Installment 3 ended at the buildings; this one follows the electrons backward to the deals that buy them. The hyperscalers are acquiring reactors, sites, and grid access the way earlier builders acquired land — and the signings are not all they look like.
The cleanest number from the Microsoft–Constellation deal: if regulators approve the restart of Three Mile Island's Unit 1, the plant would provide Microsoft with 835 megawatts of energy.[1] The agreement, signed in September 2024, ran 20 years, rerouted the plant's output to data centers, and renamed the unit the Crane Clean Energy Center.[1] It is the first big-ticket proof that AI load is priced against legacy assets, not new ones: restarting a 1970s Pennsylvania reactor was cheaper than building anything else.
Constellation's original target was 2028; by June 2025 the reboot was fast-tracked to 2027.[2] The company ordered its main transformer and fuel, restored the water systems the plant needs, and completed infrastructure inspections required for permitting approvals.[2] In mid-2026 the 262-ton transformer arrived for installation — the physical layer of the physical layer: an old reactor, a new transformer, and a hyperscaler's checkbook.[2] The restart is a bet that a plant with 45-year-old turbines can be cheaper and sooner than any small modular reactor in the queue. (my read, not a sourced claim)
Two days of October 2024 changed the shape of the market. Google signed what The Hindu calls "the world's first corporate agreement to buy power from multiple small modular reactors" — 500 megawatts total from six to seven Kairos Power reactors, the first unit online 2030, the rest deployed through 2035.[3] By November 2025 the site had a name: Oak Ridge, Tennessee, under a long-term power purchase agreement with the Tennessee Valley Authority, delivering to Google data centers in Tennessee and Alabama from 2030.[8]
The coverage itself is part of the story. ETTelecom's report on the Oak Ridge deal described "the 50-gigawatt small modular nuclear power plant" — a figure three orders of magnitude off the 500 megawatts Google actually contracted.[8] It is a small sample of how the land-rush narrative inflates in transit: the correct number sits in the same article's framing, the wrong one in its lede.[3]
This is where the Bulletin of the Atomic Scientists' July 2026 analysis bites: "these press releases are all show, with little substance."[4] Amazon's Energy Northwest agreement funded only "the initial feasibility phase of an SMR project" in exchange for the right to purchase electricity from a first project of four modules expected to generate 320 megawatts — a right, not an obligation.[4] The Kairos–Google deal similarly sells energy under power purchase agreements without specifying the terms of the purchase.[4]
The check amounts tell the scale: the October 2024 consortium behind X-energy — including Amazon's Climate Pledge Fund, Citadel founder Ken Griffin, Ares Management affiliates, NGP, and the University of Michigan — invested "approximately $500 million."[4] TerraPower raised $650 million in June 2025 including NVIDIA's venture arm.[4] Against that, one estimate puts TerraPower's proposed 345 MW Natrium plant at $9.4 billion; the cancelled 462 MW NuScale project in Idaho was $9.3 billion when it died in 2023.[4] The Bulletin's arithmetic: even the largest announced investment is "not anywhere near the amounts needed to bring new nuclear power online."[4]
The timeline record makes the gap concrete. Russia's KLT-40S floating reactor — the only operating SMR precedent — took 13 years from construction start to electricity, against a promised three.[4] China's twin Shidao Bay high-temperature units took more than twice the promised 50 months.[4] And the US goalposts moved: in 2020 the Energy Department wanted the Natrium and Xe-100 reactors operational "within 5-7 years"; by November 2025 a Natrium executive promised "power on the grid" by 2031. As of July 2026, there were no small modular reactors under construction in the United States.[4]
Counterpoint, one month after the Bulletin's "no SMRs under construction": in August 2026 the Energy Department announced Oklo's Groves Isotope Test Reactor in Texas had reached "criticality" — a self-sustaining chain reaction — the fifth reactor under a DOE pilot program, after Aalo Atomics, Deployable Energy, Antares Nuclear, and Valar Atomics.[5] The same report is careful about what a test reactor is: "small modular reactors still face a number of hurdles and there are none operating commercially in the US."[5] Policy is pushing: a US executive order wants ten large conventional reactors under construction by 2030, toward a goal of 400 gigawatts of nuclear — four times today's fleet.[5]
One SMR is by definition under 300 MW — less than a third of a gigawatt.[4] The demand side is another magnitude. Reports cited by the Bulletin found 10% of data centers already require more than 1 gigawatt, heading to 20% by 2030; Stargate's Abilene project wants up to 5 gigawatts; Meta is planning a 5-gigawatt Louisiana site.[4] In January 2026 Pacifico Energy received a Texas air permit for 7.65 gigawatts of gas-fired generation — the largest permit granted in the United States.[4] The reactors being signed for today would not fill one of these campuses.
The February 2026 earnings calls put the asymmetry in one sentence: Amazon and Google said they expected to spend up to $200 billion and $185 billion respectively on data centers in the coming year.[4] Nuclear commitments measured in the hundreds of millions are, the Bulletin computes, "a fraction of one percent" of one year's data-center capex — which the authors read as greenwashing: announcement-budget money, not operating-budget money.[4] The frame cuts both ways: it also means the land rush has barely started — the buying power is there and the supply chain is not.
The raw-material constraint arrived in August 2026: X-energy and Centrus signed a definitive, long-term agreement for high-assay low-enriched uranium (HALEU) supply and TRISO fuel fabrication for the Xe-100 advanced reactor.[7] The fuel — not the reactor — is increasingly the binding physical layer for the whole SMR category.
The restart arc is the real 2020s physics: 835 MW from a 1970s plant beats any SMR on schedule, and hyperscalers priced it accordingly.
Purchase rights are not purchases. The signed deals are options on feasibility, and the Bulletin is right that the press-release layer obscures that.
The numbers that drift in transit (50-gigawatt vs 500 MW) are a warning for anyone reading the land rush from headlines: go to the agreement.
First criticality is not first commercial power; the same week's reporting said both things plainly.
Nuclear's world share is not rising with the hype: 8.8% of world electricity in 2025, roughly half its share 30 years earlier, while solar and wind reached 19.5%.
[1] Hindustan Times (Reuters): Constellation to restart Three Mile Island nuclear plant in Microsoft AI power deal (20 Sep 2024)
[2] The Hindu: Three Mile Island nuclear plant reboot fast-tracked to 2027 (26 Jun 2025)
[3] The Hindu: Google to buy power for AI needs from SMR company Kairos (15 Oct 2024)
[4] Bulletin of the Atomic Scientists: Data centers powered by next-gen nuclear? Don't fall for Big Tech's PR hype (20 Jul 2026)
[5] Bloomberg via Yahoo: US Says Oklo's Small Nuclear Reactor Reaches Criticality (6 Aug 2026)
[7] GlobeNewswire via Business Insider: X-energy, Centrus sign HALEU supply agreement for Xe-100 (6 Aug 2026)
[8] ETTelecom: Google chooses Tennessee for groundbreaking nuclear reactor to power data centers (Nov 2025)