Nuclear power for AI comes from old reactors first. The new ones wait on fuel.

Some of the earliest new nuclear power for AI in America is due from a reactor that shut down in 2019.

Constellation is restarting it, the former Three Mile Island Unit 1, for Microsoft, and upgrading eleven more reactors for Google. Holtec and NextEra are restarting two others. Meanwhile the new reactors, from Oklo, TerraPower and Kairos Power, have permits, ground broken and buyers signed.

Yet the old reactors come first, because nuclear power for AI arrives in the order of what each project already holds. An old reactor has a licensed site, a place on the grid and ordinary fuel. Most of the new ones lack a fuel that America has almost stopped making.

Explain it like I don't work in finance

You want a cooker working this month. The old one in the garage still has its socket on the wall and its gas pipe in place, so it needs only a service and a safety check.

The new model is better, but it runs on cartridges only one shop sold, and that shop has stopped. The new cooker is the future. The old one cooks dinner first.

A reactor needs four things before it can power an AI campus: a place on the grid, a working reactor, a licence and fuel. Each runs on its own clock. A restart means repairs and inspections. An uprate, which raises a running reactor’s output, means new turbines fitted during refuelling shutdowns.

A new reactor starts every clock at zero. The newest designs also need HALEU, high-assay low-enriched uranium, concentrated to just under 20 per cent of the kind of atom that splits. Ordinary reactors use up to about 5.

Enrichment means spinning uranium gas through long lines of fast centrifuges. The only American line that has made HALEU at the level these reactors need is a demonstration of sixteen centrifuges in Ohio. Its contracted production finished in mid-June 2026, and the next contracted tonne is due by July 2032. Nobody builds a bigger plant until enough reactors exist to burn its output.

Even the old reactors lean on Russia, which sold American reactor owners about a quarter of their enrichment last year. That gap can be bridged from Europe and from stockpiles. The HALEU gap, for now, cannot.

Crane, the restarted reactor, needs none of it. It is 835 megawatts on ordinary fuel and still targets 2027. The difference is not the reactor design. It is what each project already has.

---

For investors

Each project starts on the rung of what it already holds

Line the projects up by what they already hold, and the megawatts sit at the top of the ladder. Restarts reuse a licensed site, a grid place and ordinary fuel. Uprates add output to reactors already running. A new reactor on an old site keeps ordinary fuel. Three advanced reactors start from nothing and need HALEU.

Seven in every eight megawatts on the ladder run on ordinary fuel

Capacity of each project in megawatts, with the date its owner, vendor or regulator gives

About 3,440 MW of the projects on this ladder run on ordinary fuel, against 470 MW that need HALEU. The running measure through this piece is the date each owner or regulator gives for first power. On the top rungs the grid place and the reactor work set that date. On the bottom rungs the licence class and the fuel do.

For a restart, the hard part is the plug

Crane shows that a closed reactor can lose its place on the grid while keeping its ability to run. Constellation announced its return on 20 September 2024, with Microsoft buying the output for 20 years, at about $1.6 billion of capital spending.

Closing had cost Crane its grid place. Constellation’s waiver request to FERC of 31 March 2026 says PJM’s first study tied Crane to hundreds of miles of new lines. Through a fresh connection, Crane could be fully deliverable no sooner than 31 December 2030. The reactor could generate in the second half of 2027.

Borrowed grid rights pull Crane's date forward by about three years

Crane restart, from its announcement to power on the grid, by route

So it borrowed a plug. FERC agreed on 1 June 2026 that 760 MW of capacity interconnection rights, a plant’s right to deliver a set amount onto PJM’s grid, can move to Crane from Eddystone Units 3 and 4. Those are two oil and gas units now running only under federal emergency orders. Constellation still targets 2027, with NRC approvals expected early that year.

A held plug makes a restart possible, not quick

Palisades kept its plug, and the reactor now sets its date. The 800 MW Michigan plant closed in May 2022 but was suspended rather than retired. In MISO, the Midwest grid operator, that keeps a connection for at most 36 cumulative months in five years, and FERC granted Holtec a one-time extension to 1 March 2027.

