AI has created a new, price-insensitive buyer of nuclear power, and that means a new buyer of uranium. The IEA projects global data centre electricity use will more than double, from about 415 TWh in 2024 to about 945 TWh in 2030. The four largest hyperscalers have already signed roughly 9.8 GW of nuclear deals.
- Demand shock: US data centre consumption alone is projected to rise by about 240 TWh (+130%) by 2030.
- Nuclear is the preferred fix: it is the only proven source that delivers carbon-free power 24/7 at over 90% availability.
- Fuel math: each gigawatt of reactor capacity burns roughly 0.42M lbs of U3O8 a year, and a new reactor's first core needs about three times that up front.
- Policy alignment: Washington's target is to quadruple US nuclear capacity from about 100 GW to 400 GW by 2050.
- Where value lands: utilities and tech buyers want secure, non-Russian, North American pounds. That puts a premium on US and Canadian developers with permitted or permittable deposits.
The AI power problem
AI is turning electricity into the main constraint on computing. In its base case, the IEA sees data centres going from about 1.5% of world electricity in 2024 to just under 3% by 2030 (IEA).
- Accelerated servers are the engine. The IEA expects power use by AI-accelerated servers to grow about 30% a year.
- The US carries the largest share. US data centre consumption is projected to rise by about 240 TWh, or 130%, from 2024 to 2030.
- Longer runway. Some forecasts have US data centre demand tripling or quadrupling by 2035. A mid-range estimate is about 560 TWh a year (Carnegie Endowment).
The added demand is also a different kind of load. AI training clusters run around the clock at near-constant draw. That suits power sources that never switch off and strains ones that depend on weather.
Why nuclear wins the AI load
Nuclear is the only power source that is firm, carbon-free and already proven at gigawatt scale. Hyperscalers have to meet both their uptime targets and their net-zero pledges, and nothing else does both today.
- Nuclear: ~90%+ capacity factor. It runs 24/7 and is carbon-free, which makes it a strong fit: firm, compact, with decades-long life.
- Natural gas (combined cycle): ~55–60% capacity factor and runs 24/7, but it is not carbon-free. It is fast to build, but conflicts with net-zero goals and is exposed to fuel prices.
- Wind: ~35% capacity factor. It is intermittent, so it needs storage plus backup to serve flat load.
- Solar: ~25% capacity factor. It is intermittent with a daily cycle, so it has the same limitation as wind.
- Fuel is a small share of cost. Uranium is a small part of what nuclear power costs to generate, so buyers can pay far more for fuel without changing the economics. That makes nuclear demand price-insensitive.
- Contracts are long. Microsoft's Three Mile Island (Crane) agreement runs 20 years, and Meta's Clinton deal runs through 2047. Each locks in decades of fuel buying.
- Existing plants live longer. Hyperscaler contracts fund relicensing and uprates, which keeps reactors that might have closed buying uranium for another 20+ years.
Capacity factors are rounded US averages from public EIA data and vary by year.
Hyperscaler nuclear deals
The four largest cloud companies have signed about 9.8 GW of nuclear deals across 13 projects as of May 2026 (SMR Intel). Carnegie counts up to 13 GW once all announced partnerships are included (Carnegie).
- Microsoft × Constellation: restarting Three Mile Island Unit 1 (Crane), 835 MW. 20-year PPA worth ~$16B, power expected in 2027.
- Meta × Constellation: Clinton, Illinois, ~1,100 MW. PPA through 2047, already operating.
- Meta × Vistra: existing fleet plus uprates, up to ~3 GW. PPAs, operating now and expanding from 2030.
- Meta × TerraPower: Natrium sodium reactors, 345 MW first unit, with a fleet option reported up to ~4 GW. Fleet agreement, 2032–2035.
- Meta × Oklo: Aurora powerhouse campus, ~1.2 GW. Campus development, 2030 onward.
- Amazon × Talen Energy: Susquehanna, Pennsylvania, ~960 MW. Co-located campus plus PPA, already operating.
- Amazon × X-energy / Energy Northwest: Xe-100 SMRs, ~960 MW. ~$700M equity investment plus offtake, 2030s.
- Amazon × Dominion: Lake Anna SMR, ~960 MW potential. Development study, 2032 onward.
- Google × Kairos Power: KP-FHR SMR fleet, 500 MW. Corporate SMR agreement, 2030.
- Google × NextEra: restarting Duane Arnold, Iowa, ~600 MW. PPA, 2028 onward.
Two patterns stand out:
- Restarts and life extensions come first. Crane and Duane Arnold bring closed reactors back into the fuel market, so their uranium demand is fully new.
- SMRs are the second wave. Kairos, X-energy, TerraPower and Oklo designs need first fuel loads before 2035, and several of them use HALEU.
Trackers differ on some capacity figures. Numbers shown are as reported and should be checked against company filings before external use.
