Scott Nolan from General Matter on Energy Bottlenecks

ELI5/TLDR

If AI is a factory, the data center is one room and the power plant feeding it is the much bigger room next door. Right now the bigger room is the choke point — the US grid has barely grown in 20 years, but AI demand is asking it to go nearly vertical. The only base-load power source clean and dense enough to keep up is nuclear, and the US makes essentially zero of its own nuclear fuel. Scott Nolan’s company, General Matter, is rebuilding the missing middle step — uranium enrichment — in Kentucky, and just won a $900M Department of Energy contract 24 months after starting.

The Full Story

The bottleneck behind the bottleneck

Most conversations about scaling AI stop at compute — chips, GPUs, data centers. Nolan’s lecture is about what comes one step earlier on the supply chain. A finished data center sitting without power is just a building. And right now, getting power to where the data center wants to live is the hardest step.

The numbers make this concrete. US grid expansion has been roughly flat for two decades. Forecasts now say AI alone will push electricity demand toward a terawatt-scale jump in roughly ten years. Drawn on a chart, the required curve is “nearly vertical” — a slope steeper than China’s recent build-out, from a near-standstill. The country has to do something it has not done in any living engineer’s career.

Nolan stacks the testimonies up: Sam Altman told the Senate that everything in AI eventually converges to the cost of energy. Jensen Huang, who has every incentive to say chips are the bottleneck, said on Joe Rogan that energy is. Elon Musk has said the same. The Financial Times has caught up. So have the engineers building the data centers.

Why nuclear, by elimination

The lecture walks through the options with the patience of someone who has actually had to build them. Solar and wind are intermittent — making them base-load means stacking enough batteries to make the math fall apart. Natural gas turbines have become the recent default, but turbines are now sold out two-plus years ahead and the manufacturers are not ramping fast enough.

“Maybe we don’t want to put out a lot of carbon. Maybe we want it to be pretty safe… nuclear is pretty good. Um, it’s actually lowest carbon emission of any of them and it’s essentially tied for safest with wind.”

That is the punchline of a chart most people have never looked at — deaths per terawatt-hour, including every accident on the books. Nuclear, despite Three Mile Island, Chernobyl, and Fukushima, sits at the bottom alongside wind. Three Mile Island had no measurable direct deaths. Fukushima had perhaps one fatality from radiation; the thousands killed that day were killed by the tsunami. Pricing in safety and emissions, the search basically funnels you to nuclear.

The hyperscalers — Microsoft, Google, Amazon, Meta — have figured this out. Hence all the recent power purchase agreements with existing reactors and the deals with small modular reactor (SMR) startups. The catch is that nuclear is not a one-year build. New reactors are a 5-to-10-year timeline. So the next two years are a scramble — stranded power, turbines, anything you can hook up — while the longer arc bends toward nuclear.

The bottleneck behind the bottleneck behind the bottleneck

If electricity is the bottleneck to AI, and nuclear is the bottleneck to electricity at scale, then what is the bottleneck to nuclear?

Fuel. Reactors are not perpetual motion machines — they need refueling every one to two years (five to ten for some advanced designs). Fuel comes from a five-step pipeline: mine the ore, convert the powder to a gas (UF6), enrich the gas to concentrate the U-235 isotope, convert it back to a solid, then form fuel pellets.

Step three — enrichment — is where the US has, in Nolan’s words, “less than 0.1% market share.” Zero. The country that built the Manhattan Project and ran 86% of world enrichment capacity in the 1980s now imports from Europe and, even after sanctions, still from Russia.

How did this happen? Path dependency. After the Berlin Wall fell, the US and Russia ran a program nicknamed “megatons to megawatts” — taking warheads, down-blending the highly enriched uranium, and burning it as reactor fuel. Combined with cheap European enrichment, the US’s older, more expensive plants couldn’t compete. The last one shut down in 2013. Plans probably existed to bring it back “when we needed to.” The need arrived faster than anyone planned for.

General Matter’s bet

Nolan was at Founders Fund for over a decade looking at hard-tech. He noticed in 2020 that every interesting nuclear startup he met had the same problem: no fuel, and the only fuel they could get came from Russia. He spent most of 2023 marinating on whether enrichment was the right thing to work on. By January 2024 he had a company. By January 2026 — 24 months in, with a team of about 100 — General Matter had been awarded a $900 million DOE contract to rebuild US enrichment capacity at a 100-acre site in Paducah, Kentucky, the same town that hosted the last commercial enrichment facility before 2013.

“We chose to work on a problem that we knew was completely we believed would not be solved and was completely important and required urgent action. And fortunately the DOE Congress had already been thinking about this. They had already funded these programs. We didn’t have to convince them.”

