Big Tech's AI Reactors Need 40 Tons of Fuel a Year by 2030. America Makes Less Than One.
Meta, Google, Amazon, and Microsoft have committed to nearly 10 gigawatts of nuclear power for their data centers. Most of the reactor designs behind those deals run on a fuel the U.S. only started making commercially last year — at a small fraction of what the pipeline will need.
The fuel gap: capacity vs. demand, 2025 → 2035
Pick a year or hit Play to watch U.S. HALEU production capacity grow against the Department of Energy's demand estimate. Bar heights use a square-root scale so the 2025 figure stays visible next to 2035 — read the printed numbers for exact values. Only 2025, 2030, and 2035 are sourced data points (see callout); the animation is a visual device, not a year-by-year forecast.
The plain version
Uranium fuels a reactor only after it's been "enriched" — its rare, fissile isotope U-235 concentrated up from the roughly 0.7% found in nature. Today's conventional power plants run on uranium enriched to about 4–5%. But the new generation of small, futuristic reactors that tech companies are betting on for AI data centers — from Oklo, X-energy, Kairos Power, and TerraPower — need uranium enriched further, to somewhere between 5% and 20%. That fuel has a name: HALEU, for High-Assay Low-Enriched Uranium.
Think of enrichment like distilling alcohol: natural uranium is a very weak wash, today's reactor fuel is beer-strength, and HALEU is closer to moonshine — the same basic process, just run harder. Only a handful of "distilleries" on Earth are licensed and trusted to run it that far.
For years, the only one selling HALEU commercially was Russia's state nuclear company. That got politically uncomfortable fast, so in 2024 the U.S. banned imports of enriched uranium from Russia. Around the same time, AI's power hunger pushed Meta, Google, Amazon, and Microsoft to sign deals for close to 10 gigawatts of nuclear power — and many of the reactor designs behind those deals are exactly the HALEU-hungry kind.
There's just one problem: America's only company licensed to make HALEU at scale, Centrus Energy, produced its first real batch only in 2025 — 900 kilograms, roughly what a minivan could haul. The government estimates the country will need around 40 metric tons a year by 2030. Centrus's expansion plan gets it to about 6 tons annually, and not until 2029. The reactors might get built on schedule. The fuel to run them is the part still catching up.
The expert version
Enrichment separates the fissile isotope U-235 from the far more abundant U-238, typically via cascades of gas centrifuges spinning uranium hexafluoride (UF6). Commercial light-water reactors run on Low-Enriched Uranium (LEU) at roughly 3–5% U-235. High-Assay LEU (HALEU) extends that to 5–20% — still well below the 20% threshold that defines weapons-usable Highly Enriched Uranium, but dense enough to let advanced reactor designs shrink core size, extend refueling intervals, and hit the power densities their economics assume. X-energy's Xe-100 (TRISO pebble bed), Kairos Power's Hermes (fluoride-salt-cooled, TRISO), Oklo's Aurora (metal-fueled fast reactor), and TerraPower's Natrium (sodium-cooled fast reactor) are all designed around it.
Until 2024, the only enricher selling HALEU commercially was Rosatom's TENEX. The Prohibiting Russian Uranium Imports Act (signed May 2024, effective August 11, 2024) banned unirradiated Russian LEU imports, with a Department of Energy waiver process running through January 1, 2028, and unlocked $2.72 billion for domestic enrichment — roughly $700 million earmarked for HALEU specifically.
Centrus Energy's Piketon, Ohio facility, built around a DOE-owned AC100 centrifuge design at the site of a former Cold War gaseous-diffusion plant, brought a 16-machine cascade to a sustained 900 kg/year rate in 2025, extended under contract through mid-2026. In January 2026, DOE awarded Centrus a roughly $900 million task order, with Fluor as EPC contractor, to build a full 120-machine cascade targeting 6 metric tons/year by 2029, alongside direct supply contracts with Oklo (deliveries beginning 2029) and Radiant.
DOE's own fuel-cycle planning estimates put domestic HALEU demand at roughly 40 metric tons/year by 2030 to support the announced advanced-reactor fleet, rising to about 50 metric tons/year by 2035. Other entrants — Urenco's Capenhurst, UK line (targeting 2031) and smaller ventures like General Matter — add capacity, but none clears meaningful volume this decade. A reactor is a decade-scale capital project; the enrichment cascade meant to fuel it is proving almost as slow to build.
Why it matters for tech + supply chain: the constraint on AI's nuclear pivot isn't reactor licensing or construction schedules — it's a slow-to-build enrichment step now controlled by two governments, soon a third, which means whether Meta's or Google's 2030-era reactors actually have fuel depends on a Department of Energy contract timeline, not a Silicon Valley one.
Why it matters for tech + supply chain: HALEU enrichment capacity is now the long-pole item in every advanced-reactor PPA tech buyers have signed — it reframes "nuclear for AI" capex risk away from siting and licensing (real, but increasingly de-risked by state and federal streamlining) and toward a single-digit list of enrichment projects, most pre-revenue, that must scale 40–80x this decade to match announced offtake.