The wire was ready first. The switchyard was reconnected on 11 September 2026. Fuel loading, begun on 30 August, stopped when an assembly tilted after the refuelling machine released it. Holtec recovered it by 21 October and gave no new date, although its power contracts can be ended if the plant is not in service by March 2027.

NextEra’s 615 MW Duane Arnold restart in Iowa is due in the first quarter of 2029 at the latest, with its NRC review still open.

Uprates add megawatts one outage at a time

An uprate needs no new plug, but it arrives in steps. On 6 October 2026 Constellation agreed to sell Google 890 MW of new capacity from uprates at eleven units in Illinois, Pennsylvania and New Jersey. The first lands by 2028 and all of it, Google says, before the end of 2032.

Heavy equipment is swapped during refuelling outages, which Constellation’s units take every 18 to 24 months. The NRC’s list of expected applications shows twelve from Constellation, the last due to be filed in late 2030. It withholds plant names, so it is not Google’s schedule. It does show a staircase, not a block.

A new reactor on ordinary fuel keeps pace with the old ones

The order is not old against new; it is what each project holds. Ontario Power Generation is building four 300 MW BWRX-300 reactors at Darlington, beside an existing nuclear station, on ordinary fuel. Canada’s regulator licensed the first on 4 April 2025 and received OPG’s application to operate on 25 March 2026.

The plug is new but short: Hydro One planned a 22 km line inside an existing corridor. GE Vernova spoke in January 2025 of operation by the end of 2029, the same years as the American restarts. OPG’s later schedule is in press reports only, so I make no claim on it.

The advanced reactors have permits and buyers, but not fuel

The three HALEU reactors have cleared the steps that once looked hardest. What none has is a commercial supply of fuel that is being made today.

The NRC authorised TerraPower’s construction permit for Kemmerer Unit 1 in Wyoming, a 345 MW Natrium reactor, on 4 March 2026, and TerraPower expects it complete in 2030. Its first cores are contracted from ASP Isotopes, which plans enrichment at Pelindaba in South Africa. ASP’s quarterly report of 14 August 2026 says it has not yet applied its technology to uranium-235.

Oklo plans to own its 75 MW Aurora reactors and sell the electricity. Aurora-INL in Idaho broke ground on 22 September 2025 under Energy Department authorisation, and Oklo targets 2028. Its first core rests on a 5 tonne award of government HALEU recovered from old reactor fuel. Beyond it, Oklo has only a letter of intent with Centrus for up to five Auroras from 2029.

Kairos Power sold first. On 18 August 2025 TVA agreed to buy up to 50 MW from Hermes 2 at Oak Ridge from 2030, with Google taking the clean energy attributes. But its NRC permits cover two test reactors of 35 thermal megawatts each, which cannot make 50 electrical megawatts. Kairos contracted for government HALEU on 20 January 2026, only for Hermes 1, its demonstration reactor. Hermes 2’s fuel source is still not named.

America has stopped making the fuel, and stock and promises stand in

The strongest evidence comes from the only American maker in its own filing. Centrus’s quarterly report says its contracted HALEU production finished in mid-June 2026. On 30 June the Energy Department exercised a three-month option for maintenance and storage only, with no production.

America's only HALEU line has stopped, and what stands behind it is stock or promises

HALEU for American advanced reactors: dated selections, letters, orders and deliveries, against the reactors' own target dates

Everything behind that line is stock or a promise. Centrus is working with the department on agreements to run the cascade privately, with no assurance they are reached. Its new federal task order asks for one tonne by July 2032. ASP has not yet enriched uranium, and its contract with TerraPower runs from 2028.

Government stock is handed out by allocation: ten conditional selections in three rounds, the latest on 23 July 2026, with no tonnage published for any. A selection is a conditional promise, not fuel. It becomes fuel through a contract and a delivery, and Kairos’s has reached the contract, for Hermes 1. This stock covers a few first cores, such as Oklo’s, not the fleets behind them. Gas is the opposite case, where the new pipe lands alongside the turbine; here the fuel runs years behind the reactor.

