In January 2026, Meta signed deals with three nuclear energy companies — Vistra Corp., Oklo, and TerraPower — locking in more than 6 gigawatts of nuclear power. In a single month.
Google, Amazon, and Microsoft have all signed comparable commitments. The reason is the same for all of them: AI data centres need electricity that never stops, and the grid cannot deliver it fast enough.
This article covers who has signed what, why nuclear specifically, and what the OpenAI/Nvidia demand picture reveals about where this is heading. It is part of our comprehensive nuclear power’s AI renaissance series, which explores every major dimension of how the AI power crisis is reshaping energy strategy. The strategic implications — particularly what this means for your infrastructure strategy — are explored in the companion articles.
What is actually driving tech companies toward nuclear power?
AI training and inference run 24 hours a day, every day of the year. They cannot pause for clouds, calm weather, or peak demand windows. A single large AI training cluster can demand 500 megawatts of power continuously — comparable to a small power plant running flat out.
The requirement is baseload power: always-on electricity that does not depend on weather or time of day. Nuclear plants run at approximately 90–95% capacity — the proportion of time they operate at full output. No other clean energy source comes close.
Meta’s Mark Zuckerberg put it this way: “power will be the bottleneck that will limit AI growth, especially as it takes years for utility providers to build new power plants and expand the grid.”
Meta is building two data centres: the 1 GW Prometheus in Ohio and the 5 GW Hyperion in Louisiana. The grid cannot scale fast enough to supply them. US data centre electricity consumption is projected to more than double from 2024 to 2030. These deals are supply security decisions, not sustainability gestures.
Why can’t solar and wind meet AI data centre demand?
Solar and wind are intermittent. Solar generates when the sun shines. Wind generates when wind blows. Each achieves output maybe 20–35% of the time. AI workloads cannot be scheduled around weather.
The instinct is to solve this with batteries — Battery Energy Storage Systems (BESS). The problem is scale. At December 2024 prices of $115 per kWh plus installation, the batteries alone for a five-gigawatt facility would cost over $5 billion. Tim Fist, senior technology fellow at the Institute for Progress, is blunt: pairing solar and wind with batteries “isn’t feasible yet” for a 500 MW data centre.
Amazon, Microsoft, Meta, and Google are the four largest purchasers of corporate renewable energy PPAs, having contracted over 50 GW combined — equal to Sweden’s generation capacity. Still not enough for 24/7 AI baseload needs.
The Uptime Institute’s 2026 global data centre forecast names power availability — not chip supply or land — as the binding constraint on data centre expansion: “Developers will not outrun the power shortage — AI-driven load growth will intensify pressure on already constrained grids.”
Geothermal is one emerging firm-power alternative — Google has backed Fervo Energy as a near-term complement. The grid saturation that is forcing the nuclear turn is covered in depth in the companion article on grid infrastructure bottlenecks.
Who has signed what nuclear deals — and at what scale?
Meta leads on volume. Its January 2026 deals total more than 6 GW across four counterparties. Here is the full picture across the major hyperscalers:
Meta — 6+ GW (January 2026) Vistra Corp.: 20-year PPA, 2,176 MW from the Perry and Davis-Besse plants in Ohio, plus 433 MW in capacity upgrades. Existing plant — power flows fastest. Constellation Energy: 20-year PPA, 1.1 GW from the Clinton Clean Energy Center in Illinois, previously slated to retire in 2027. Oklo: 1.2 GW from Aurora Powerhouse reactors in Ohio, targeting 2030. New build — higher risk, longer timeline. TerraPower: first two Natrium reactors at 690 MW, targeting early 2032, with rights to six additional units totalling 2.8 GW.
Google Kairos Power: the world’s first corporate SMR (small modular reactor — factory-built units up to ~300 MW) purchase agreement, 500 MW across six or seven molten salt reactors, first unit by 2030. NextEra Energy: restart of Iowa’s Duane Arnold plant, 615 MW, 25-year PPA, back online by 2029 — the first nuclear plant restart in the US. Kairos Power/TVA: Hermes 2 plant in Oak Ridge, Tennessee, 50 MW from 2030.
Amazon (AWS) X-Energy Xe-100 SMR modules (320 MWe) in Washington State with Energy Northwest; broader goal of 5 GW by 2039.
