Nuclear’s Second Life Is Being Funded by Tech, Not Utilities

in short

In the past eighteen months, hyperscalers have signed roughly 9.8 gigawatts of nuclear power purchase agreements across thirteen disclosed deals — more private nuclear procurement than the industry saw in the previous twenty years combined. The companies writing the checks aren't utilities. They're the same four or five names running the world's AI infrastructure.

Reactor restarts used to be a utility’s decision, made on a regulator’s timeline, funded by ratepayers over decades. That model has quietly been replaced. Microsoft signed a 20-year, roughly $16 billion agreement with Constellation Energy to restart a reactor at the former Three Mile Island site — now renamed the Crane Clean Energy Center — bringing 835 megawatts online specifically for its own use by 2027 (Presenc AI). Amazon committed roughly $700 million to X-energy and separately signed for 1.9 gigawatts with Talen Energy (Presenc AI). Across the sector, this is now the largest wave of private nuclear procurement in half a century, and it is unmistakably tech capital, not utility capital, doing the underwriting.

The Deal Map

The roughly 9.8 gigawatts spans two distinct strategies: restarting or extending existing large reactors, and financing an entirely new category of small modular reactors (SMRs) that don’t exist commercially yet. Microsoft was first to actually bring power online through the Crane restart; Meta signed the largest and longest-dated commitment in the group, agreeing to a 20-year deal and separately arranging for more than 2 gigawatts through Entergy (SMR Intel). Meta’s broader nuclear exposure runs to roughly 6.6 gigawatts, and Amazon’s X-energy commitment carries a $50 billion, 960-megawatt structure (EnkiAI). Google has its own restart PPA covering 615 megawatts. Oracle has committed to three separate SMRs (Internet Pros).

The SMR side of this is the more speculative bet. No commercial SMR fleet is operating at scale anywhere in the world yet; these are procurement commitments against reactor designs still working through licensing and first-of-a-kind construction risk, with delivery timelines stretching toward 2030 (IDTechEx). More gigawatt-capacity has been signed in the past eighteen months than in the prior twenty years combined, and the enriched uranium fuel these designs need — HALEU — is itself in short supply, a bottleneck that could throttle the whole SMR wave regardless of how much capital is behind it (Core Insights Review).

Why Tech, Not Utilities, Is Underwriting This

The honest explanation is that utilities have already been burned once, badly, and haven’t forgotten it. Georgia’s Vogtle plant — the last large reactor built from scratch in the United States — ran roughly seven years late and cost about $36.8 billion, becoming the most expensive power plant in American history and driving a documented 23.7% rate increase, with pre-construction charges alone adding close to $1,000 per household before the plant ever generated a watt (NuclearCosts.org; The Current). Westinghouse, the reactor’s builder, went bankrupt in the middle of construction. Regulators and ratepayer advocates absorbed that lesson, and utilities have been reluctant to repeat it since.

Hyperscalers face none of the same political exposure, and they have balance sheets utilities don’t. A tech company can sign a twenty-year PPA the way it signs a cloud contract — as a bet on future demand it already believes in — without asking a public utility commission’s permission first. That’s the entire shift in one sentence: the customer with the deepest pockets and the least regulatory friction became the financier.

The Challenge: Ratepayers and Regulators Are Pushing Back

This is not friction-free, and treating it as such would understate a real and ongoing fight. In November 2024, FERC rejected an interconnection agreement between Talen and Amazon that would have let Amazon’s data center draw power directly from the Susquehanna nuclear plant, outside the normal grid queue — a structure regulators worried would shift roughly $140 million a year in costs onto ordinary ratepayers (Utility Dive; Canary Media). FERC denied a revised version of the same arrangement again in April 2025 (ANS Nuclear Newswire). The market noticed both times — Constellation, Vistra, Talen, and Amazon shares all moved on the news (TipRanks).

The regulatory fight is really about who pays for grid upgrades. Industry estimates put the interconnection cost for linking a large data center directly to a nuclear plant at $150–250 million, and the unresolved question is whether that cost gets absorbed by the hyperscaler alone or socialized across the wider ratepayer base (Canary Media). Seven hyperscalers have since signed onto a broader Ratepayer Protection Pledge, committing publicly that their power demand won’t raise household electricity bills — a direct policy response to exactly this fight (ANS Nuclear Newswire). Whether that pledge holds once the SMR wave actually needs to connect to the grid at scale is, honestly, untested.

Conclusion

Watch who signs the nuclear PPA, not who owns the reactor. Utilities used to control the pace of nuclear buildout because they controlled the financing; that control has now passed to a handful of technology companies with both the capital and the demand to make these commitments on their own terms. Total data center electricity demand is projected to approach 1,000 terawatt-hours by 2026, and federal policy — including the ADVANCE Act’s SMR licensing reforms — is actively being shaped around this demand (Informed, Clearly). The 9.8 gigawatts already signed is not the ceiling. It’s the opening bid in a much larger repricing of who controls power infrastructure in an AI economy — and Vogtle’s lesson, that nuclear construction can go catastrophically over budget and over schedule, hasn’t disappeared just because the customer changed.

source

covers

Core deal tracker; Microsoft/Crane and Amazon/Talen detail

Deal-by-deal breakdown; restarts vs. SMRs

Skeptical assessment vs. US energy realities

IDTechEx — Data Centers Go Nuclear

SMR market forecast and investment thesis

Meta, Amazon, Google commitment figures

Timeline and participant map

Execution risk; HALEU fuel bottleneck

Amazon, Google, Oracle deal specifics

Co-location regulatory risk

Why utilities stopped building nuclear

Vogtle cost and timeline detail

Ratepayer impact

Interconnection cost quantification

Second FERC rejection; Ratepayer Pledge

Vogtle ratepayer cost figures

Market reaction to FERC rejection

Demand projection; ADVANCE Act

Disclaimer!

The content in this article is provided for general informational and educational purposes only. It does not constitute financial, investment, legal, tax, or professional advice of any kind, and should not be relied upon as a basis for any investment or financial decision.This piece reflects independent research and analysis. It is not a recommendation to buy, sell, hold, or otherwise transact in any security, token, asset, or investment product, and is not an offer or solicitation of any kind in any jurisdiction. Nothing in this article is affiliated with, sponsored by, or intended to promote any specific company, fund, token, or investment offering.Figures, forecasts, and third-party data cited here are drawn from publicly available sources believed to be reliable as of the article’s publication date. They have not been independently audited, may be revised or superseded by later reporting, and should not be treated as guaranteed or complete. Forward-looking statements involve inherent uncertainty and actual outcomes may differ materially.Readers should conduct their own due diligence and consult qualified, licensed professionals before making any financial, investment, or legal decisions.

Share

The AI buildout resembles an arms race driven by the belief that falling behind is permanent. Learn why nobody racing has answered whether the compounding is real.
Chip fabs use millions of gallons daily yet water costs almost nothing. Learn why water pricing, recycling, and desert politics shape where AI chips get built.
The US semiconductor industry faces a 157,000 worker shortfall by 2030. Learn why the talent gap is the one AI bottleneck that capital alone cannot fix.