What could change the order

The case against has two halves, and both are real. The new reactors could come sooner: Aurora-INL targets 2028 on government stock, Centrus may run its cascade privately, and the department may hand out more. And the old reactors have a fuel gap of their own.

Russia supplied 26% of the enrichment American owners bought in 2025, and every waiver under the ban ends by 1 January 2028. Uranium is thinly contracted further out. At the end of 2025 owners had at most 174 million pounds contracted for 2026 to 2035 and 186 million not yet contracted, about half. For 2031, 23 million pounds was unfilled against at most 10 million contracted.

Neither half reverses the order. Aurora-INL alone, at 75 MW, would not bring HALEU reactors to the 100 MW my claim tests, and the others’ fuel is unnamed or not yet enriched. For the running fleet the gaps are bridgeable. US-origin enrichment was 23% of 2025 purchases, Urenco’s Eunice plant adds about a fifth of the Russian volume by early 2027, and owners held 118 million pounds of uranium against 40.9 million loaded that year. Utilities usually buy uranium a few years forward, and a year earlier the unfilled total was a similar 184 million pounds, so it is a buying pattern, not a shortage.

The old cooker still has its socket and its gas pipe, and the service is booked. The new one is the better machine, but its cartridges come from a shop that has stopped and a few promised shops that have not opened. Nuclear power for AI arrives in the order of what each project already has, and this decade that order starts with the old reactors.

The next American tonne is a fifth of one small reactor’s fuel award

Here is the number that is easy to pass over. Oklo’s first 75 MW Aurora rests on 5 tonnes of government HALEU. The only American HALEU production under contract with a stated size is Centrus’s federal tonne, due by July 2032. That tonne is a fifth of the award one small reactor’s first core rests on.

America's next contracted tonne is a fifth of the award behind one small reactor's first core

HALEU in tonnes: what Centrus made, what it must make next, and the government stock behind Oklo's first reactor

So this decade’s first cores come from stock, not from new enrichment. The stock was recovered from old reactor fuel and is handed out by allocation, with no tonnage published. That makes the Energy Department, not a market, the gate for the new reactors’ early years. The nine holders of its selections gain, and Hermes 2, whose fuel source is unnamed, waits. The loop is plain: nobody builds an enrichment plant before reactors exist to burn its output, and no reactor runs without fuel. Two events would start to break it: Centrus agreeing to run its cascade privately, and ASP Isotopes enriching its first uranium.

Ordered today, a big turbine and the next HALEU tonne both arrive in 2032

The second idea joins this piece to its gas sibling. A heavy gas turbine slot sold now ships in 2032. America’s next contracted tonne of HALEU is due by July 2032. Google’s uprates are due in full before the end of the same year.

The gas queue and the fuel queue run out in the same year

From agreement or order to the date each owner, maker or contract gives

That is why the old reactors’ lead should last the decade, not just open it. On the contracts in place today, a buyer wanting firm power from a big new gas turbine is looking at 2033 at the earliest, and advanced reactors beyond the stock wait on a tonne due in 2032. Until then, existing reactors are among the few firm, carbon-free sources that can grow without queuing for a new reactor or gas turbine. Their owners, Constellation, NextEra and Holtec, hold that window. Buyers who waited pay in years. The quicker gas routes, and Aurora-INL on government stock, are the exceptions worth watching.

The HALEU gap is public in Centrus’s own filing. That the gas queue and the fuel queue land in the same year is my framing, and I have not seen it made elsewhere.

What would prove me wrong

The first test is the bottom rung. If HALEU reactors with 100 MW or more of capacity between them send electricity to American grids by the end of 2030, fuel was not the gate it looks like today. Aurora-INL alone, at 75 MW, would not break it.

The second is the top rung. If none of Crane, Palisades or Duane Arnold has sent power to the grid by the end of 2029, an existing site is not the fast lane I claim.

