
The company says it is still capacity constrained.
Microsoft expects its overall data center footprint to roughly double over a two-year period. Earlier in fiscal 2026, the company said its total AI capacity was expected to increase by more than 80 percent during the year.
Fairwater in Wisconsin alone is expected to eventually scale toward two gigawatts.
This buildout supports far more than Microsoft's own AI assistant. Azure provides infrastructure for businesses, governments, developers, research organizations and AI companies around the world. Microsoft 365 Copilot runs on the same broader cloud ecosystem, as do many AI tools being added to enterprise software.
Then there is OpenAI.
As part of the evolving relationship between the two companies, Microsoft disclosed that OpenAI had contracted for an additional $250 billion in Azure services. That represents future cloud consumption rather than a $250 billion payment arriving all at once, but it demonstrates the amount of computing infrastructure Microsoft expects some of its largest AI customers to need.
Every layer of that expansion eventually leads to the same basic requirement.
The computers need power.
When a Technology Problem Becomes an Energy Problem
Processors have become much more efficient over time, but efficiency alone cannot offset unlimited growth.
If one generation of hardware uses electricity more efficiently but a company installs vastly more hardware, total power demand can still increase.
AI accelerators also pack enormous amounts of computing capability into a relatively small space. That creates another challenge: heat.
Microsoft's newer facilities use advanced liquid-cooling systems that move heat away from high-performance chips more efficiently than traditional air cooling. Networking has been redesigned. Server racks are becoming denser. Even the buildings themselves are being reconsidered to fit more computing power into a manageable footprint.
The result is a technology industry beginning to think in units once more familiar to utilities and power-plant operators.
Megawatts.
Gigawatts.
Transmission capacity.
Grid interconnections.
Power purchase agreements.
For Microsoft, securing enough computing capacity now involves securing enough energy to operate it.
Why Nuclear Power Fits Microsoft's Plans
Microsoft has already contracted large amounts of renewable energy around the world, including wind and solar. Those sources remain an important part of the company's clean-energy strategy.
Large data centers also need dependable power around the clock.
A cloud region cannot tell a hospital, bank or business that its applications need to wait for better weather. AI training jobs may run continuously for days or weeks, and inference services have to remain available whenever customers use them.
Nuclear power has an advantage in that environment. A nuclear generating station can produce a large amount of electricity continuously while generating very little direct carbon dioxide during operation.
Microsoft is openly looking at nuclear as part of its energy strategy. Its 2026 Environmental Sustainability Report specifically points to reliable clean energy for growing cloud and AI demand, including nuclear power and even longer-term technologies such as fusion.
Its agreement in Pennsylvania is one of the clearest examples of that strategy moving from discussion into infrastructure.
Why Three Mile Island?
The name still carries baggage, so it is worth clearing up an important point.
The reactor being restarted is not the reactor involved in the 1979 Three Mile Island accident.
Three Mile Island originally had two reactors. Unit 2 experienced the partial meltdown in March 1979 and never returned to service. Unit 1 was a separate reactor that continued operating for another four decades.
Unit 1 eventually shut down in September 2019 for economic reasons.
Constellation Energy, which owns the facility today, has renamed it the Christopher M. Crane Clean Energy Center and is now working to reverse that shutdown.
In September 2024, Constellation announced a 20-year power purchase agreement with Microsoft. Under the agreement, Microsoft will purchase energy from the renewed plant to help match the electricity consumed by its data centers within the PJM electric grid with carbon-free generation.
The reactor is expected to return approximately 835 megawatts of power to the regional grid.
Microsoft is not running a dedicated cable from the reactor directly into one of its data centers. Electricity enters the broader grid and is distributed throughout the region. The agreement gives Constellation a long-term buyer for the plant's output while giving Microsoft a large source of reliable carbon-free generation to offset the demand created by its data center operations in PJM territory.
That long-term commitment changed the economics of a plant that had been considered finished only a few years earlier.
The Restart Is Moving Forward
Bringing a retired nuclear reactor back into operation is not as simple as performing maintenance and turning it on again.
When Unit 1 closed, fuel was removed and the plant entered the decommissioning process. Systems that once supported an operating nuclear facility now have to be restored, inspected and approved for service again.
The Nuclear Regulatory Commission has created a dedicated restart panel to oversee the work.
On July 1, 2026, the NRC approved Constellation's request allowing the facility to receive and possess new nuclear fuel and new sealed neutron sources. NRC inspectors are also reviewing plant systems, security, licensing requirements and other areas that must be addressed before commercial operation can resume.
Constellation currently plans to return the reactor to service in 2027, although final timing remains dependent on regulatory reviews, inspections and the completion of plant work.
The federal government is involved financially as well. A Department of Energy loan guarantee supports up to $1 billion in financing for the restart, with Constellation funding the remaining project costs.
A plant that closed because it struggled economically in 2019 now has a major technology company committed to purchasing its electricity for two decades.
The plant did not suddenly discover a new purpose. Demand around it changed.
Pennsylvania Is Seeing the Other Side of the Data Center Boom
The Crane restart is only one part of a much larger infrastructure discussion now taking place across Pennsylvania.
On August 18, 2026, Governor Josh Shapiro signed an executive order placing new requirements on proposed data center developments in the Commonwealth. Pennsylvania says more than 100 potential projects have appeared in publicly sourced databases, although many remain speculative. Fifty-eight projects had engaged with the Department of Environmental Protection at some level, while only five had obtained all permits needed for their first phase.
