The Three Mile Island Restart Is Ahead of Schedule. So Is the Rest of Big Tech's Nuclear Bet.

Three Mile Island Unit 1 is coming back online a year ahead of schedule. Constellation Energy confirmed earlier this year that the reactor, rebranded as the Crane Clean Energy Center, is now targeting grid synchronization in 2027 rather than the 2028 date announced when Microsoft signed a 20-year offtake deal in September 2024. The project is roughly 80% staffed, with more than 500 employees on-site, and the NRC license amendment hearing notice went up in the Federal Register on February 24.

The Three Mile Island restart is the flagship of a larger bet. Microsoft, Amazon, Google, and Meta have collectively committed more than 10 gigawatts of nuclear power to AI data centers, and the pace accelerated sharply in the first four months of 2026. Meta alone announced a 6.6 GW portfolio across Constellation, TerraPower, Oklo, and Commonwealth Fusion in early January. Amazon inked 1.9 GW with Talen Energy at Susquehanna. Google has Kairos Power building small modular reactors for Tennessee and Alabama deliveries through the mid-2030s.

The headlines read like a clean-power triumph. The grid-level math is messier. The International Energy Agency projects global data center electricity consumption to hit 1,100 TWh in 2026, equivalent to Japan's entire national demand. All of the nuclear deals announced so far, combined, do not close that gap.

This piece walks through what the restart actually looks like in April 2026, why every other hyperscaler made similar bets within eighteen months of Microsoft's, and what still has to go right before any of this electricity flows.

What the Restart Looks Like in April 2026

The Three Mile Island restart is further along than the public timeline suggests, which is exactly why the opposition is showing up now.

The plant that most of the public still calls Three Mile Island has a new name: the Crane Clean Energy Center, after Chris Crane, the late CEO of Exelon, Constellation's former parent company. The restart cost is roughly $1.6 billion. A $1 billion Department of Energy loan, granted in late 2025 to fast-track "domestic AI energy security," underwrites the federal share. The reactor, rated at approximately 835 MW, is fully contracted to Microsoft for twenty years.

As of early 2026, the project is running ahead of its original 2028 schedule. Constellation is targeting 2027 for grid synchronization. On-site staffing sits around 80% of plan, with more than 500 employees now working through the restart checklist. The engineering work underway covers cooling system restoration, control room modernization, and replacement of large primary-side components including the main power transformers.

The NRC review is the rate-limiting step. Constellation has filed three license amendments covering the facility license and permanently-defueled technical specifications, the radiological emergency preparedness plan, and the physical security plan. The NRC hearing notice was published in the Federal Register on February 24, 2026. Public comments were due March 26. The deadline for intervention petitions is April 27, 2026, which means the window for formal opposition is closing the same week as this piece.

One persistent source of public confusion is worth clearing up. TMI Unit 1, the reactor being restarted, operated commercially from 1974 to 2019 and was shut down for economic reasons, not safety. Cheap natural gas made the unit uncompetitive. TMI Unit 2 is the 1979 partial meltdown reactor. It is a different physical reactor on the same site, and it is being decommissioned separately, with the Department of Energy's Idaho Operations Office handling spent fuel storage under a separate Federal Register notice filed in February. Most of the "zombie reactor" framing in coverage of the restart is actually drawing on Unit 2's history to describe Unit 1's future. The distinction matters because Unit 1 has an operational safety record. Unit 2 is the one in the cautionary tale.

Local opposition is concentrated but experienced. Eric Epstein, a longtime anti-nuclear advocate who runs TMI Alert, has been the most quoted critic. His characterization of the NRC, "a captured agency, an appendage of the industry," is the sharpest public skepticism on record. The intervention filing window closing in April is the moment when organized opposition either formalizes or doesn't. Whether it slips the 2027 schedule depends less on the opposition's arguments and more on the commission's willingness to expedite.