Research notes

Two initiation notes rest on these claims: Constellation, on the Crane claim, and Oklo, on the Oklo claim. Whether any of this is already in share prices is a question for these notes, not for the journal.

The claims on record, word for word

The test this piece stands behind, quoted word for word from the claims page:

Nuclear for AI: restarts first, new reactors later (30 October 2026)

Claim: “Nuclear power for AI arrives in the order of what each project already has: restarts and uprates of existing American reactors, which reuse a licensed site, a grid place and ordinary fuel, deliver this decade, while reactors fuelled with HALEU deliver little before 2031, because government stock covers only a few first cores and commercial enrichment at the level they need has stopped in America and not started at the plants contracted to replace it.”

Breaks if: “Reactors fuelled with HALEU together send electricity to American grids from 100 MW or more of capacity on or before 31 December 2030; or none of Crane, Palisades or Duane Arnold has delivered power to the grid by 31 December 2029.”

Two earlier claims on these reactors:

Constellation can bring a closed reactor back by 2027. The grid connection was the hard part. (30 October 2026)

Claim: “The fastest nuclear power Constellation can bring to AI this decade comes from reactors that already hold a place on the grid: at Crane the reactor can be ready by 2027, but a fresh grid connection would not have been fully ready until around 2030 or later.”

Breaks if: “Crane, with the 760 MW of grid rights transferred from Eddystone, has still not delivered power to the grid at full output by 31 December 2028, which would mean the reactor work and not the connection set the pace; or restarts and uprates that need new grid capacity are connected as quickly as those that reuse an existing one.”

Where later evidence bears on it: nothing cuts against it, and the test is open. Palisades adds a caution: a held grid place makes a restart possible, not quick. The uprates arrive in outage-sized steps through 2032, not as one block.

Oklo will not sell you a reactor. It wants to sell you nuclear electricity for decades. (29 September 2026)

Claim: “Oklo is a seller of electricity, not reactors, and the model stands or falls on building its first plant.”

Breaks if: “Oklo starts selling reactors outright to raise cash, or Aurora in Idaho slips well past 2028 or produces power uncompetitive with gas, or the letters of intent stay letters.”

Where later evidence bears on it: I was wrong. The piece said Meta’s agreement “includes a binding prepayment to secure fuel”. Oklo’s quarterly report calls it a mechanism for Meta to prepay, discloses no amount, and does not call it a power purchase agreement. Its exposure map also says Centrus “enriches uranium to HALEU levels”. Centrus’s contracted HALEU production finished in mid-June 2026, the Energy Department has since ordered only maintenance, and its next contracted tonne is due by July 2032. Its supply to Oklo is a letter of intent, deliveries from 2029. The first half of the claim holds: Oklo still sells power, not reactors, and still targets 2028. But the limb “the letters of intent stay letters” currently points towards a break: no customer megawatt is under a binding power purchase agreement, and Oklo’s fuel beyond its first core rests on a letter of intent too.

The figures behind the charts

Each project on the ladder, with dates as the owner or regulator gives them:

Rung	Project	Grid place	Reactor work	Licence	Fuel	Date given
Restart, borrowed plug	Crane, 835 MW, Constellation	760 MW of rights moved from Eddystone (FERC, 1 June 2026)	Can generate in the second half of 2027	Existing; NRC approvals expected early 2027	Ordinary	2027 (6 August 2026)
Restart, held plug	Palisades, 800 MW, Holtec	Held to 1 March 2027 (FERC, 28 February 2025)	Fuel loading began 30 August; halted, assembly recovered by 21 October 2026	Existing, restored	Ordinary	None (21 October 2026)
Restart	Duane Arnold, 615 MW, NextEra	Not stated by NextEra	Inspections, engineering	NRC review open	Ordinary	First quarter 2029 at the latest (8 September 2026)
Uprate	890 MW at eleven Constellation units, for Google	Existing	New turbines, fitted in refuelling outages	NRC amendment where heat rises	Ordinary	2028 to 2032 (6 October 2026)
Small reactor, old site	Darlington 1, 300 MW, OPG	New 22 km line, existing corridor	Licensed to build, 4 April 2025	Licence to operate applied for, 25 March 2026	Ordinary	Vendor: end of 2029 (27 January 2025)
Advanced, permit in hand	Kemmerer 1, 345 MW, TerraPower	Not a stated gate	Construction from April 2026	Construction permit, 4 March 2026	HALEU, supplier not yet enriching	2030 (4 March 2026)
Advanced, own and operate	Aurora-INL, 75 MW, Oklo	Not disclosed	Ground broken 22 September 2025	Energy Department, two of five steps	HALEU, 5 tonnes of government stock	2028 (7 August 2026)
Advanced, sold first	Hermes 2, up to 50 MW, Kairos Power	TVA’s grid	Removing legacy items before excavation	Permits for two 35 MWt test reactors	HALEU, source not named; government stock contracted for Hermes 1 only	2030 (18 August 2025)

