Microsoft is paying to restart a reactor that melted down partially in 1979. The Three Mile Island Unit 1 deal, signed with Constellation, runs 20 years, costs around $16 billion in committed offtake, brings back 835 MW, and is scheduled to come alive in 2028 under a rebranded name so nobody has to say “Three Mile Island” in a press release. Big tech signed more than 10 GW of new US nuclear capacity in the year running up to December 2025. That number is the whole story and the whole con at once, and untangling which half is which is the reason I finally sat down to write the piece I said I would leave alone.
I parked the nuclear angle in my last post on grid interconnection because that one was about the wire and this one is about the generator. The wire is the bottleneck nobody can buy their way out of fast. The generator is the thing everybody is now trying to buy their way out of the wire with: if you can drop a reactor next to the racks, you stop begging the grid operator for a queue slot. Co-location is the whole pitch. And the moment you look at who can actually pour concrete on a reactor next to a data center before 2030, the map stops looking like the press releases and starts looking like something much more lopsided.
The American approach: capital first, reactors later
The US strategy is money. An absurd amount of it, moving fast, into a sector that a few years ago I would have called a graveyard. Amazon is the aggressive one here. It bought a 960 MW nuclear-adjacent campus from Talen Energy for $650 million back in early 2024, 1,200 acres in Pennsylvania sitting next to the Susquehanna plant, and it has since committed north of $20 billion to build that corridor out into an AI campus fed straight off the reactor. Amazon also put roughly half a billion into X-energy and pledged to deploy 5 GW of its Xe-100 reactors by 2039. Google went a different route and signed with Kairos Power for 500 MW, structured as the first corporate SMR fleet deal in the country, with the first unit hoped for around 2030. Meta, late as usual to the party it will eventually dominate, stitched together deals with Oklo and TerraPower on the reactor side and Vistra on the existing-plant side, adding up to as much as 6.6 GW, including funding two TerraPower units worth about 690 MW.
Read those deals carefully, and a pattern falls out. The gigawatts that are real and near-term are almost all restarts and power-purchase agreements off existing or already-built plants. Three Mile Island is a restart. Susquehanna is an existing plant. Vistra is selling capacity it already has. The brand-new small modular reactors, the Xe-100s and the Kairos Hermes units and Oklo’s Aurora, are the parts dated 2030, 2035, 2039. The AI-nuclear story is being sold in the present tense and delivered in the future tense, and the gap between those two tenses is where I lose patience with most of the coverage.
The one moment that actually moved the needle for me was X-energy’s IPO. The company priced at $23 a share on April 23, 2026, four dollars above its own marketing range, raised about $1.02 billion, and landed a valuation near $9.12 billion, the largest pure-play advanced-nuclear listing in US history. Demand ran past $10 billion at the top of the range. That is not a vaporware valuation. Its Xe-100 is a high-temperature gas pebble-bed design at 80 MWe per unit, and the Amazon 5 GW commitment underwrites the order book in a way NuScale never had. When sovereign wealth funds and hyperscaler treasury desks are fighting over an SMR book that hard, the technology has at least crossed from PowerPoint into something investable.
But I cannot write about the American SMR dream without the cautionary tale that sits underneath it, because everyone in the industry pretends it did not happen. NuScale is the most advanced US design, the only one with a Standard Design Approval, and its scaled-up US 460 unit cleared the NRC in May 2025, two months ahead of schedule. NuScale is also the company whose flagship project collapsed. The Carbon Free Power Project with the Utah municipal utilities died in November 2023 because the price kept climbing: $4.2 billion in 2018, $6.1 billion in 2020, $9.3 billion by 2023, even after the plant was scaled down to 462 MW in 2021. The subscribers walked. First-of-a-kind SMRs have no economies of scale to lean on, which is the entire economic premise of the “modular” part, and the first units are always the ones that eat the overrun. I keep that graveyard in mind every time a hyperscaler announces a number with a 2035 date attached to it.
France holds the best hand and keeps folding early
Here is the part that actually annoys me, because it is my home turf. France has the strongest nuclear card on the continent by a distance. Around 70% of its electricity is nuclear; it has the deepest operating fleet and the deepest bench of reactor engineers in the Western world, and it has EDF, a national champion that has built more reactors than anyone outside China and Russia. If any Western country should own the co-located-nuclear-plus-compute future, it is this one. And France is playing the hand slowly, expensively, and with a nervous flinch every time it gets close to committing.
The SMR bet is Nuward, EDF’s small modular design. The problem is that Nuward has already been redesigned once under fire. In July 2024, after feedback from potential European customers, EDF hit pause and “optimized” the design, throwing out some of the more novel engineering in favor of existing and proven technology, and pushed the conceptual design finalization out to mid-2026. The reworked reactor is now a 400 MWe plant, sold as deliverable in 48 months, with a 1,150 MWth core convertible to 400 MWe plus 115 MW of high-temperature heat for industry. EDF is talking about up to 30 SMRs by 2050 and a first fleet of three or four units across one or two European countries. Those are fine numbers. They are also numbers that start in the 2030s, which means France’s SMR is not powering anything, compute or otherwise, this decade.
