Big Tech Signed $30 Billion in Nuclear Deals for AI. The Reactors Don't Exist Yet.
Microsoft, Google, Amazon, and Meta have collectively committed over $30 billion to nuclear energy for data centers. There are two operational SMRs on Earth. NuScale's flagship project was cancelled after costs doubled. The earliest new reactor comes online in 2030. The data centers need power now.
Data centers consumed 6% of all U.S. electricity in 2025. The Department of Energy projects that number will hit 12% by 2028 — roughly 260 terawatt-hours, more than the entire residential electricity consumption of Texas. An estimated $580 billion was spent on AI data center infrastructure in 2025 alone.
To power this demand without incinerating their carbon commitments, the four largest hyperscalers — Microsoft, Google, Amazon, and Meta — have signed a blitz of nuclear energy deals since mid-2024. Restart old plants. Fund new startups. Build small modular reactors on-site. The combined commitments exceed $30 billion.
Here is what those commitments have produced so far: zero watts.
The Deal Sheet
The nuclear shopping spree has been breathless. Microsoft signed a 20-year power purchase agreement with Constellation Energy to restart Three Mile Island Unit 1 — yes, that Three Mile Island — at a cost of $1.6 billion, with an additional $1 billion DOE loan. The reactor, renamed the Crane Clean Energy Center in an act of aggressive rebranding, is now targeting a 2027-2028 restart.
Google partnered with NextEra Energy to reopen the Duane Arnold plant in Iowa, closed since 2020, targeting a 2028 restart. Google also purchased 500 MW from Kairos Power, an SMR startup whose first reactor isn't expected to come online until 2030.
Amazon secured power from the Susquehanna nuclear plant through 2042 and invested in X-Energy, another SMR developer. Meta signed a 20-year agreement with a Constellation nuclear plant in Illinois, then announced deals with three additional nuclear companies for over 6 gigawatts of power. OpenAI CEO Sam Altman's Oklo secured a deal to supply 12 MW to Switch data centers. Bill Gates' TerraPower partnered with Sabey Data Centers.
The political wind is at their backs. President Trump signed executive orders to boost U.S. nuclear capacity from 100 GW to 400 GW by 2050. The DOE is distributing loans like party favors.
On paper, nuclear AI is a done deal. On the grid, it's a press release.
The Reality: Two Operational SMRs on Planet Earth
For all the excitement about small modular reactors — the technology most frequently cited as the silver bullet for AI power demand — the global inventory is bleak. There are exactly two operational SMRs in the world as of early 2026: Russia's floating reactor Akademik Lomonosov (a 70 MW barge reactor serving a remote Arctic town) and China's HTR-PM (a 210 MW high-temperature gas-cooled reactor in Shandong). Neither is in the United States.
The closest the U.S. came was NuScale's VOYGR, the only SMR design to receive full NRC certification. In November 2023, NuScale's flagship Carbon Free Power Project at Idaho National Laboratory was cancelled after projected costs ballooned from $5.3 billion to $9.3 billion — a 75% overrun that vaporized the entire economic argument for small modular over conventional nuclear.
The fundamental problem is the one nuclear has never solved: economies of scale run backwards when you shrink the reactor. Making a reactor smaller means higher costs per megawatt, not lower. The theory was always that factory manufacturing at volume would overcome this penalty. That serial manufacturing has never materialized. Not once, in the entire history of commercial nuclear power.
The Timeline Problem
Consider the actual calendar these deals operate on:
| Project | Type | Earliest Power-On | Capacity |
|---|---|---|---|
| Palisades (Holtec) | Restart | Early 2026 → delayed | 800 MW |
| Three Mile Island (Constellation/Microsoft) | Restart | 2027-2028 | 837 MW |
| Duane Arnold (NextEra/Google) | Restart | 2028+ | 601 MW |
| Kairos Power (Google) | New SMR | 2030+ | 500 MW |
| X-Energy (Amazon) | New SMR | 2030+ | TBD |
| TerraPower (Sabey/Gates) | New SMR | 2030+ | 345 MW |
| Oklo (Switch/Altman) | New SMR | 2030+ | 12 MW |
The restarts are the only projects with a prayer of delivering power before 2030. And even those are slipping. Palisades, which was supposed to make history as the first-ever restart of a decommissioning nuclear plant in the U.S., has already been pushed from late 2025 to early 2026 while Holtec works through NRC approvals, despite a $1.52 billion DOE loan guarantee and 600 full-time workers on payroll.
Three Mile Island's restart involves refurbishing a reactor that has been cold for seven years. The $1.6 billion price tag is a preliminary estimate. If nuclear history teaches anything, it's that preliminary estimates are fiction. The last U.S. reactor to come online — Vogtle Units 3 and 4 in Georgia — took over 14 years to build and cost more than double its original budget.
