The Nuclear Bid

Global nuclear output set a record in 2025.¹ So, by our count, did the gap between what the industry announced and what it connected to a grid. Both facts belong in the first paragraph of any honest read of this market, because the second one is where the money hides.

The new fact of the past two years is the buyer. Microsoft, Amazon, Google and Meta have signed twenty-year commitments against reactors that are switched off, reactors that exist only as licence applications and in one case a technology that has never produced a commercial watt. Power is the binding constraint on the AI buildout: compute is capital and capital is available, whilst power is physical, permitted and queued, and the queue is measured in years. That makes firm clean generation the scarcest input in the largest capex cycle on earth, and it has put a standing bid under the world's slowest-moving energy asset.

Our starting position, held before this piece and unmoved by writing it: the technology is not the constraint and never was. Fission has worked for seventy years. What does not work is building it on time and on budget in a Western regulatory environment, which makes nuclear a political and financial story wearing a technology costume. Technology signals are therefore weak evidence here, and announcement signals are weaker still. The strong evidence sits in three places: the terms of the offtake (who carries the cost overrun tells you who actually believes the schedule), the physical fuel cycle (where balances clear regardless of narrative) and the small set of programmes anywhere on earth that deliver reactors on contract.

This piece reads the market in that order. A short history of how the industry got here, then the bid, then the conversion problem, then the commodity underneath, then the builders, then where we think it settles and who gets paid.

Two acts and a long stall

The first act was fast. Obninsk put five megawatts on the Soviet grid in 1954; Calder Hall followed in 1956; by the early 1970s the United States had over two hundred reactors ordered or under construction. France, panicked by the 1973 oil shock, launched the Messmer plan in 1974 and built out fifty-six reactors in roughly fifteen years, the fastest decarbonisation of a large grid ever executed and still the standing proof of what serial, standardised state delivery does to nuclear economics.²

The stall came in three shocks. Three Mile Island in 1979 killed the American order book without killing anyone; Chernobyl in 1986 froze Europe; Fukushima in 2011 took the entire Japanese fleet offline and pushed Germany into a full exit, completed in April 2023. Nuclear's share of global electricity peaked at 17.5 per cent in 1996 and has drifted to roughly 9 per cent since, not because reactors closed en masse but because the world electrified around a fleet that stopped growing.³

The second act began quietly in the East. China connected its first reactor in 1991 and now runs the programme everyone else measures themselves against, approving roughly ten units a year since 2022 and building them in six to seven years at costs the West last saw in the 1970s. The stall, in other words, was never global. It was a Western delivery failure wearing a global reputation.

What actually broke

The Western numbers ended the first act, and any position in this market starts by looking at them without flinching.

Vogtle units 3 and 4, the only reactors completed in the United States in three decades, were sanctioned at around $14bn and finished around $36.8bn, fifteen years in the building.⁴ Flamanville 3 reached the French grid in December 2024, twelve years late and roughly four times its budget. Olkiluoto 3 in Finland ran fourteen years behind schedule. Hinkley Point C, sold to the British public at £18bn with first power promised for 2017, was re-costed by EDF in February 2026 at £35bn in 2015 prices (roughly £46bn in today's money) with unit 1 now confirmed for 2030, and Sizewell C reached its final investment decision in July 2025 at around £38bn, close to double the original estimate.⁵

Now the control group. Barakah in the UAE: four Korean APR-1400 reactors, first concrete 2012, all four in commercial operation by late 2024, delivered materially on the terms KEPCO contracted in 2009.⁶ China builds the same class of machine in six to seven years, in parallel, dozens at a time. Same physics, same steel, same neutrons. The difference is delivery: regulatory churn, first-of-a-kind engineering re-litigated on every site, atrophied supply chains and, above all, financing cost compounding schedule risk. At a Western cost of capital the clock is the budget; every year of slip capitalises interest onto a balance sheet that was already the most expensive part of the machine.

And because the honest version of this scene needs its own counter-example: the failure is specific to new-build, not to nuclear engineering in the West generally. Ontario Power Generation finished its four-unit Darlington refurbishment in February 2026 ahead of schedule and C$150m under budget, on the same site where the West's first SMR now rises.³⁵ Repeat work on a known machine delivers; first-of-a-kind is where the disease lives.