The Rest of Big Tech Is Already Doing the Same Thing

The Microsoft deal was a first mover. Eighteen months later, every major hyperscaler has a nuclear portfolio, and Meta's is the biggest of them.

Amazon moved second and moved hard. The company has a 1.9 GW PPA through 2042 from Talen Energy's Susquehanna nuclear plant in Pennsylvania, with additional SMR deals signed with Energy Northwest and X-Energy, plus a separate Virginia SMR exploration agreement with Dominion Energy. The Talen deal is the most visible because it's colocated: Amazon is building data centers directly on the Susquehanna site rather than pulling power over the broader PJM grid. That architecture made Amazon the test case for the FERC co-location proceedings that concluded in December 2025.

Google signed with Kairos Power to take 500 MW of small modular reactor capacity across six or seven reactors to be delivered between 2030 and 2035, with sites in Tennessee and Alabama. The company has separate agreements with NextEra for additional nuclear capacity. Google's approach is the most timeline-distant of the four hyperscalers, reflecting both its confidence in the Kairos design and its longer-dated AI capacity planning.

Meta's 6.6 GW nuclear portfolio announcement in early January 2026 is the largest single-quarter move by any hyperscaler. The portfolio covers Constellation's Clinton plant in Illinois, TerraPower (the Bill Gates venture pursuing advanced reactor designs), Oklo (Sam Altman's small-reactor company), and Commonwealth Fusion. Taken together, Meta's package is roughly the size of eight Three Mile Islands. It is also the clearest validation to date that capital is flowing to the SMR and advanced-reactor tier, not just to existing reactor restarts.

Microsoft's position has widened beyond TMI. The company's Nucleation program, its internal nuclear partnerships initiative, has added SMR agreements alongside the Crane PPA. Oracle, which previously sat outside the hyperscaler nuclear conversation, announced SMR commitments as part of its $500 billion AI infrastructure plan. Every major US AI infrastructure buyer now has a nuclear line item on its capital plan.

The shape of these deals is converging. Twenty-year PPAs are becoming the standard term. Offtake pricing is typically fixed with inflation escalators. Hyperscalers are increasingly taking on delivery risk, writing penalty clauses into contracts when reactors miss commercial operation dates, because the alternative, waiting for a reactor vendor to take that risk, has historically meant deals that never close. The result is that reactor vendors like Kairos, Oklo, and TerraPower are effectively being capitalized by hyperscaler PPAs. That is a structural shift in how nuclear gets built.

The Energy Demand Behind All of It

AI data center electricity demand has gone from "fast-growing" to "grid-restructuring" in eighteen months, and the numbers keep getting revised upward.

The IEA's energy demand projection for AI puts global data center electricity consumption at roughly 1,100 TWh in 2026. That is 18% higher than the agency's December 2025 estimate, and it is equivalent to Japan's total national electricity consumption. US data center demand alone is projected at 260 TWh in 2026, which works out to roughly 10% of the US grid in the first quarter, up from about 4% in 2023. European data center demand sits around 150 TWh.

The mix inside those numbers matters. Accelerated servers, the GPU-heavy infrastructure that runs AI workloads, are the dominant growth driver. IEA's Base Case has accelerated-server electricity consumption growing roughly 30% annually, versus about 9% for conventional servers. By 2030, the same report projects global data center consumption to roughly double from 2024 levels to 945 TWh, reaching just under 3% of global electricity.

The reason nuclear specifically, rather than any other clean generation, keeps showing up in hyperscaler deals has to do with the load profile. Frontier AI model training is a constant-draw workload that runs for weeks, sometimes months, at close to peak capacity. It is not load-following, and it does not tolerate intermittency well without substantial overbuild and storage. Natural gas peakers do not fit the constant-draw profile at the scale hyperscalers need. Utility-scale solar-plus-battery at the gigawatt level requires land and interconnection queue position that most data center builders do not have. Nuclear is the generation type that best matches the load shape.