The 3,440 MW on ordinary fuel includes Darlington’s 300 MW in Canada, outside the claim’s test; without it the figure is 3,140 MW. The 470 MW on HALEU counts Hermes 2 at its full 50 MW. Both are my arithmetic.

More on the rungs. Palisades: Holtec’s power contracts can be ended by either side if the plant is not in service by March 2027 (registration statement, 8 September 2026); the tilted assembly is in an NRC filing of 30 September 2026; MISO’s grid agreement lists reactor startup testing by 1 February 2027. Uprates: the NRC list of 8 September 2026 shows seven small and five extended uprates from Constellation, and covers all its uprates, not only Google’s. Darlington: the regulator lifted its hold point for the reactor building’s foundation on 30 March 2026. Aurora-INL: two of five Energy Department steps approved by 11 June 2026; its 5 tonnes are HALEU recovered from old EBR-II fuel; the Centrus letter of intent of 18 June 2026 is subject to definitive agreements (Centrus, 6 August 2026). Meta’s agreement with Oklo of 5 January 2026, for a 1.2 GW campus in Ohio, gives Meta a mechanism to prepay for power, which Oklo will use to secure fuel. Kemmerer: the Energy Department has conditionally promised TerraPower HALEU from government stock in an undisclosed amount (9 April 2025), and ASP has not won approval to test its technology on uranium beyond a services contract with Necsa, South Africa’s state nuclear company. Hermes 2: the permits of 21 November 2024 were issued under the Atomic Energy Act’s research section, which caps energy sales at half the plant’s yearly running costs (42 U.S.C. 2134(c)); the NRC has not ruled on a larger reactor and no operating licence application is on its record. Hermes 1’s latest completion date moved to 30 April 2029 at Kairos’s request (NRC, 13 May 2026). Kairos was a first-round Energy Department HALEU selection (9 April 2025). On 1 October 2026 Kairos was still removing legacy items to prepare for site excavation.

Enrichment, for old reactors and new:

Supplier	Product	Capacity or output	Status, dated
Russia	Ordinary	3.28 million SWU, 26% of US owners’ 2025 purchases	Waivers end by 1 January 2028 (EIA, August 2026)
Urenco, Eunice	Ordinary	About 4.3 million SWU; 700,000 more by early 2027	Licensed for 10 million since 2015; new plant producing 2032 (29 June 2026)
Orano, France	Ordinary	2.5 million SWU extension	First production 2028 (10 October 2024)
Orano, Oak Ridge	Ordinary	Not stated	NRC review due by 30 April 2027
Centrus, Piketon	Ordinary	New centrifuges	First online 2029 (6 August 2026)
Centrus, demonstration cascade	HALEU	16 centrifuges; 900 kg in phase two alone, to June 2025	Contracted production finished mid-June 2026; maintenance only since 30 June
Energy Department stock	HALEU	Ten conditional selections in three rounds; no tonnage published	Latest round 23 July 2026; Kairos’s contract, for Hermes 1, 20 January 2026
Centrus, federal task order	HALEU	1 tonne by July 2032	Signed 1 July 2026
General Matter	HALEU	Federal task order	Awarded 5 January 2026
ASP Isotopes, Pelindaba	HALEU	Up to 150 tonnes to TerraPower, 2028 to 2037	Has not enriched uranium (14 August 2026)