This is the base a new demand shock arrives into: a technology with a working Eastern production line, a broken Western one and a customer base, until recently, of regulated utilities that could neither pay for the risk nor wait out the schedule.

The buyers arrive

What changed is not sentiment. It is the balance sheet on the other side of the table.

Between September 2024 and January 2026 the four largest buyers of electricity on earth committed to nuclear offtake at tenors utilities stopped writing decades ago. Microsoft signed twenty years against the restart of Three Mile Island unit 1, now the Crane Clean Energy Center, 835 MW against a $1.6bn revamp.⁷ Amazon amended its Talen arrangement at Susquehanna up to 1,920 MW.⁸ Meta took 1.1 GW at Clinton for twenty years from mid-2027.⁹ Google signed a 25-year deal to restart Duane Arnold in Iowa, 615 MW targeted for early 2029, and separately a 500 MW master agreement with Kairos Power's molten-salt SMRs.¹⁰

Then the frontier moved from restarts to paper reactors. Amazon set a 5 GW ambition with X-energy. Meta's January 2026 request for proposals landed at up to 6.6 GW across TerraPower (two firm Natrium units, six optioned), Oklo (1.2 GW in Ohio) and Vistra's existing Perry and Davis-Besse plants.¹¹ Google bought 200 MW of fusion offtake from Commonwealth Fusion Systems; Microsoft holds a 2028 delivery contract with Helion, the boldest date in the energy business. One caveat on the headline arithmetic, and one exclusion. Roughly a third of Meta's 6.6 GW preserves existing output rather than adding any (Vistra's 2,176 MW runs today), which is valuable to Meta and invisible to the supply balance. And we exclude the sector's largest number entirely: Oklo and Switch's 12 GW arrangement is explicitly non-binding, and a non-binding gigawatt is a press release with units attached.

Why this matters more than any policy signal of the past thirty years: the historical missing ingredient was a counterparty that could pay above wholesale for firm clean power and could wait a decade for it. Hyperscalers can do both, thereby absorbing exactly the two risks (price and schedule) that regulated utilities could not. The desk's discipline is to weight the structure over the headline: a twenty-year PPA on an existing reactor is a different asset class from a framework ambition on an unlicensed design, and the table above deliberately mixes them because the market's own announcements do.

Announcement is not megawatt

Here is the base rate the industry would prefer you not to anchor on. The 2000s "nuclear renaissance" produced roughly thirty new-reactor licence applications in the United States and, twenty years later, two operating units, both at Vogtle. Nuclear announcements convert to delivered megawatts at a rate close to zero, and that base rate applies to every new announcement by default until its offtake terms say otherwise.

The three freshest milestones in the file all cleared in 2026, all real, and none of them is a watt. FERC granted Constellation a waiver transferring 760 MW of interconnection rights to the Crane restart on 1 June 2026: a genuine regulatory obstacle removed, whilst the reactor stays exactly as switched off as it was, targeted for the second half of 2027.¹² Ontario Power Generation set the 953-tonne basemat for Darlington unit 1 on 1 May 2026, the first Western grid-scale SMR into main construction, targeted at the grid for end-2030.¹³ Commonwealth Fusion became the first fusion company to file for PJM interconnection on 28 April 2026, for a 400 MW plant whose physics milestone has not yet occurred; the interconnection study alone runs four to six years.¹⁴

Scaffolding, all of it. Necessary, checkable and not the thing itself. The same test embarrasses the financing headlines: the World Bank lifted its decades-old nuclear financing ban in June 2025 to considerable fanfare, and thirteen months later had financed nothing at all, its furthest-advanced activity being exploratory talks with South Africa's Eskom.³¹

Precision matters here, so state the live exception before the rule. Deals riding reactors that already existed do operate: Amazon has drawn Susquehanna power since 2024 and Meta's Clinton term starts mid-2027 on a running unit. Those are re-contractings of existing supply. The rule is about additions: not one megawatt of capacity restarted or newly built for this demand is in commercial operation anywhere on earth, so the funnel below, which is what the equity market is actually pricing, remains entirely a claim on the future.

The fastest megawatt is one already built

If the bid is real, the first place it clears is not a new design. It is the small inventory of intact, recently retired reactors, because a restart carries no first-of-a-kind engineering and no unproven licensing pathway, only refurbishment and paperwork. The market agrees: all three American restart projects sold their output forward the moment they were credible.

