Microsoft’s Three Mile Island Deal Marks Shift: Energy Now Limits AI Scaling, Not Chips
As data center electricity demand surged 17% in 2025 alone, hyperscalers are bypassing utilities and financing nuclear restarts to secure the baseload power that frontier models require.
Microsoft’s 20-year agreement to purchase the entire output of Three Mile Island’s Unit 1 reactor marks the point where energy availability, not semiconductor supply, became the binding constraint on AI development. The facility, offline since 1979, will restart by 2027—a year ahead of the original 2028 target—backed by $1.6 billion from Constellation Energy and a $1 billion federal loan guarantee issued in November 2025. The accelerated timeline reflects computational scaling demands that have outstripped traditional grid infrastructure.
From Chip Bottleneck to Grid Bottleneck
Global data center electricity demand is projected to more than double from 415 TWh in 2024 to 945 TWh by 2030, according to the International Energy Agency. That trajectory accelerated sharply in 2025, when demand jumped 17% in a single year. US facilities alone consumed 183 TWh in 2024—4.4% of total electricity generation—and are on track to reach 426 TWh by 2030, per Pew Research analysis.
The constraint is no longer access to H100 chips or fabrication capacity. Anthropic estimates that training a single frontier model will require 5 GW of dedicated power by 2027, Brookings Institution noted in early 2026 analysis. At that scale, securing reliable baseload supply becomes a precondition for computational scaling—and traditional utility interconnection queues spanning 5-7 years cannot match hyperscaler deployment cycles of 2-5 years.
“Powering industries critical to our nation’s global economic and technological competitiveness, including Data Centers, requires an abundance of Energy that is carbon-free and reliable every hour of every day, and Nuclear plants are the only energy sources that can consistently deliver on that promise.”
— Joe Dominguez, CEO, Constellation Energy
Hyperscalers Become Infrastructure Investors
Microsoft’s arrangement with Constellation is not a standard power purchase agreement. The tech giant is effectively underwriting the entire restart—securing dedicated output from a facility that will generate enough electricity to power approximately 800,000 homes. That model is proliferating across the sector. Meta announced 6.6 GW of nuclear capacity agreements in January 2026, partnering with TerraPower, Oklo, and Vistra to bring reactors online by 2035, Meta confirmed. Amazon signed a 1.92 GW deal with Talen Energy for output from the Susquehanna facility, while Google is restarting Iowa’s Duane Arnold plant, according to Perkins Coie analysis.
These are not passive offtake contracts. Hyperscalers are financing and building their own on-site generation behind the meter, bypassing the traditional grid entirely, Latitude Media reported. The shift reflects both operational necessity—avoiding interconnection delays—and a strategic pivot toward controlling energy supply chains as directly as semiconductor procurement.
Grid Economics Under Strain
The concentration of hyperscale load is reshaping wholesale electricity markets. Wholesale costs in areas near US data centers increased 267% compared to five years prior, as of September 2025, Bloomberg analysis showed. In the PJM market covering 13 states, data center demand drove a $9.3 billion price increase in the 2025-26 capacity auction, translating to average residential bill increases of $18 per month in Maryland and $16 per month in Ohio.
Those price signals are accelerating behind-the-meter strategies. Tech companies’ capital expenditure surged past $400 billion in 2025 and is set to increase by 75% in 2026, the International Energy Agency reported. A significant portion of that spending is now directed toward energy Infrastructure—nuclear restarts, small modular reactors, and natural gas peaker plants—rather than servers or networking equipment.
Three Mile Island’s Unit 1 reactor operated safely until its retirement in 2019 for economic reasons. Unit 2, which experienced a partial meltdown in 1979, remains permanently shut down. The facility’s reactivation will mark only the second time a shuttered US commercial reactor has been brought back online, following Michigan’s Palisades plant expected later this year.
Nuclear Renaissance Driven by AI Economics
The Three Mile Island restart would have been economically unviable under traditional utility models. Wholesale power prices in 2019—when Constellation shuttered the reactor—did not support the operational costs of a single-unit facility competing against subsidised renewables and cheap natural gas. Microsoft’s 20-year commitment provides revenue certainty that changes the investment calculus entirely.
That dynamic is eroding regulatory barriers that have stood for decades. The Trump administration’s $1 billion loan guarantee and executive orders streamlining nuclear permitting reflect recognition that energy availability is now a national competitiveness issue, Fox News reported. Constellation CEO Joe Dominguez framed the decision as “the most powerful symbol of the rebirth of nuclear power as a clean and reliable energy resource,” per CNBC.
| Company | Capacity | Partners/Facilities | Timeline |
|---|---|---|---|
| Microsoft | ~835 MW | Three Mile Island (Constellation) | 2027 |
| Meta | 6.6 GW | TerraPower, Oklo, Vistra | By 2035 |
| Amazon | 1.92 GW | Susquehanna (Talen Energy) | TBD |
| ~600 MW | Duane Arnold (Iowa) | TBD |
Staffing for the restart is proceeding ahead of schedule, with the facility more than 65% staffed as of mid-2026, Pennsylvania Capital-Star reported. Many former workers have returned, drawn by the facility’s resurrection after seven years offline.
What to Watch
Three Mile Island’s restart timeline will test whether nuclear infrastructure can be accelerated to match AI deployment cycles. If Constellation meets the 2027 target, it establishes a precedent for rapid reactivation that could apply to other retired facilities with intact infrastructure. Gartner projects AI-optimised server electricity usage will nearly quintuple from 93 TWh in 2025 to 432 TWh by 2030—a pace that will require multiple gigawatts of new baseload capacity annually.
The regulatory environment remains in flux. Federal Energy Regulatory Commission Chair Laura Swett noted in March 2026 that hyperscalers “don’t speak FERC” and show “a lack of understanding of how the utilities normally function,” per Fortune. That tension will intensify as tech companies push for behind-the-meter arrangements that bypass traditional rate regulation and grid cost allocation.
The next inflection point arrives when a hyperscaler cannot secure sufficient power to train a planned frontier model. At current growth rates and infrastructure timelines, that constraint could emerge within 18-24 months—forcing either a slowdown in model scaling or further acceleration of direct energy investment. Energy Secretary Chris Wright’s comment that “AI is going to be transformative technology, but we can’t be second place” signals federal willingness to expedite permitting. Whether that translates to faster timelines for new nuclear, or a temporary reliance on natural gas generation, will determine which companies can sustain computational scaling through 2028.