Constellation inked a 20-year power purchase agreement to restart Three Mile Island Unit 1. 835 megawatts. 2028 delivery. Microsoft on the other end of the wire. Every headline called it the dawn of the nuclear-AI era.
Here's what the headlines buried: the uprate pipeline โ the cheapest, fastest nuclear capacity America can actually add โ is nearly dry. Not dry in 2035. Dry now. The uprate was supposed to be the bridge, cheap megawatts from machines already standing. Instead, it's the bottleneck. And the people who already figured that out aren't the ones in the boardroom. They're the ones who spent the last four years fighting for grid capacity in a market nobody was watching.
I've watched this movie before. Chasing the white whale in the 2017 ether rush, I learned that the crowd always prices the narrative before it prices the constraint. The ICO boom ran on whitepapers; the physics arrived later and took everyone's money. Nuclear is running the same play, and the constraint is already here โ it's just not in the reactor.
Let me frame the numbers. The United States operates about 94 commercial reactors, roughly 97 gigawatts, running at a 92-93% capacity factor. That fleet supplies 18-20% of national electricity and โ this is the part that matters โ about 47% of all carbon-free generation.
Now the demand side. Data centers consumed roughly 176 TWh in 2023, about 4.4% of US power. LBNL and EPRI forecasts put 2028 consumption at 325-580 TWh. Multiple analysts push the 2030 data-center share toward 9% and beyond. Single hyperscalers are projecting gigawatt-scale annual additions.
The gap between those two curves is the entire trade.
Into that gap stepped the "uprate" thesis โ squeezing more megawatts out of reactors that already exist. It's the cheapest nuclear capacity available. No new site. No new license to build from scratch. No decade of environmental review. Just run the existing machine harder.
The problem is that uprates come in tiers, and the cheap tiers are exhausted.
This isn't a new problem. Crypto mining ran the same energy gauntlet between 2021 and 2024. Cheap stranded power dried up, jurisdictions banned new load, and after the fourth halving, block rewards collapsed while hash rate kept climbing. Miners who couldn't secure long-term power contracts got squeezed out. The survivors weren't the fastest โ they were the ones who locked in the cheapest, most dispatchable megawatts years in advance. Big Tech is now learning the lesson crypto already paid for.
Here's the stratification nobody publishes on a chart.
Measurement Uncertainty Recapture (MUR) uprates add roughly 1.0-1.7% capacity and cost in the low single-digit millions per reactor. Stretch uprates add 2-5%. Both are essentially free money โ modest instrumentation upgrades, better flow measurement, minor thermal tweaks.
Then there's the Extended uprate. Extended uprates add 5-15% or more, but they require replacing the turbine, the generator, and often the main transformer. That's not a tuning job. That's a capital project with new-build economics.
Since 1977, the NRC has approved roughly 170 uprates totaling about 8 GW โ against a 97 GW fleet, that's an 8% uplift accumulated over nearly five decades. The MUR and Stretch fruit has been picked clean. What's left is Extended, and Extended costs what it costs.
Run the math. A single hyperscaler's annual demand growth can hit gigawatts. The entire historical US uprate program delivered 8 GW across 48 years. The mismatch is an order of magnitude, not a rounding error.
And here's the bottleneck the nuclear bulls never mention: it isn't nuclear physics. It's manufacturing. Extended uprates need steam turbines, generators, and transformers from a handful of suppliers โ GE Vernova, Siemens Energy, Mitsubishi Power โ whose production slots are booked solid, largely by gas turbine orders, through 2028. You can't uprate a reactor if nobody can build the conventional island.
The order books tell the story. Gas turbine demand from AI data centers and grid operators has pushed lead times to multi-year highs, and the same industrial base that builds combustion turbines builds the steam-side equipment nuclear uprates need. The two demands are competing for the same scarce factory floor.
The crypto parallel is exact, and it's why this feels familiar. After the 2024 halving, miner revenue per unit of hash fell hard. Hash rate didn't fall with it โ it kept rising, concentrating into a handful of pools and well-capitalized operators. Nuclear uprates are the same story in different units: the marginal capacity keeps getting more expensive while demand keeps climbing. Both industries are running headlong into the same wall โ the cheapest increments always go first.
This is where my trader's lens kicks in. Hunting spreads while the market sleeps taught me that the real opportunity is never in the obvious asset โ it's in the constraint upstream of it. Everyone's buying nuclear operators. Nobody's buying the forge.