The scoreboard as of mid-August 2026, and it is more honest than the press releases. Palisades in Michigan (800 MW, Holtec) was to be America's first restart; a February 2026 target has already slipped, fuel is on site and not yet loaded, more than five thousand work activities remain open and the hard contractual obligations begin in March 2027.¹⁵ Crane (835 MW) holds its second-half-2027 target with the interconnection waiver banked and management hinting at earlier. Duane Arnold (615 MW) targets the first quarter of 2029 against Google's twenty-five-year commitment. Three projects, roughly 2.3 GW, spread across 2026 to 2029: none has produced a watt, all three are the cheapest risk-adjusted megawatts in the complex, and the sequencing matters, because if Palisades, the easiest project in the set, slips past its contract date, everything behind it reprices.

Japan is running the same trade at fleet scale. Fifteen reactors were operating as of February 2026, Kashiwazaki-Kariwa unit 6 the latest and the largest single capacity restoration since Fukushima, and every returned unit displaces imported LNG at the margin: the quiet macro link between this market and the gas curve.¹⁶

The commodity remembered first

Physical markets price expectations earlier and more honestly than press releases, and uranium began repricing before a single hyperscaler deal existed.

The shape of the decade first. After Fukushima the market drowned in inventory, and spot touched $18/lb in late 2016, a price at which almost no mine on earth earned its keep. Supply discipline followed (Cameco suspended McArthur River in 2018; Kazatomprom held volumes back), then financialisation arrived in July 2021 when the Sprott Physical Uranium Trust began sweeping surplus pounds into a listed vehicle; it holds 81.7 Mlb today.¹⁷ The 2022 energy shock reset the political premium and spot ran to roughly $106/lb in late January 2024, the highest since 2007. A second spike came in January 2026, around 24 per cent in a month to prints above $100 on some reporters' marks (TradeTech marked $100.25 on 28 January; UxC's month-end was $94.28: quote the reporter or do not quote the number), driven in part by a Washington proclamation discussed below.¹⁸

Now the tell. Since February 2026 spot has gone flat, sitting at $86.38/lb at end-July, whilst the long-term price has risen every single month of the year to $95.50/lb, an 18-year high.¹⁹ A term market making highs above a becalmed spot market is a decoupling, and it says something specific: utilities are paying up on price whilst still refusing to commit on volume. The contracting book proves it. Around 37 Mlb had been term-contracted globally by 10 August, on track for a fourteenth consecutive year below the roughly 150 Mlb replacement rate, and over the past five years about 589 Mlb were contracted against roughly 815 Mlb consumed in reactors.²⁰ US utilities' coverage falls to 60 per cent of requirements by 2030 and 9 per cent by 2033. Every year of under-contracting defers rather than destroys demand, and deferred inelastic demand is the most reliable fuel a commodity rally has. The 2025 numbers show what the eventual repricing looks like: American utilities' term deliveries averaged $55.91/lb against a market term price of $95.50.²¹

Where the pounds come from

The supply side is more concentrated than oil ever was. Kazakhstan alone mines around 40 per cent of the world's uranium, which makes Kazatomprom the swing producer, and the company's position deserves precise reading because it is widely garbled. The August 2025 "10 per cent cut" reduced a regulatory production ceiling, not output: actual 2026 guidance sits at 27,500 to 29,000 tU, below even the reduced ceiling, whilst first-half production ran 9 per cent above a weak 2025 base and the company warns of depletion-driven declines after 2027.²² Cameco runs the Western counterweight from Saskatchewan, holding 2026 guidance at 19.5 to 21.5 Mlb through a year in which a sulphuric-acid plant failure idled the McClean Lake mill and a partially collapsed bridge halted Key Lake: a reminder of how physically fragile a concentrated supply chain is.²³ Niger, historically a mid-single-digit share of global supply and a material share of French feed, has left the Western market entirely: SOMAIR was nationalised in June 2025 and an ICSID tribunal has since ordered Niger not to sell the uranium it seized.²⁴