Microsoft Constellation Energy: 20-year PPA signed in 2023 to restart Three Mile Island Unit 1 in Pennsylvania, 835 MW of carbon-free power by 2028.
Oracle Planning a 1 GW data centre powered by three SMRs.
The overall picture: tech giants have committed over $10 billion to nuclear partnerships, with 22 gigawatts in development globally. The specific reactor technologies behind these deals — Oklo’s Aurora Powerhouse, TerraPower’s Natrium, Kairos’s molten salt design, and X-Energy’s Xe-100 — are covered in the companion article. And for the record: Microsoft’s SEC filing named “community opposition, local moratoriums, and hyper-local dissent” as material operational risks alongside power availability. Power procurement is now a board-level disclosure item.
What is a power purchase agreement and how do these deals work?
A power purchase agreement (PPA) is a long-term commercial contract in which a buyer commits to purchasing electricity from a generator at a fixed price over a set term. For nuclear deals, that term is typically 15 to 25 years.
The structure works for both sides. The buyer — Meta, Google, Microsoft — gets price certainty and supply security. The generator — Vistra, Constellation, Oklo — gets the guaranteed revenue stream that justifies the capital investment. Tech companies do not own the reactors. They are long-term electricity customers.
The capital cost of a nuclear plant ($5–20 billion) makes direct ownership unattractive. A PPA transfers construction risk and operational responsibility to the energy specialist.
Two risk profiles matter here. The Meta/Vistra deal is an existing-plant PPA: Perry and Davis-Besse are already operating, so power flows relatively quickly. The Meta/Oklo deal (1.2 GW, targeting 2030) is a new-build SMR PPA: Oklo has yet to receive full Nuclear Regulatory Commission approval. Longer timeline, higher risk.
TerraPower targets $50–60 per megawatt-hour for scaled Natrium production; Oklo targets $80–130 per MWh. The bet is that manufacturing scale brings costs down. These 15–25 year terms are worth noting: they outlast most corporate strategy cycles by a significant margin.
What does the OpenAI/Nvidia compute plan add to the power demand picture?
OpenAI and Nvidia announced “the biggest AI infrastructure deployment in history”: OpenAI will deploy at least 10 gigawatts of Nvidia systems. The output of roughly ten large nuclear reactors, from a single partnership.
Sam Altman, OpenAI’s CEO: “Building this infrastructure is critical to everything we want to do. This is the fuel that we need to drive improvement, drive better models, drive revenue, drive everything.” He envisages facilities going “way beyond” 10 GW. SoftBank, OpenAI, Oracle, and MGX plan to spend $500 billion in four years on US data centres; the first broke ground in Abilene, Texas.
Unlike Meta, Google, Amazon, and Microsoft, OpenAI and Nvidia have not yet secured comparable long-term nuclear power commitments. The compute ambition is stated. The electricity to run it has no committed generation source yet. The share of US electricity going to data centres may triple from 4.4% in 2024 to 12% by 2028.
What does the nuclear rush reveal about the future of AI infrastructure?
Power procurement has moved from facilities management into board-level strategy. Microsoft’s SEC filing — naming “community opposition, local moratoriums, and hyper-local dissent” alongside power availability as material operational risks — is the clearest signal of how far this has shifted. When power supply appears in a 10-K, it is no longer an infrastructure team problem.
The deals reveal a timeline problem. Data centres can be built in two to three years. But new SMRs targeting 2030–2032 are still years away. There is currently a seven-year wait on some US grid connection requests. The gap will be filled with gas turbines and on-site generation.
Power is becoming a competitive moat. Microsoft’s Three Mile Island PPA, signed in 2023, established the commercial model that others followed. First movers locked in supply and price certainty late entrants cannot easily replicate. AI data centre demand contributed to an 833% increase in PJM’s capacity market auction price for 2025–2026. Seven of 13 major US grid regions are projected to operate below their safety margins by 2030.
For the full context on the nuclear buildout reshaping AI infrastructure, including every major dimension from grid policy to reactor technology to energy alternatives, see the series overview. For what this means for your infrastructure strategy — how to assess, communicate, and act on AI power risk without being a hyperscaler — see the companion framework article.
FAQ
Why are Google and Amazon building nuclear reactors for their data centres?