SWU, separative work units, measure enrichment work. The Energy Department’s three rounds: 9 April 2025 (TRISO-X, Kairos Power, Radiant, Westinghouse, TerraPower), 26 August 2025 (Antares, Standard Nuclear, Abilene Christian University with Natura) and 23 July 2026 (NASA, Radiant). Radiant was selected twice, so nine recipients.

Uranium, from the EIA’s 2025 Uranium Marketing Annual Report (released 29 July 2026, re-released 13 August 2026), all as U3O8 equivalent. Contracted deliveries for 2026 to 2035 were at most 174 million pounds (Table 10, maximum; minimum 132.5 million). Unfilled requirements to 2035 were 186 million (Table 11), about half the 360 million maximum anticipated requirement. For 2026, under 1 million pounds is unfilled. For 2030, more is contracted (at most 17.9 million) than unfilled (12.0 million). For 2031, 23.1 million is unfilled against at most 10.1 million contracted, so most of the uranium the fleet will burn in 2031 to 2035 is not yet bought. At the end of 2024 the unfilled total for 2025 to 2034 was 184 million. The EIA’s summary says “2025 through 2035”, but its end-2025 table starts in 2026.

Inventories at the end of 2025 (Table 22, preliminary): owners held 118.3 million pounds, including 51.8 million as enriched uranium and 7.8 million in fabricated fuel not yet loaded, against 40.9 million loaded in 2025; suppliers held 51.7 million more.

Chart dates: Crane’s two routes start at its announcement of 20 September 2024. In the HALEU chart, years given without a month are placed mid-year, except Centrus’s letter to Oklo and the start of ASP’s window, placed at the start of 2029 and 2028; ASP’s line is cut at the chart’s edge.

Who is exposed if old reactors come first and new ones wait on fuel

This maps who operates in each layer. It is not a recommendation, and naming a company is not a view on its shares. Figures in the piece are disclosures used as evidence for the mechanism.

Plant owners

Constellation, NextEra Energy and Holtec run, restart and upgrade existing reactors.

Reactor makers

Oklo, TerraPower and Kairos Power design advanced reactors; GE Vernova Hitachi Nuclear Energy designs the BWRX-300.

Enrichers

Centrus Energy, Urenco, Orano and ASP Isotopes enrich or plan to enrich uranium.

The buyers

Microsoft, Google and Meta run AI data centres on long nuclear contracts.

On the other side

AI campuses counting on HALEU reactors this decade, and Rosatom, which loses its American customers.

---

Sources: Constellation releases (20 September and 4 November 2024; 6 August and 6 October 2026), its 2025 annual report and FERC waiver request (31 March 2026); FERC orders (28 February 2025, 1 June 2026); MISO's Palisades grid agreement; Holtec's registration statement (8 September 2026), NRC filing (30 September 2026) and releases (21 September, 21 October 2026); NextEra (8 September 2026); Google (6 October 2026); the NRC's uprate list, Hermes permits and orders, Kemmerer release, Eunice licence and Orano letter; 42 U.S.C. 2134(c); the Canadian Nuclear Safety Commission, Hydro One (27 March 2023) and GE Vernova (27 January 2025); Oklo (7 August 2026); TerraPower (9 January, 4 March 2026); Kairos Power, TVA and Google (18 August 2025) and Kairos (20 January, 1 October 2026); Centrus (6 August 2026) and ASP Isotopes (14 August 2026) quarterly reports; Energy Department HALEU pages to 23 July 2026; Public Law 118-62; the EIA's 2025 Uranium Marketing Annual Report, tables 10, 11 and 22; Urenco (29 June 2026) and Orano (10 October 2024). Leads came from Leo Cui's YouTube videos on small reactors and the energy sector. Personal research, not investment advice.