The upstream squeeze runs deeper. Enrichment is concentrated in Rosatom (historically 40%+), Urenco, and Orano. The US banned Russian low-enriched uranium imports in 2024 but built in a waiver transition โ an admission that near-term decoupling is impossible. And advanced reactors need HALEU, high-assay low-enriched uranium under 20% enrichment, where Western commercial capacity is essentially blank. Centrus runs demonstration-scale only.
So the binding constraint isn't the reactor, the fuel, or even the turbine. It's the interconnection queue. LBNL tracking shows a terawatt-scale backlog of projects waiting for grid connection, with average waits stretching into years. Extended uprates add output that still needs grid and equipment accommodation. Data centers need to connect somewhere, and the transmission corridors are jammed.
There's a quieter problem the nuclear-AI narrative refuses to price. When a hyperscaler signs a PPA that "reserves" existing nuclear output, that carbon-free electricity gets reassigned on paper. Under location-based accounting, the local grid still sees the same electrons. Under market-based accounting, the data center claims clean power while the public grid fills the gap with higher-carbon generation. The result is an accounting win and a physical wash โ possibly a physical loss. Total system emissions can rise even as one company's sustainability report improves. Nobody signing these deals wants to say that out loud, but the math doesn't care about the press release.
Watch the demand side too, because it's where the real flexibility lives. Virtual power plants, grid-scale storage, and demand response can't run 24/7 like a reactor, but they deploy in months, not decades. When nuclear supply misses, the gap doesn't stay empty โ it gets filled by whatever can show up fastest. In practice, that's gas peakers and batteries, backed by aggressive efficiency programs. The nuclear headline grabs attention; the demand-response contract quietly closes the deal. Speed beats scale when the grid is the bottleneck.
Volatility is just noise until it becomes signal. The signal here is that nuclear's "easy wins" aren't running out in the reactor hall. They're running out in the enrichment plant, the forge, and the utility's queue spreadsheet.
Now the angle nobody's trading.
The consensus is that SMR developers win. Oklo, NuScale, X-energy โ the pure-plays catching every headline. I audited AI-agent revenue-sharing mechanisms on Solana in 2025 and found a fee-distribution flaw that created centralization risk across 15 agents. The lesson transfers: the entity that controls the settlement layer captures the value, not the entity with the best story. In nuclear, the settlement layer is enrichment, heavy forgings, and grid access.
Look at the timelines and the mismatch screams. SMR developers target commercialization in the 2030s. NuScale, Kairos, X-energy, GE-Hitachi's BWRX-300 โ all penciled into the next decade. The hyperscaler energy gap is a 2026-to-2030 problem. You cannot solve a five-year problem with a ten-year asset. That gap gets filled by gas, storage, and efficiency โ not reactors. The nuclear trade is a 2030s option dressed up as a 2020s solution.
The "sell shovels" thesis โ enrichment, components, turbines โ is the highest-certainty trade in this cycle. The "gold miners" โ SMR developers โ are burning cash and delaying revenue into the 2030s. NuScale's UAMPS cancellation in 2023 was the warning shot.
That's the trade nobody's positioning for. Not the reactor operator, not the SMR dreamer โ the component supplier whose order book is already full and whose pricing power is quietly compounding while everyone else argues about megawatts.
And here's the crypto-native tell. The RWA crowd is already pitching tokenized power, on-chain energy credits, nuclear-backed yield. I've watched that story run for three years. Traditional institutions don't need a public chain to sign a PPA โ Constellation and Microsoft proved that with a PDF and a press release. The tokenization narrative is a story retail tells itself while the institutional money moves through private contracts.
The real crypto winners are quieter. Bitcoin miners who survived the post-halving reward cut by pivoting to AI and HPC hosting already own the scarce asset โ energized sites with interconnection queue positions. They bought the grid access years ago, when nobody wanted it. Now hyperscalers are paying premiums for exactly that. The miners who read the physics first are the ones holding the door.
Stop watching who signs the next nuclear press release. Start watching who controls carbon-free, dispatchable, grid-connectable power โ and who owns the queue positions to deliver it. The winners won't be the loudest storytellers. They'll be the ones who bought the shovels, the fuel, and the line into the substation before the crowd showed up. The physics always prices in last. The next bull case isn't nuclear. It's whoever already owns the wire.