The state's new requirements focus on electricity costs, environmental protection, workforce development, transparency and community involvement. AI data center projects have also been removed from Pennsylvania's expedited permitting program.
There are legitimate reasons for Pennsylvania to want the investment. Large technology projects can bring construction work, tax revenue, infrastructure improvements and high-paying technical jobs.
There are equally legitimate reasons to ask how quickly they should be built.
Electric-grid upgrades cost money. Large projects can consume significant electricity and water. Communities have to consider land use, transmission infrastructure and how much permanent employment remains after construction is finished.
Those questions become more important when data center proposals move from a handful of facilities to dozens of potential projects.
Microsoft Is Becoming an Infrastructure Company
Microsoft is still a software company, but describing it only that way no longer captures what the business has become.
Azure requires land, buildings, fiber networks, power infrastructure, cooling systems and enormous quantities of computing hardware. AI is accelerating every part of that requirement.
The company's newest data centers are being engineered almost like giant computers themselves. Microsoft is designing the networking between GPUs, developing its own processors, improving cooling systems and connecting facilities hundreds of miles apart so they can work on the same AI problems.
Now energy procurement is becoming another piece of that stack.
A shortage of GPUs can limit AI growth.
A shortage of data center space can limit AI growth.
A shortage of transmission capacity or dependable electricity can do exactly the same thing.
That puts utilities, electrical grids and power generation much closer to the center of the technology industry than they were a decade ago.
The York Tech Services Perspective
For those of us in Pennsylvania, the Microsoft and Three Mile Island story makes the physical side of AI unusually easy to see.
The AI boom is often discussed through new models, smarter assistants and faster processors. Those advances are important, but they represent only the visible end of a much larger system.
Behind the software are data centers that Microsoft is expanding at a pace that would have been difficult to imagine only a few years ago. Fairwater-style facilities can connect hundreds of thousands of GPUs. Microsoft added roughly a gigawatt of capacity in a single quarter and is preparing to spend around $190 billion on capital investments this year. Even after that expansion, the company says it still expects demand to exceed the capacity it can provide.
Follow that chain far enough and a 1970s nuclear reactor in Pennsylvania suddenly becomes relevant to the future of artificial intelligence.
That is what makes this story worth watching.
Three Mile Island Unit 1 closed in 2019 because the economics no longer worked. A few years later, Microsoft is helping create the long-term demand needed to bring it back.
The computers changed.
The demand changed.
Now parts of the power grid are changing with them.
Sources
- Microsoft, "Inside the world's most powerful AI datacenter" - Microsoft's detailed look at Fairwater, its 315-acre footprint, hundreds of thousands of GPUs and the company's global cloud infrastructure. Source
- Microsoft FY2026 Third Quarter Earnings Conference Call - Microsoft's current figures for capital spending, new gigawatt-scale capacity, infrastructure growth and ongoing capacity constraints. Source
- Microsoft FY2026 First Quarter Earnings Conference Call - Details on Fairwater's planned scale, growth in Microsoft AI capacity and OpenAI's additional Azure commitment. Source
- Microsoft, "Infinite Scale: The Architecture Behind the Azure AI Superfactory" - Technical overview of how Microsoft is connecting Fairwater facilities and hundreds of thousands of GPUs into distributed AI systems. Source
- Microsoft 2026 Environmental Sustainability Report - Microsoft's current clean-energy strategy and its explanation of how the Crane agreement supports reliable carbon-free power for growing data center demand. Source
- Constellation Energy, Crane Clean Energy Center Announcement - Original details of the 20-year Microsoft power purchase agreement and the planned return of approximately 835 MW to the grid. Source
- U.S. Nuclear Regulatory Commission, Christopher M. Crane Clean Energy Center - Current NRC licensing, inspection and restart status, including the July 2026 approval involving new nuclear fuel. Source
- Commonwealth of Pennsylvania, Executive Order on Data Center Development, August 18, 2026 - Current Pennsylvania requirements for proposed data centers and state figures on the number of projects under discussion. Source
Key Takeaways
- Microsoft's global cloud includes more than 400 data centers across 70 regions, and its infrastructure footprint continues to expand rapidly.
- Microsoft expects roughly $190 billion in capital expenditures during calendar year 2026, much of it tied to cloud and AI infrastructure.
- The company added roughly one gigawatt of computing capacity in a single recent quarter and still expects to remain capacity constrained through 2026.
- Microsoft's Fairwater AI facilities are designed to connect hundreds of thousands of GPUs into massive distributed AI systems.
- OpenAI has contracted for an additional $250 billion in Azure services, adding to Microsoft's long-term demand for AI computing capacity.
- Microsoft signed a 20-year power purchase agreement with Constellation supporting the restart of the former Three Mile Island Unit 1.
- The reactor, now named the Christopher M. Crane Clean Energy Center, is expected to return approximately 835 megawatts of power to the PJM grid.
- The reactor involved in the 1979 accident was Unit 2, not the Unit 1 reactor being restarted.
- The NRC approved the facility to receive and possess new nuclear fuel in July 2026, and Constellation is targeting a 2027 restart.
- Pennsylvania is simultaneously imposing stricter requirements on proposed data centers as officials weigh new investment against energy, environmental and community concerns.