That said, the accounting of these deals is not clean. Carbon Brief and other independent analysts have noted that some of the hyperscaler nuclear PPAs represent shifted demand rather than strictly new clean generation. When a reactor that was going to keep running gets reassigned from the grid to a single data center customer, the grid still has to replace that capacity with something else, often gas. The environmental balance sheet depends on whether the new PPA displaces fossil capacity or simply relocates existing capacity. That distinction rarely makes it into the press releases.

The Regulatory Plumbing Nobody Wants to Read

The Three Mile Island restart became feasible in its current form because of one federal ruling in December 2025 that most people outside the grid business haven't noticed.

The Federal Energy Regulatory Commission issued an order on December 18, 2025 directing PJM Interconnection, which runs the grid covering Pennsylvania and twelve other states, to create new rules for data center colocation at power plants. The FERC ruling on co-location established three new transmission service categories: Firm Contract Demand, Non-Firm Contract Demand, and Interim Non-Firm. It also overhauled PJM's Behind-the-Meter Generation rules, which previously treated colocated load ambiguously.

Prior to the order, deals like Amazon's direct colocation at Susquehanna sat in regulatory limbo. After it, there is a defined legal path for data-center-to-reactor arrangements, whether colocated or grid-connected. PJM's initial compliance filing was due January 17, 2026; the balance of tariff revisions and PJM's initial brief were due February 16. The rules are now in the system. The Crane PPA structure, Amazon-Susquehanna, and every future hyperscaler-reactor deal that involves direct generation coupling is shaped by this order.

The federal policy layer is running in parallel. Trump administration executive orders have designated AI data centers as "critical defense facilities," which gives the Department of Energy authority to expedite reactor siting. Three advanced reactor pilots are targeting criticality by July 4, 2026, as a public showcase. Four federal sites, Idaho National Lab, Oak Ridge Reservation, Paducah Gaseous Diffusion Plant, and Savannah River, have been selected for public-private data center plus nuclear development projects that bypass state and local permitting friction by using federal land.

Energy Secretary Christopher Wright has framed the combined build as "a new Manhattan Project." That framing is politically useful inside the administration and aggressive enough to worry opponents. It is also, read charitably, an accurate description of the scale of the coordination problem.

Not everything is flowing in the direction of the deals. The White House published a "Ratepayer Protection Pledge" in March 2026, signaling that the administration sees residential electricity bill pressure as a political exposure ahead of the November midterms. CNBC's reporting on Trump's "AI data center problem" frames the tension precisely: push nuclear and AI data centers fast, and residential bills rise in the PJM footprint; slow them down, and the administration's signature tech-infrastructure narrative wobbles. The state-level front is where the fight is most acute. Pennsylvania's PUC, Virginia's State Corporation Commission, and New York's PSC are all running proceedings on whether data center load should cross-subsidize other ratepayers, or be walled off from the retail rate base. Those decisions will shape 2027 timelines more than any NRC filing.

What Could Still Derail the 2027 Timeline

The 2027 date is realistic only if three things go right, and each has independent failure modes.

NRC approval is the first. The license amendment hearing window closes April 27, 2026. If intervenors, led by TMI Alert and joined by national organizations like Beyond Nuclear, file substantive challenges, the restart can slip by quarters rather than months. The commission has stated publicly that it sees the Crane review as precedent-setting, which cuts both ways: it wants the review to be thorough enough to withstand scrutiny, and it feels pressure to move quickly. Those two pressures produce a narrow path.

Supply chain is the second. Large power transformers are the bottleneck commodity for nuclear restarts and for the broader grid buildout. Lead times stretch into years, and the US transformer manufacturing base has not kept pace with demand. Constellation has queued orders for the Crane restart, but a single missed delivery on a main transformer can add quarters to the schedule. Control system replacements and primary-side components face similar lead-time risks. The Hill op-ed arguing the restart "will never happen" overstates the case, but its underlying supply-chain concerns are not invented.

