Scott Nolan runs a uranium enrichment startup that won a $900M DOE contract 24 months after founding. He walks a chain of bottlenecks from AI down to one missing industrial step the US no longer does.
Everything in AI converges to the cost of electricity, because chips and models keep getting cheaper but power is what you actually consume to run models. Follow that constraint down and you get a chain: AI needs power, clean base-load power at scale means nuclear, nuclear needs fuel, and fuel needs enrichment, a step where the US holds under 0.1% of world capacity and still imports from Russia. Nolan's argument is that enrichment is the deepest bottleneck on AI scaling on a five-year horizon.
Power over chips
The class frames AI as a factory manufacturing intelligence, with compute as one bottleneck. Nolan zooms out: powering the data centers is a separate, larger problem. Even with a data center ready, if you cannot get power to it, you cannot train models. He cites Sam Altman testifying to the Senate that all costs converge to the cost of energy, Jensen Huang and Elon Musk agreeing, and Balaji arguing all costs should be denominated in joules.
The demand shock
ChatGPT in late 2022 was the first consumer killer app and triggered a compute and energy crunch by early 2023, because taping out chips and standing up data centers takes about two years. Nolan marks Claude 4.6 in December 2025 as the enterprise killer app: adults came back from winter break, used it at work, and enterprise demand spiked. US electricity demand is now super-linear while grid growth has been near flat for 20 years, so the country must go from a standstill to a China-like or steeper build rate.
From stranded power to nuclear
Five years ago the play was stranded energy: unused hydro, geothermal, or wind in remote areas, first monetized by Bitcoin miners like Crusoe, which needed little connectivity. Those cheap resources are now claimed. Data centers need uptime and base load, so the field moved to natural gas turbines, but turbine lead times have stretched to years. On safety and emissions, nuclear scores best, lowest carbon and roughly tied with wind for safest, which is why hyperscalers are turning to it despite a 5 to 10 year ramp.
The enrichment gap
Every reactor burns fuel and must be refueled every year or two. Fuel comes from five steps: mine, convert to gas, enrich, convert back to solid, form pellets. Enrichment is the middle step, and the US holds under 0.1% of world capacity today, down from about 86% in the 1980s. The country cannot produce its own nuclear fuel at scale and still imports from Europe and Russia despite sanctions. Enrichment is also the biggest cost in advanced fuel, so it becomes the bottleneck all the way up to AI.
How the US lost it
After the Berlin Wall fell, the Megatons to Megawatts program down-blended Russian warheads into reactor fuel, and cheaper European and Russian supply plus free trade made US enrichment uneconomical. The last US commercial enrichment site shut in 2013. Nolan calls it path dependency: an expensive domestic technology, disarmament supply, and free trade combined to sunset the whole capability, and the need came back faster than anyone planned for.
Building the primitive
General Matter treats enrichment as a fundamental primitive, like SpaceX reducing space to dollars per kilo to orbit. Founded January 2024 by a first-time hardware founder, staffed with people from national labs, Tesla, and SpaceX, it won a $900M DOE contract 24 months later with a ~100-person team. Its facility is in Paducah, Kentucky, the same city as the last US commercial enrichment site, on 100 undeveloped acres. Nolan's advice: ignore surface narratives and memes, go many clicks deeper, and work on the important problem your skills uniquely fit.
- Chips and models keep getting cheaper, so AI costs converge on the cost of electricity, which is what you consume to run models.
- A ready data center is useless if you cannot get power to it, which is why Nolan draws electricity larger than compute on his factory map.
- US electricity demand is now super-linear while grid expansion has been nearly flat for 20 years, requiring a jump from standstill to near-vertical build.
- Stranded power, first used by Bitcoin miners, has mostly been claimed, pushing builders to gas turbines with multi-year lead times.
- Nuclear has the lowest carbon emissions of any source and is roughly tied with wind for safest, but ramps over 5 to 10 years, not overnight.
- The US holds under 0.1% of world enrichment capacity today versus about 86% in the 1980s and still imports fuel from Russia despite sanctions.
- General Matter won a $900M DOE enrichment contract 24 months after founding with a team of about 100 people, evidence AI is creating new physical-economy jobs.
In their words
“Everything is going to converge to the cost of energy, to the cost of electricity.”
“Even if you have a data center ready to go, if you can't get power to it, doesn't matter. It's over.”
“The US is actually unable to produce its own nuclear fuel at any scale whatsoever.”
Terms to know
- Enrichment
- Refining mined uranium to raise the concentration of the fissile isotope U-235 to the level a reactor needs; the middle of five fuel-making steps.
- Base load
- Power that runs continuously and reliably, the always-on floor of supply that data centers need for uptime, as opposed to intermittent solar or wind.
- Stranded energy
- Power supply with no nearby demand, like remote hydro, geothermal, or West Texas wind, cheap to grab but now mostly claimed.
- HALEU
- High-Assay Low-Enriched Uranium, a more enriched fuel grade that advanced small and micro reactors require.
- Megatons to Megawatts
- A post-Cold War program that down-blended Russian warhead uranium into US reactor fuel, which helped make domestic enrichment uneconomical.
- Terawatt
- A unit of power equal to a trillion watts; a benchmark for how much new generation AI demand may need within a decade.
Scott Nolan at Stanford CS 153: Frontier Systems
New to this? Come build with us.
Reading is good. Building with people is better. Our drop-ins are free and open to total beginners.