Microsoft, Google, Amazon, and Meta have each signed real nuclear deals for AI power. But the power flowing soonest — Three Mile Island's restart under Constellation, existing fleet PPAs — comes from old, large reactors, not new technology. Every genuine small modular reactor deal in this letter targets 2030 or later, and not one is under construction.
"Big Tech goes nuclear" has been one of the year's most repeated headlines, and it is not wrong — the deal volume is real, the dollar figures are enormous, and the strategic logic is sound: AI data centres need firm, round-the-clock power that solar and wind alone cannot reliably provide, and nuclear is the only zero-carbon source that can run at near-100% capacity factor. What the headline compresses away is a distinction that matters enormously to anyone actually trying to invest around this theme: almost none of the power arriving in the next two years comes from the new reactor technology the coverage is actually about.
Sort the announced hyperscaler nuclear deals by what they actually are, and a clean line appears. Microsoft's flagship deal with Constellation Energy — an 835 MW, $16 billion, 20-year power purchase agreement — is a restart of Three Mile Island's Unit 1, an existing large reactor that had already operated safely for decades before being shut down for economic reasons in 2019. FERC approved the necessary transmission changes in June 2026, and the plant has been renamed the Christopher M. Crane Clean Energy Center; it is targeting online status in the second half of 2027 and is genuinely under construction (in the sense of restart works) today. Meta's arrangements with Vistra and Constellation for up to 6.6 GW across their existing nuclear fleets are the same category of deal: buying power from reactors that already exist and already run.
Every other major deal in this space — the ones actually built around new small modular reactor (SMR) technology — sits in a different category entirely. Google's 500 MW agreement with Kairos Power, using KP-FHR reactor technology, targets 2030 and is described even by trackers following the deal closely as the "first-ever US corporate SMR fleet deal" — a first, not a proof of concept. Amazon's $700 million-led investment in X-energy and Energy Northwest, covering up to 960 MW across twelve Xe-100 units with a first phase of four units (320 MW), targets "the 2030s." Meta's TerraPower deal, up to 4.0 GW across eight Natrium plants, targets its first two units (690 MW) by 2032. Meta's Oklo deal, 1.2 GW at the Aurora SFR campus in Ohio, has no confirmed online date in current tracking. Not one of these SMR projects has a reactor under construction and producing power today.
| Deal | Technology | Capacity | Target Online |
|---|---|---|---|
| Microsoft / Constellation (Three Mile Island) | Existing large reactor | 835 MW | H2 2027 |
| Meta / Vistra & Constellation fleet PPAs | Existing large reactors | up to 6.6 GW | Already operating |
| Google / Kairos Power | New SMR (KP-FHR) | 500 MW | 2030 |
| Amazon / X-energy | New SMR (Xe-100) | up to 960 MW | 2030s |
| Meta / TerraPower | New SMR (Natrium) | up to 4.0 GW | 2032-2035 |
| Meta / Oklo | New SMR (Aurora SFR) | 1.2 GW | Not yet confirmed |
None of this means SMR technology is a mirage — Kairos, X-energy, and TerraPower are real companies with real engineering, real regulatory applications in progress, and, in Kairos's case, an operating test reactor. But "signed a deal with a hyperscaler" and "is generating power for a hyperscaler" are very different claims, and every SMR deal in this letter is currently the first kind, not the second. Investors reading "Amazon goes nuclear" as a near-term revenue event for the SMR developer involved are, in every case documented here, reading a 2030-or-later story as though it were a 2026 one.
The nuclear renaissance the headlines describe is really two renaissances at different clocks: one restarting plants that already worked, arriving within eighteen months, and one betting on plants that have never run at scale, arriving in five to ten years if the engineering, financing, and regulatory path all hold.
The near-term, lower-risk trade is the existing fleet operators actually collecting power-purchase revenue now or within the next two years — Constellation Energy and Vistra, both of which have publicly traded equity and disclosed hyperscaler contracts that show up in reported financials rather than in a press release alone. The SMR developers themselves are a different risk category entirely: NuScale Energy is the most established publicly traded pure-play with an actual NRC-approved design, though even NuScale has no operating commercial SMR yet; Oklo is public but pre-revenue; Kairos Power, X-energy, and TerraPower remain privately held as of this letter, meaning most retail investors cannot buy direct exposure to the three largest hyperscaler SMR partnerships even if they wanted to.
SMR technology carries real, unresolved engineering and regulatory risk — no company in this category has yet built and operated a commercial-scale unit in the US, and nuclear construction has a long, well-documented history of cost overruns and multi-year delays even for conventional reactor designs. The existing-fleet trade carries a different risk: regulatory approval, like the FERC transmission ruling Three Mile Island needed, is not guaranteed on every restart, and the economics of these long-dated PPAs depend on AI power demand actually materialising at the scale contracted. A slowdown in AI infrastructure spending would pressure both categories, though the existing-fleet operators have a diversified base business to fall back on that the pure-play SMR developers do not.
Treat this as two separate trades, not one nuclear-renaissance theme. Own Constellation and Vistra for the power-purchase revenue that is real and near-term. Treat any SMR pure-play — public or, more often, still private — as a venture-stage bet on 2030-plus technology, sized accordingly, and reread every future "hyperscaler signs nuclear deal" headline for which of these two categories it actually belongs to before assuming it changes anything in the next two years.