Labor is the third. Nuclear-qualified operators, reactor engineers, and licensed technicians are scarce. The 500-person on-site count is ahead of the original baseline. Peak restart labor needs may exceed what Constellation can fully source from the existing workforce without competing against other reactor restart and new-build projects that are also staffing up. Every reactor restart in the Palisades-Crane-Duane Arnold pipeline is drawing on the same talent pool.

Political pressure cuts both ways. The Trump administration is pushing the build as a matter of strategic priority. But residential electricity bill increases attributed to data center demand are showing up in state-level polling, and if state PUCs or the FERC revisit cost allocation in late 2026, the economics of single-customer PPAs get harder. The cost overruns, not the safety incidents, will likely be the headline if the project slips. Nuclear restarts have historically run over budget more often than they have triggered incidents. The Crane project has the advantage of operating history, a well-known reactor design, and federal backing, but none of those eliminate cost risk.

Market risk is the final variable. All of these PPAs assume AI data center electricity demand keeps growing at something close to its current trajectory. If model efficiency improvements or a plateau in frontier-scale training reduces the demand curve, twenty-year PPAs look different. Every hyperscaler deal has escape clauses; most are not fully disclosed. The timeline math works if the demand math holds. Both are contingent.

What This Means for Anyone Not Building a Data Center

The tech companies get firm power. Everyone else inherits the infrastructure, the bill, and the precedent.

The quieter fight is over who pays for the transmission network upgrades that support hyperscaler load. FERC's December 2025 order partially insulates hyperscalers from direct subsidy of broader grid costs, but the transmission-upgrade socialization question is still live at the state level. Virginia's data center buildout has already produced formal consumer advocate filings at the SCC. Pennsylvania will face similar proceedings as Susquehanna and Crane come online. New York is working through a similar calculation for Meta's Clinton-area commitments as they scale.

Precedent matters more than any individual deal. If the Crane restart succeeds as a single-customer PPA with minimal rate-base exposure, expect Palisades, operated by Holtec in Michigan, and potentially Duane Arnold in Iowa and Clinton in Illinois, to follow the same template. The reactor-restart pipeline for US nuclear depends on this one working cleanly. A cost overrun or a delay that gets blamed on the single-customer structure will reshape the terms of every follow-on deal.

The structural implication is larger than any one reactor. Direct data-center-to-reactor PPAs represent a quiet shift from a shared-resource grid to a private-generation model. Firm power is increasingly sold to single customers on long-term contracts rather than dispatched to a common pool. That is efficient for the hyperscalers. It is less obviously good for reliability, cross-subsidy, and long-run planning in the rest of the grid. Regulators at both the federal and state level are working out what that shift means in real time, under political pressure on both sides.

For anyone not running a hyperscaler, the practical question is whether the firm nuclear capacity being contracted today eventually flows back to retail customers, or whether it stays behind the meter indefinitely. The twenty-year term of these PPAs suggests the latter. What happens after 2046 is a policy question the current negotiation doesn't resolve.

The Test Case

The Three Mile Island restart is the test case for whether Big Tech's nuclear thesis can graduate from press release to grid-connected generation. If Constellation hits 2027, the 10+ GW of hyperscaler nuclear commitments signed through early 2026 become a credible path rather than a collection of hopeful deals. If the restart slips substantially, every other PPA in the portfolio gets re-underwritten, and the SMR-vendor capitalization model that Meta, Google, and Amazon have been writing large checks against gets re-examined. The NRC intervention window closing April 27, the FERC co-location rules settling into PJM's tariff, and the state PUC rulings on cost allocation are the three near-term gates. For anyone trying to make sense of the overlapping deal flow, regulatory action, and political narrative, keeping a usable record of how each thread moves is the difference between confident decision-making and responding to headlines. Tooling that behaves like an AI-native second brain is genuinely useful here, because the thing being tracked is not a single deal, but the shape of a decade of infrastructure. Three Mile Island is the bellwether. The rest of the nuclear bet rides on it.