Neither is building reactors — they are signing long-term PPAs with reactor builders and operators. The driver is baseload power: AI workloads run continuously and require uninterrupted electricity that solar and wind cannot reliably provide. Google has partnered with Kairos Power (seven SMRs, 500 MW, first by 2030) and NextEra Energy to restart Iowa’s Duane Arnold plant (615 MW, back online 2029). Amazon has committed to four X-Energy Xe-100 modules (320 MWe) in Washington State.
What does nuclear power have to do with ChatGPT and AI?
Running ChatGPT requires massive, continuous electricity — every query draws power around the clock. Training is more intensive still: GPT-3 consumed 1,287 megawatt-hours during training, enough to power about 120 average US homes for a year. Today’s models are far larger. Nuclear is the only clean source that can guarantee 24/7 supply at this scale.
How much power does an AI data centre actually require?
A large AI training cluster can demand 500 megawatts of continuous power. Meta’s Hyperion data centre in Louisiana is planned at 5 GW — equivalent to the continuous demand of roughly 5 million average homes.
Is there really a problem getting power for AI data centres?
Yes. The Uptime Institute’s 2026 forecast names power availability as the binding constraint on data centre expansion, ahead of chips, land, and connectivity. In the US, some grid connection requests are waiting seven years. Between April and June 2025, $98 billion worth of data centre projects were blocked or delayed across 11 states — power and zoning working as two separate but intersecting constraints.
Why is baseload power important for AI and what is it?
Baseload power is electricity generated continuously at constant output, regardless of weather or time of day. AI training and inference cannot be paused or throttled without affecting quality and latency. Solar and wind are weather-dependent, and battery storage at the required scale does not yet exist.
Why is Microsoft’s Three Mile Island deal significant?
Microsoft signed a 20-year PPA with Constellation Energy in 2023 to restart Three Mile Island Unit 1 in Pennsylvania — 835 MW of carbon-free power by 2028. It demonstrated that a major tech company would make a 20-year commitment to nuclear, and the deal structure it used became the model that Meta, Google, and Amazon subsequently adopted.
What is a small modular reactor and why does it matter for these deals?
An SMR (small modular reactor) is a next-generation nuclear design up to ~300 MW per unit — roughly one-third of a traditional nuclear plant — designed for factory prefabrication and modular scaling. Most hyperscaler new-build deals are with SMR developers: Oklo, TerraPower, Kairos Power, X-Energy. The specific reactor technologies behind these deals — including cost targets and regulatory timelines — are covered in the companion article.
How long will it take for these nuclear deals to deliver power?
Existing plant PPAs (Vistra, Constellation, NextEra restarts) can deliver power within one to three years. New SMR construction is slower: Oklo is targeting 2030, TerraPower early 2032. Demand exists now; new nuclear supply is five to ten years away. The gap will be bridged with gas turbines and on-site generation.
What is Vistra Corp. and why is it central to the Meta deal?
Vistra Corp. is one of the largest US electricity generators. Meta’s largest single nuclear commitment is a 20-year PPA with Vistra for 2,176 MW from the already-operating Perry and Davis-Besse plants in Ohio — the fastest path to nuclear baseload for Meta’s Prometheus data centre.
Why are tech companies funding nuclear instead of building their own power plants?
The capital cost of a nuclear plant ($5–20 billion) makes direct ownership unattractive. A PPA transfers construction risk and operational responsibility to the energy specialist; the tech company gets contractual rights to a specific volume of power at a known price without becoming an energy company.
What is the PJM Interconnection and why does it matter for AI data centres?
PJM Interconnection manages the electricity grid across 13 US Mid-Atlantic and Midwest states — the largest electricity market in the world by load, serving 65 million people. Virginia (“Data Center Alley”) sits within PJM territory. AI data centre demand contributed to an 833% increase in PJM’s capacity market auction price for 2025–2026, with load growth forecast to increase 48% from 2025–2045.
How does power availability affect companies that don’t own data centres?
If your organisation runs on cloud infrastructure rather than owned data centres, power constraints still reach you — indirectly, through pricing and regional availability. If hyperscalers face higher long-term power costs from nuclear PPAs, those costs will be reflected in cloud compute pricing over time. Power constraints also determine which cloud regions can expand — a provider that cannot secure grid connection cannot add capacity there.