The team is the SpaceX-Tesla playbook applied to fuel: pull experienced people from National Labs and existing nuclear firms, cross them with operators from Tesla and SpaceX who know how to break into a “capital-intensive incumbent dominated stagnant industry.” The first months were 100-hour weeks, “living and sleeping at the headquarters.” The technology pitch is enrichment that works for both LEU (the standard ~5% fuel for the existing reactor fleet that supplies 20% of the US grid) and HALEU (the higher-enriched fuel needed by SMRs and advanced reactors).

A few asides worth keeping

Bitcoin was a dress rehearsal. Crusoe started by mining Bitcoin off stranded gas — methane that was being flared into the atmosphere anyway. The company learned how to bolt power generation onto remote sites with minimal connectivity. That same playbook is now powering the Stargate data center in West Texas. The host’s point: the cultural backlash against crypto sometimes throws out genuinely useful infrastructure primitives that were developed during it.

Germany is the cautionary tale. It shut down working reactors planning to replace them with renewables. What actually replaced them was coal and gas. Air-quality maps tell the story — France (heavily nuclear) is blue, Germany is red. Nolan calls it “self-defeating to turn off affordable energy, very cheap energy once it’s built that’s clean.”

Public opinion has flipped. The polling on nuclear has crossed from majority-negative to majority-positive in the past few years. The political support has been bipartisan since the Biden era and continued into the current administration. Of all the surprises in the lecture, this might be the quietest one.

On working at SpaceX. Nolan joined as an intern when SpaceX was 35 people, worked on propulsion test stands, and left after the first Falcon 1 launches because he thought the company “had become big” at around 100 people. He calls this his clearest professional mistake — 100 people, he says now, is exactly when the truly hard scaling work starts. The friends who stayed got to work on the coolest projects of the past decade.

Key Takeaways

Energy, not compute, is the operative bottleneck for AI scaling over the next decade. Compute can be built faster than the grid that powers it.
US grid capacity has been roughly flat for 20 years; AI alone now needs an almost-vertical demand curve, steeper than China’s recent build-out.
By elimination — solar/wind need too many batteries, gas turbines are sold out two-plus years — the long-term answer for base load is nuclear.
Nuclear is the lowest-carbon source on the chart and statistically tied with wind for safest, even after counting every historical accident.
Reactor fuel is a five-step pipeline; the US has effectively 0% global share in step three (enrichment) despite running 86% of world capacity in the 1980s.
The cause was path dependency: post-Cold War “megatons to megawatts” plus cheap European competition killed older US plants by 2013.
General Matter is rebuilding US enrichment in Paducah, KY, won a $900M DOE contract 24 months after founding, and aims to be online before 2030.
Hyperscalers know this — recent reactor deals (Microsoft/Three Mile Island restart, etc.) are not PR; they reflect a multi-year supply scramble.
Useful career framing from Nolan: pick the most important problem that won’t get solved by someone else and that your skill set fits — more important than which org type (startup, big co, government, nonprofit).
“100 people is actually just where you get into the really hard stuff of scaling a business” — leaving early is often the wrong move.

Claude’s Take

This is a good lecture because it does what most AI-infrastructure talks don’t: it walks the supply chain backward instead of forward. Most people stop at “we need more GPUs.” Nolan walks past the GPUs, past the data center, past the grid, past the reactor, into the enrichment plant. Each step earns its place in the argument. By the time you arrive at “the US makes none of its own nuclear fuel,” you’ve been given enough connective tissue to understand why that one obscure fact actually constrains how big AI can get.

The honest question is whether the bottleneck framing is overstated. Two pieces of skepticism worth flagging. First, “nuclear is the only answer” is partly true and partly a function of who’s doing the talking — Nolan runs a nuclear fuel company, and natural gas plus storage is going to do most of the heavy lifting through the late 2020s no matter what nuclear does. Second, the construction timelines for new reactors in the US have a long history of disappointing even sober estimates. Vogtle units 3 and 4 took 15 years and ran tens of billions over budget. The “5-to-10-year timeline” Nolan quotes for new gigawatt-scale builds is the optimistic case, not the base case. SMRs may compress this, but they haven’t proven it yet at commercial scale.

That said, the enrichment argument is the strongest part of the lecture and it does not require you to believe nuclear will win — only that nuclear will be a meaningful fraction of the answer. If even a fraction of the planned SMRs and reactor restarts happen, the US will need fuel, and right now it can’t make any. That gap is the real product. An 8 because the framing is genuinely useful — bottleneck-of-the-bottleneck thinking applies far beyond energy — and because Nolan is unusually clear-eyed about what his company is and isn’t solving. Half a point off for the “nuclear is the only answer” simplification and the implied timeline optimism.