The bigger French story is the EPR2 program, and this is where my frustration turns into something closer to alarm. EPR2 is the plan to build six new large reactors at Penly, Gravelines, and Bugey. In December 2025, EDF put the provisional cost of that program at €72.8 billion, about $85.4 billion, presented to its own board and headed for a government audit in the first quarter of 2026. Rewind to 2019 and the estimate for those same six units was around €56 billion. The number has moved €17 billion in the wrong direction before a single one is finished. And the reason everyone braces when EDF quotes a new-build price is Flamanville-3, the EPR that came in at €12.7 billion, more than triple its original budget, after well over a decade of delay. EDF’s own learning-curve pitch has Penly starting near €8,000 per kilowatt and dropping toward €5,870 by the time Bugey gets built. That is the optimistic case, and it still lands above where the Americans and the Chinese say they can build.
So France owns the fleet, the expertise, and the only credible Western industrial base for pouring reactors at scale, and it is saddled with an SMR that slipped a year and a flagship large-reactor program whose price chart only points up. This is the same maddening pattern I keep seeing across French deep tech and French tech sovereignty: world-class engineering, real strategic assets, and an execution cadence that lets faster and cheaper players eat the actual market. The fleet is the strongest card the continent holds. Nobody in Paris seems to be in a hurry to play it before someone else defines the terms.
China is the one actually connecting reactors to the grid
While the Americans sign offtake agreements and the French audit their cost estimates, China is doing the boring thing that wins: building. It is running or constructing two separate SMR lines, the HTR-PM high-temperature gas-cooled design and the ACP100, marketed as Linglong One, an integral pressurized-water reactor. Linglong One is a 125 MW unit at Changjiang in Hainan, launched in 2021, and through 2025 it cleared cold functional testing and a non-nuclear steam turbine run, the last major drills before fuel loading and grid connection. The HTR-PM has been feeding the grid since late 2021. These are not renderings. They are reactors with steam running through them.
The cost and speed gap is the part the West does not like to talk about. CNNC built the first Hualong One in 68.7 months and is targeting a simplified Hualong Two that drops construction cost from roughly CNY 17,000 per kilowatt, about $2,600, to around CNY 13,000, about $2,000, while cutting build time from five years to four. Compare that $2,000 to $2,600 per kilowatt against EDF’s €5,870 to €8,000 and the picture is brutal. China standardized on modular construction, factory-prefabricating whole cooling-system and steam-generator assemblies, the same industrial playbook that made its solar and battery dominance inevitable. Taipingling, six reactors in the Greater Bay Area, cost about 120 billion yuan, roughly $17.6 billion, and went up in five years. The country that can build a reactor in four years for a third of the Western price is not going to lose the co-location race on economics. Whether its SMRs can be exported past the countries already inside its political orbit is a different question, and one I am deliberately not opening here because the geopolitics of nuclear export is its own post.
The number nobody says out loud
Now the cost reality, because “nuclear is expensive” is true but useless without figures. Idaho National Lab’s own range puts SMR overnight capital cost at $4,000 per kilowatt on the low end, $6,000 in the middle, and $7,000 on the high end, against roughly $3,800 for a conventional large reactor. That is the uncomfortable truth buried in the modular dream: per kilowatt, small is currently more expensive than large, not less, because you throw away the economies of scale that make big reactors cheap. The levelized cost lands somewhere between $60 and $100 per MWh for SMRs versus about $50 for large nuclear in the same modeling. There are rosier academic models; one Gen IV economic study pegs an SMR at $2,901 per kilowatt against $6,936 for a conventional plant, but those are nth-of-a-kind projections assuming a factory line that does not exist yet. The EIA’s 2025 outlook is blunter: its reference case builds no new US nuclear at all, citing overnight costs around $7,821 per kilowatt. SMRs get cheaper than large reactors only after enough of them get built to create a learning curve, which INL and the IAEA both put somewhere in the 2040s, not now.
Set that against demand, and the mismatch is almost comic. Hyperscale data centers pulled roughly 50 GW globally in 2024 and are tracking past 100 GW by 2030, with one projection putting data-center demand at 106 GW by 2035. Total SMR capacity in development right now sits around 22 GW. Deloitte figures new nuclear could cover about 10% of the projected data-center demand increase by 2035. The SMR market itself is real and growing fast, from about $3.5 billion in 2025 toward $14 billion by 2035 at nearly 15% CAGR, with Asia-Pacific already holding 42% of it. But “fast-growing market” and “can power the AI buildout this decade” are not the same sentence. The reactors that hyperscalers are signing for cannot, on their own, feed the compute they are signing for. This is the data center power crunch meeting physics, and physics is not moving its timeline for anyone’s earnings call.
The nuclear-for-AI story is real as a 2030s bet and mostly theater as a 2025 one. The gigawatts arriving before 2030 are old reactors switched back on, not new ones co-located with racks. The brand-new small reactors that would let a hyperscaler bolt power straight onto a campus are years and one or two brutal first-of-a-kind overruns away from mattering. China is the only player actually connecting new units to a grid on schedule and at a price that works; France has the best industrial hand in the West, and the least urgency to play it, and the US has the capital, the deals, the IPO, and a real chance of building the thing, right after it finishes explaining away the last project that collapsed.
If I had to bet, the first data center actually running on a purpose-built SMR at commercial scale lands closer to 2032 than 2030, and it is at least even odds it happens in China or the Gulf before it happens in Pennsylvania or Normandy. I would love to be wrong about the France part.