AI Data Centers & Nuclear Power: Every Deal, in Megawatts and Pounds

60-second answer: The largest technology companies have signed a wave of nuclear deals to power AI data centers: Microsoft with Constellation to restart Three Mile Island Unit 1 (the Crane Clean Energy Center, ~835 MW), Amazon with Talen at the Susquehanna plant plus an X-energy small modular reactor (SMR) program, Google with Kairos Power for SMRs, and Oracle and Meta both signaling large nuclear procurement. These are mostly power purchase agreements and early-stage SMR commitments, not overnight capacity. Nuclear is a genuine structural catalyst for uranium demand — very roughly, a gigawatt of operating capacity implies on the order of ~0.4–0.5 million pounds of U3O8 per year — but the timelines run years to over a decade. Track every announced deal on our reactor tracker.

Type "AI data centers nuclear power" into a search bar and you get a wall of quotes and a vague sense that Big Tech is buying reactors. What you rarely get is the actual deal book: who signed what, how the deal is structured, roughly how much capacity it involves, and where it sits on the path from press release to electrons.

This is that deal book. We catalogue the major agreements, translate the buildout into approximate annual pounds of uranium using our demand model, and explain why AI demand is a real catalyst even though almost none of it produces power tomorrow.

Why AI Went Shopping for Nuclear

Training and serving large AI models is enormously power-hungry, and the load is concentrated, constant, and carbon-sensitive. Hyperscalers want electricity that is (1) available around the clock, (2) low-carbon to meet their own climate commitments, and (3) firm enough to sit next to a multi-billion-dollar data center. Nuclear checks all three boxes better than intermittent renewables alone.

That combination is why the same names that dominate cloud computing — Microsoft, Amazon, Google, Oracle, Meta — have all moved on nuclear within roughly the last two years. Some are reviving existing reactors. Some are underwriting the first commercial SMRs. The structures differ, and the difference matters for how quickly any of it turns into demand for uranium.

The Deal Book

The table below catalogues the headline AI and data-center nuclear deals. Capacities are approximate and, for SMR programs, reflect announced targets rather than firm operating capacity. Status is as of mid-2026 and moves quickly.

A few structural notes so the table isn't misread:

• Restarts (Microsoft–Constellation) are the fastest path to real power because the reactor already exists — Crane is a recommissioning, not new construction. It still needs regulatory sign-off and years of work.
• Co-location (Amazon–Talen) attaches a data center to an already-operating plant. It reallocates existing generation to a new customer; it does not, by itself, add new reactors to the grid. The arrangement drew FERC scrutiny over how co-located load interacts with the wider grid.
• SMR programs (Amazon–X-energy, Google–Kairos) are the genuinely new capacity, and also the most speculative. These are first-of-a-kind commercial reactors with development, licensing, and supply-chain risk. Announced megawatts are targets, not guaranteed.
• RFPs and permits (Meta, Oracle) are intent, not contracts. Treat signaled gigawatts as a demand indicator, not delivered capacity.

We maintain the live, sourced version of this list — with links to each announcement and status changes — on the reactor deal tracker.

From Megawatts to Pounds: What the Buildout Means for Uranium

Here is the part almost no one translates. If AI drives gigawatts of nuclear onto the grid, how much uranium does that actually require?

A useful rule of thumb: a large light-water reactor running at roughly 1 GW of capacity consumes on the order of ~0.4 to 0.5 million pounds of U3O8 per year once operating, after accounting for typical capacity factors and refuelling cycles. The exact figure depends on enrichment level, tails assay, burnup, and reactor type — so treat it as an order-of-magnitude frame, not a precise conversion. Our reactor requirements glossary entry walks through the assumptions.

Apply that qualitatively to the deal book:

• The Microsoft–Constellation restart (~835 MW) implies very roughly ~0.3–0.4 Mlb/yr of fresh demand once it is running — a single reactor's worth.
• If the announced SMR programs and RFPs from Amazon, Google, Meta and Oracle collectively delivered on the order of several gigawatts over the 2030s, that is broadly 1–2+ Mlb/yr of incremental annual demand layered on top of the existing fleet — meaningful, but arriving gradually.

Two honest caveats. First, SMRs use different fuel: many advanced designs require HALEU (high-assay low-enriched uranium), which tightens the enrichment bottleneck as much as it lifts raw pounds. Second, initial reactor cores need more uranium up front than steady-state refuelling, so first-fill demand is lumpy. The direction is unambiguous — up — but the annual pounds accrue over years, not on the day a deal is announced.

For the full picture of how this demand meets a structurally short supply side, see our uranium supply and demand model.

Why It's a Real Catalyst — and Why Timelines Are Long

The bull case is straightforward. AI power demand is large, growing, and being met partly by companies with the balance sheets to underwrite long-dated contracts. That is exactly the kind of durable, price-insensitive demand that supports the uranium term price and encourages new reactors. It is a structural catalyst, not a headline that fades.

The reality check is equally straightforward. Restarts take years. New SMRs take longer — licensing, first-of-a-kind construction, and fuel supply all sit on the critical path, and several announced units target the late 2020s to 2030s. Co-location deals reshuffle existing generation more than they add it. An investor buying "the AI nuclear trade" is buying a multi-year to decade-long thesis, with the near-term uranium impact modest relative to the eventual buildout. Our investment thesis lays out how we weigh the catalyst against the timeline.

How to Get Exposure

If the AI-nuclear story is what you want to own, the exposure splits along the same lines as the rest of the sector. Utilities running the reactors capture the power-price and PPA upside. Uranium miners carry the fuel-demand leverage. SMR developers offer the highest potential return and the highest chance of permanent loss.

• Utilities and the broad universe: our best nuclear energy stocks guide sorts the field by risk.
• Reactor builders: the SMR stocks guide covers order books and dilution, with deeper dives on NuScale Power and Oklo.
• Fuel-side leverage: the uranium stock screener ranks every listed producer and developer.

Frequently asked questions

Which tech companies have signed nuclear deals for AI data centers? Microsoft (with Constellation), Amazon (with Talen and X-energy), Google (with Kairos Power), plus Oracle and Meta, which have signaled large nuclear procurement via permits and an RFP. See the deal table above and our live reactor tracker.

Is Microsoft restarting Three Mile Island? Effectively yes — Microsoft signed a long-term power purchase agreement with Constellation to restart Three Mile Island Unit 1, rebranded the Crane Clean Energy Center (~835 MW). It is a restart of an existing reactor, targeted for later this decade, not new construction.

How much uranium does an AI data center's nuclear power actually need? Depends entirely on the capacity. As a rough frame, a ~1 GW reactor consumes on the order of ~0.4–0.5 million pounds of U3O8 per year once operating. A single restart adds one reactor's worth; the full slate of announced SMR programs would layer on demand gradually through the 2030s.

Will SMRs for data centers move the uranium price soon? Not immediately. Most announced SMRs target the late 2020s to 2030s and face licensing and construction risk, and many need HALEU fuel that also strains enrichment capacity. The demand is real and directional, but it arrives over years.

Are these deals a reason to buy uranium stocks? They strengthen the long-term demand thesis, but they are a multi-year catalyst, not a same-day one. How you get exposure — utilities, miners, or SMR developers — depends on your risk tolerance; see our best nuclear energy stocks guide.

This article is for informational purposes only, not investment advice.