Translating megawatts into pounds
Each gigawatt of reactor capacity burns about 0.42M lbs of U3O8 a year. That figure comes from the World Nuclear Association's generic 189 t U3O8 per GWe per year (WNA). For scale, today's roughly 400 GWe global fleet needs about 69,000 tU a year, or about 180M lbs U3O8.
- Reactor restarts already signed (Crane + Duane Arnold): ~1.4 GW, or ~0.6M lbs U3O8 a year (~0.3% of today's global demand).
- All hyperscaler deals: ~9.8 GW, or ~4.1M lbs a year (~2%).
- Nuclear supplies 20% of the IEA's 2024–2030 data centre growth: ~13 GW, or ~5.6M lbs a year (~3%).
- US reaches its 400 GW goal (+300 GW): ~125M lbs a year (~70% of today's global demand).
Our calculations: 0.42M lbs per GW-year, and 90% capacity factor (about 7.9 TWh per GW-year). The 20% scenario is illustrative, not a forecast.
- Initial cores multiply near-term buying. A new reactor's first fuel load is roughly three annual reloads, so about 1.25M lbs per GW, bought 1–2 years before startup.
- Even contracts for existing plants add demand. Most of the 9.8 GW is existing reactors. The real effect there is decades of extended operation, not new pounds today.
- The upside is in the build-out. At the margin, AI demand is what turns policy targets like 400 GW into financed projects. That is where the step-change in uranium demand sits.
SMRs and the next wave
Small modular reactors are how hyperscalers plan to move from buying nuclear power to putting reactors next to their campuses. The first commercial units are targeted for about 2030.
- Built at the campus. Units of 50–350 MW can sit beside a data centre, so they don't wait years to connect to a congested grid.
- The pipeline is real money. Amazon has put about $700M into X-energy. Google has a 500 MW Kairos order, and Meta has agreements with both TerraPower and Oklo.
- New fuel needs. Several advanced designs (Kairos, TerraPower, Oklo, X-energy) use HALEU, enriched to 5–20%. Producing HALEU takes much more natural uranium per unit of fuel, which raises the pounds needed per gigawatt.
- Timelines are uncertain. Kairos targets 2030, and TerraPower and Oklo plan 2032–2035. First-of-a-kind manufacturing and licensing could push these back (Carnegie).
Policy tailwinds
US policy now treats nuclear as strategic infrastructure for AI and national security. The May 2025 executive orders set a goal of quadrupling US nuclear capacity, from about 100 GW to 400 GW by 2050 (White House).
- Faster licensing. The NRC has a fixed 18-month deadline to decide on new reactor licences.
- Regulatory overhaul. The NRC must rewrite its rules within 18 months. Designs tested by the Department of Energy or Department of Defense get an expedited path.
- A historic reversal. Only two new US reactors have started construction and entered commercial operation since 1978. The policy aims to end that drought.
- Fuel security. The US ban on Russian enriched uranium (Driver 4) adds to this policy push. AI-driven reactor demand will need to be fuelled from Western supply chains.
Risks and counterpoints
The AI thesis is strong but not risk-free, and investors should weigh these four points.
- AI efficiency gains: more efficient chips and models could flatten power growth (the IEA's High Efficiency case). Mitigant: cheaper AI tends to drive more usage, and nuclear contracts are already signed for 20+ years.
- Nuclear timelines: new reactors take 10–15 years, and SMRs may slip past 2035. Mitigant: restarts and uprates add demand now, and utilities buy fuel years ahead.
- Gas wins near-term: hyperscalers fill the 2025–2030 gap with gas turbines. Mitigant: gas covers speed and nuclear covers the long term, so both get built.
- Coverage gap: signed deals cover under 20% of projected US data centre demand to 2035 (Carnegie). Mitigant: the gap is upside, since every new deal adds demand on top.
Investment implications
AI adds a durable, long-dated source of demand to a market that already mines less uranium than reactors burn. The pounds that benefit most are North American ones.
- Demand is now strategic, not cyclical. Microsoft, Amazon, Google and Meta are among the world's best-funded buyers, and they are signing 20-year commitments. That puts a floor under nuclear demand in a way past cycles never had.
- Where the uranium comes from now matters. US buyers want fuel chains free of Russian content. US and Canadian deposits command a strategic premium over equivalent pounds from higher-risk jurisdictions.
- Development-stage assets gain value. New mines take 10+ years from discovery to production. Projects already advancing through permitting are best placed to supply the 2030s reactor wave, which includes restarts, uprates, large reactors and SMRs.
- The re-rating is still ahead. The deals signed so far cover a fraction of projected AI load, so each new hyperscaler or SMR announcement is a potential catalyst.
This report is for information only and is not investment advice. Forward-looking figures are estimates drawn from third-party sources and our own calculations.
Sources