The new-build SMRs — Kairos, X-Energy, TerraPower, Oklo — are even further out. Most are still in the demonstration or licensing phase. Kairos plans a non-power demonstration reactor (Hermes) at Oak Ridge before building anything commercial. TerraPower broke ground in Wyoming in 2024 but won't produce electricity until the end of the decade. These are not construction delays. These are normal timelines for first-of-a-kind nuclear technology.
Meanwhile, the Grid Burns
The mismatch is vertiginous. U.S. data center electricity demand is growing at roughly 15-20% per year. The average hyperscale data center consumes as much electricity as 100,000 homes. Some of the largest facilities under construction will use 20 times that, according to the IEA.
This demand isn't waiting for 2030. It's here now. And the power has to come from somewhere.
Here's where it's actually coming from: natural gas. More than 40% of U.S. data center electricity comes from gas-fired plants, per the IEA. Solar and wind deliver 24%. Nuclear provides about 20%. Coal still supplies 15%. The nuclear share comes almost entirely from existing plants — not new ones.
The inconvenient math: every major hyperscaler has pledged to be carbon-neutral or carbon-negative by 2030-2040. Microsoft by 2030. Google net-zero by 2030. Amazon by 2040. Meta has not specified. But their actual electricity consumption is growing so fast that their absolute carbon emissions are rising even as they buy more renewables. Google's 2024 environmental report showed emissions up 48% from 2019, driven almost entirely by data center expansion.
The nuclear deals are, in part, a way to make the math work on paper. Sign a contract now, book the clean energy credits later, and burn gas in the meantime. It's not fraud. It's how corporate carbon accounting works. But it does mean the "nuclear-powered AI" narrative is running years — possibly a decade — ahead of the physics.
What Actually Works Right Now
The honest answer to "how do you power AI data centers with clean energy today?" is: solar and batteries, and it's not even close.
Utility-scale solar costs $20-30 per MWh. Battery storage has fallen below $50/MWh for 4-hour duration systems. Together, they're deployable in 12-18 months from contract to grid connection. By contrast, a nuclear restart takes 3-5 years, and a new-build SMR takes 6-10 years, at costs of $100-200+ per MWh that make NuScale's $9.3 billion cancellation look like a rounding error on the industry's cumulative overspending.
Microsoft has quietly been signing solar and wind PPAs at volumes that dwarf its nuclear commitments. Amazon operates the world's largest corporate renewable energy portfolio. Google and Meta buy renewable energy certificates by the terawatt-hour. The nuclear deals get the headlines. The solar farms get the electrons.
There are real limitations to solar + storage for data centers. They need baseload — constant, 24/7 power — not intermittent supply. A data center can't brown out when a cloud passes. This is nuclear's genuine advantage: capacity factor above 90% versus solar's 20-25%. But pairing solar with lithium-ion storage, and increasingly with iron-air or sodium-ion batteries for longer duration, is closing that gap faster than reactors are getting built.
Enhanced geothermal — drilling hot-rock reservoirs using techniques borrowed from fracking — may be the real dark horse. Fervo Energy's Nevada pilot achieved $57/MWh with 24/7 baseload power. Google signed a deal with Fervo. No NRC licensing. No 14-year construction timelines. No cost-doubling. Just drill, pump, generate.
The Deeper Game
So why are the smartest companies on Earth writing checks for technology that won't deliver power for years?
Three reasons. First, optionality. If SMRs do work at scale — and it's an if, not a when — the company that locked in early supply will have an insurmountable advantage. Nuclear's capacity factor and energy density make it the ideal data center power source. The problem has always been cost and construction time, not physics.
Second, regulatory positioning. By committing billions to nuclear, Big Tech buys goodwill with policymakers who see nuclear as a pillar of energy independence. When you're asking states for permits to build 2-gigawatt data center campuses that will consume more power than a mid-sized city, having "nuclear-powered" in the pitch deck matters.
Third, land constraints. Data centers near major metro areas are running out of grid capacity. Dominion Energy in Virginia — which hosts the largest data center cluster on Earth in Loudoun County — has been warning since 2022 that it can't guarantee power to new facilities. An on-site SMR, if it ever works, solves that constraint permanently. That's worth paying a premium for an option that might not vest for a decade.
The Bottom Line
Big Tech's nuclear deals are venture capital bets cosplaying as power purchase agreements. They're buying lottery tickets on technology that doesn't exist at commercial scale, with delivery dates that slip further into the future with every quarterly update, for data centers that need power yesterday. The immediate reality is gas, the medium-term reality is solar + storage + geothermal, and the nuclear reality — if it arrives at all — is a 2030s story. The press releases make it sound like the grid is about to go nuclear. The grid is about to go natural gas with a green PowerPoint deck. The atmosphere, as always, doesn't grade on a curve.