Nearly 10 GW of nuclear is now under contract to power AI

The contracts, by the numbers

Track the announced deals and one figure stands out: ~9,765 MW of advanced nuclear contracted to data-centre buyers, against only 15 of 29 NRC dockets approved and one project actually under construction. The named counterparties:

• Meta — 5,200 MW. A trio of 20-year PPAs signed in January 2026 with Vistra, TerraPower and Oklo. The Vistra agreement alone covers over 2.1 GW across three existing plants (Perry and Davis-Besse in Ohio, Beaver Valley in Pennsylvania). TerraPower adds two Natrium reactors up to 690 MW with the first online as early as 2032; Oklo's Pike County, Ohio campus could add up to 1.2 GW from 2030.
• Unnamed operators — 2,200 MW. Roughly 23% of the book is booked across entries with no named hyperscaler. Charted separately precisely because neither side carries the combined figure.
• Amazon — 960 MW (including a nuclear-adjacent Pennsylvania campus and $500M into X-energy's SMRs).
• Microsoft — 835 MW. The 20-year PPA to restart Three Mile Island Unit 1, ~$1.6B, targeted online 2027 — the only deal already under construction.
• Google — 500 MW from Kairos Power, targeting 2030.

The international fork: Google went to Finland

The most instructive single deal is not American. In September 2026 Google committed at least €13 billion (~$15.1B) to AI infrastructure in Finland over two years — its largest European investment — including a 22-year agreement for up to 50% of the output of one of Finland's two nuclear plants (Fortum's Loviisa), its first nuclear deal outside the US. Google's own framing was blunt: "bring your own power."

Three things make the Finland deal different from the US ones. It is a purchase agreement on existing output, not a bet on a reactor that does not exist yet — the power is real today. It extends an operating plant's life to 2050 rather than funding new-build. And it is located where the cooling is free and the grid is stable. That is the template a company can actually execute on right now; the US new-build deals are options on the 2030s.

Why this is the compute story, not an energy story

Data-centre power demand reached 29.6 GW by late 2025 — roughly New York State's peak — and is projected to rise about 130% by 2030. An AI facility uses three to five times the energy of a conventional one. You cannot resolve that by deploying more GPUs; the constraint is upstream of the silicon.

That inverts the usual narrative. Through 2024 the bottleneck was chips. Through 2026 the bottleneck is interconnection, grid capacity and firm power — things that take five to ten years to build and cannot be expedited by capital alone. The compute landlord thesis has moved beyond real estate and hardware into energy procurement, because that is where the schedule risk now lives.

The ratepayer problem nobody wants to price

Announced contracts are not the same as delivered capacity, and the gap has a political cost. Oracle's October 2026 subscription to 125–250 MW from Wisconsin's Point Beach plant is the cleanest example of both the model and its friction. Point Beach's cost per MWh has risen from $45.94 in 2016 to $75.51 in 2026, with projections to $122.45 by 2033. The Oracle subscription is a primary driver of a proposed $176 million electric rate hike for other We Energies customers in 2027 — roughly 20% of the increase — even as the utility projects the deal will save customers about $300M in fuel costs between 2027 and 2033.

That tension is now sitting in front of public utility commissions. Wisconsin's regulator is weighing whether tech companies should cover 100% of new power-plant costs. Every deal signed in the next 24 months will be judged against that question, and the answer will set the price of "bring your own power" for everyone.

What it means for decentralised capacity

A near-10 GW nuclear book that mostly does not exist is a strong signal about where the market is heading and a weak signal about what is available now. The two things a decentralised network should take from it:

• Firm power is the moat, and it is scarce. Any node that can point to an existing grid interconnection with headroom has a real, defensible advantage over one that cannot — regardless of which GPU is installed.
• The 2026–2028 window is elastic-workload territory. While hyperscalers wait for reactors, the demand that exists today has to run somewhere. That is exactly the burst, deferrable, cost-sensitive load a distributed network is built to absorb.

The bottom line

10 GW of nuclear contracted, essentially zero new US nuclear output since 2020, and only one restart under construction. The hyperscalers are not buying power; they are buying position in a queue that clears in the 2030s. For everyone else, the computable bet is the opposite: serve the elastic demand that needs to run in the meantime, and treat a firm grid connection as the scarcest asset on the board.