Amazon Web Services and Talen Energy have expanded their nuclear power partnership, locking in electricity from the Susquehanna nuclear plant in Pennsylvania to feed data centers running artificial intelligence workloads. The deal, disclosed in an SEC filing on June 11, 2025, adds to a growing roster of tech-to-nuclear contracts that now collectively approach 10 gigawatts in announced capacity. With Microsoft, Google, and Amazon all racing to secure carbon-free baseload power, the question is no longer whether Big Tech will bet on nuclear but whether the grid and regulators can keep pace.
Why behind-the-fence nuclear deals are reshaping power markets
The core tension is structural. AI data centers demand continuous, high-density electricity that wind and solar farms cannot reliably deliver around the clock. Nuclear plants, which typically run at capacity factors above 90 percent, offer exactly that kind of output. But the grid connecting generators to customers is increasingly congested, especially in the PJM Interconnection territory that covers much of the mid-Atlantic and Midwest. Tech companies waiting years in interconnection queues have found a shortcut: buying power directly from a nuclear plant through a behind-the-fence arrangement that bypasses the public transmission system entirely.
Talen Energy’s Cumulus Data Campus illustrates how this works in practice. The campus sits adjacent to the Susquehanna plant in Luzerne County, Pennsylvania, and draws electricity through a direct-connect configuration rather than pulling from the regional grid. That physical proximity eliminates transmission bottlenecks and lets the data center scale without waiting for new high-voltage lines. The arrangement also means Talen can monetize its nuclear output at premium rates while AWS gets firm, zero-carbon power for its AI clusters.
If this model proves durable, it will pull siting decisions toward existing nuclear plant footprints rather than greenfield locations. County zoning filings near operating reactors and NRC license amendment applications over the next 18 months will serve as early indicators of whether other utilities and tech firms follow the same playbook. Communities near plants like Calvert Cliffs in Maryland, where Constellation operates, or the South Texas Project could see new data center proposals surface as developers seek the same behind-the-meter advantages.
SEC filings and federal data confirm the scale of commitments
The numbers in public filings and federal energy data anchor the scale of these deals. Talen’s latest SEC disclosure confirms the expanded relationship with Amazon through a new power purchase agreement supplying Susquehanna nuclear output to AWS data centers supporting AI and other cloud technologies. An earlier Talen registration statement filed with the SEC references both 540 MW and 960 MW as contractual and zoning conditions tied to the Cumulus Data Campus buildout, establishing a phased capacity ramp that can grow as AWS adds servers and networking equipment.
Federal data from the U.S. Energy Information Administration independently confirms two of the largest individual arrangements. The EIA identifies the AWS–Talen Susquehanna contract at 960 MW, alongside Microsoft’s 835 MW agreement with Constellation. The Microsoft–Constellation arrangement centers on restarting the Three Mile Island Unit 1 reactor, a project that would bring a shuttered plant back online specifically to serve data center load. Together, those two deals alone account for nearly 1,800 MW of committed nuclear capacity directed at tech-sector buyers, equivalent to the consumption of several million typical U.S. homes.
The 10-gigawatt figure referenced in broader industry tracking includes additional announced agreements from Google, Meta, and other hyperscale operators pursuing small modular reactors and conventional plant restarts. No single primary source aggregates all of these contracts into one verified total. The individual deal sizes documented in SEC filings and EIA records, however, show a clear pattern: tech companies are signing multi-hundred-megawatt nuclear offtake agreements at a pace that has no precedent in the U.S. power sector.
Context from other federal datasets underscores why these volumes matter. Weekly information from the EIA’s natural gas storage reports highlights how much of the current power system still depends on gas-fired generation to balance variable renewables. Every gigawatt of nuclear capacity contracted to run AI data centers instead of serving the broader grid tightens that balance, forcing grid planners to reconsider assumptions about future gas demand, reserve margins, and emissions trajectories.
Open questions on permitting, grid impact, and the path to 10 GW
Several material gaps remain between announced deals and operating capacity. The most immediate is permitting. Restarting a shuttered reactor like Three Mile Island Unit 1 requires Nuclear Regulatory Commission approval, environmental review, and potentially years of refurbishment and safety upgrades. New small modular reactor designs have not yet received full commercial licenses, and developers still need to clear hurdles on construction financing, supply chains, and long-term waste management. The timeline from signed power purchase agreement to electrons flowing could stretch well beyond 2030 for many of these projects.
Grid operators face their own dilemma. Behind-the-fence deals remove large blocks of generation from the public grid, reducing the supply available to residential and industrial customers. PJM Interconnection has already flagged concerns about reliability if too much nuclear output is diverted to private buyers, especially during heat waves and cold snaps when demand spikes. Regulators in states like Pennsylvania and Maryland will need to weigh the economic benefits of data center investment against the risk of tighter electricity markets and higher prices for everyone else.
The financial structure of these agreements also raises questions. Nuclear plants carry high fixed costs, and long-term PPAs with creditworthy tech buyers provide revenue certainty that merchant power markets do not. That stability can extend the life of existing reactors and justify new capital spending, but it can also concentrate risk. If AI demand projections soften or regulatory rules change, utilities could find themselves locked into contracts that limit flexibility to serve other customers or respond to market signals. Conversely, if wholesale prices spike, policymakers may face pressure to claw back some of the economic value now flowing to private, off-grid loads.
Local impacts add another layer of complexity. Data centers bring construction jobs, tax base growth, and follow-on investment in fiber and logistics. They also consume large amounts of water for cooling, strain local distribution networks, and can provoke community opposition over noise, land use, and perceived inequities if residents see little direct benefit from facilities powered by nearby nuclear plants. Counties hosting reactors that become magnets for AI clusters will have to navigate zoning, infrastructure upgrades, and revenue-sharing debates that go well beyond traditional utility planning.
Finally, there is the question of what happens after the first wave of marquee deals. If AWS, Microsoft, and their peers collectively reach the 10 GW mark in nuclear-linked contracts, the marginal project may face steeper political and technical challenges. Transmission planners could push for more on-grid solutions that share new nuclear output broadly rather than dedicating it to single corporate buyers. Environmental advocates divided over nuclear’s role in decarbonization may intensify scrutiny of reactors revived primarily to power AI. And as more states chase data center investment, competition over limited nuclear capacity could drive up contract prices, eroding some of the economic logic behind today’s arrangements.
The AWS–Talen expansion at Susquehanna shows how quickly the contours of the power system can shift when digital infrastructure and nuclear assets align. Whether this model becomes a durable pillar of the AI economy or a transitional strategy constrained by regulation, public opinion, and grid physics will depend on decisions made over the next few years by utility commissions, federal regulators, and local communities living in the shadow of these plants.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