New supply is real but late, which is the recurring sentence of this piece, and 2026 supplied the two data points that price it. NexGen's Rook I, the largest development project in the industry, cleared its final federal approval on 5 March 2026 and still will not produce before the early 2030s.²⁵ And Deep Yellow declined to sanction its Tumas greenfield even with term at $95/lb, which locates the industry's true incentive price somewhere above the highest term print in eighteen years. The restart cohort (Paladin's Langer Heinrich beating guidance, Boss's Honeymoon, the new American ISR operations, where enCore's cost of sales ran $68/lb against an $86 spot) adds single-digit millions of pounds at thin margins against a structural gap in the tens of millions. Nothing sanctioned today answers demand before the early 2030s, thus the balance of this decade is already largely written on the supply side; the argument is only about demand.

The chokepoint is not the ore

The popular version of this trade stops at uranium. The desk version does not, because pounds in the ground are the least constrained stage of the chain. Between mine and reactor sit conversion (U3O8 to UF6 gas) and enrichment (spinning that gas to reactor grade), and both are duopoly-to-oligopoly industries where Russia holds the largest single position.

Rosatom controls roughly 44 per cent of global enrichment capacity, and Russian material still supplied 26 per cent of American utilities' foreign-origin enrichment in 2025.²⁶ The United States answered with the Prohibiting Russian Uranium Imports Act in May 2024: waivers keep Russian material flowing under shrinking caps through 2027, and the hard stop arrives on 1 January 2028, now under eighteen months away.²⁷

The prices say the squeeze is already on, and they say it with an elegance the ore market cannot match. At end-July 2026 conversion was reported around $64.50/kgU spot against $55.50 contracted, and enrichment around $200/SWU spot against $173 contracted, alongside uranium at $86.38 spot against $95.50 term.²⁸ Read the curve shapes together: uranium carries an 11 per cent term premium (buyers securing future volume), whilst conversion and enrichment both trade at roughly 16 per cent spot premiums (buyers paying for material now). Opposite structures across three stages of one fuel chain, and the midstream is the one in present-tense scarcity. Relief is dated and late: Urenco's New Mexico expansion produces first SWU around 2032, Orano's Georges Besse II extension phases in from 2028, and Urenco's order book sits at a record EUR 21.3bn stretching into the 2040s.²⁹ The 2028 cliff is not speculative demand; it is contractual demand from utilities that must re-book Western separative work that does not yet exist at scale. For calibration, American utilities' 2025 contract book averaged $108.70/SWU against a $173 new-contract market: the repricing runway, again.²¹

Then there is HALEU, the higher-enriched fuel most advanced designs require. Centrus is the only Western producer and has been running at 900 kg a year, demonstration scale; in July 2026 it signed a $900m Department of Energy task order to commercialise the cascade at an initial 12 tonnes a year, a thirteenfold scale-up that conveys exactly how small the base is.³⁰ The DOE's first allocation round went to TRISO-X, TerraPower, Kairos, Radiant and Westinghouse, several of them the same developers the hyperscalers just contracted. A Natrium or an Xe-100 delivered on time into a fuel queue is still a reactor without fuel; the fuel cycle, not the reactor, is the schedule.

The factory thesis, finally on trial

The small modular reactor argument is an economics claim, not an engineering one: replace bespoke megaprojects with a factory series, and the cost curve that killed Western nuclear inverts, because unit two is cheaper than unit one and unit ten is cheaper again. It has never been demonstrated. What is genuinely new in 2026 is that it is finally being tested with steel rather than slideware.

Darlington unit 1 is the test case: the first Western grid-scale SMR in main construction, basemat set in May 2026, first grid connection targeted for end-2030, inside OPG's sanctioned C$20.9bn four-unit programme.²² The only number that will ultimately matter from this programme is the cost of unit 2 against unit 1; a foundation says nothing about it, and we will not pretend otherwise. TerraPower's Natrium took the first commercial Gen-IV construction permit from the NRC on 4 March 2026 and began nuclear-island construction in April, with Meta already holding forward offtake across eight planned units.²³ Kairos broke ground on Hermes 2 in April 2026 with Google's 500 MW behind it; X-energy advances at Dow's Seadrift site with Amazon's ambition behind it; Oklo carries Meta's 1.2 GW selection and Switch's framework.

The cautionary entry is recent and specific: NuScale, the first design certified in America, lost its launch customer in November 2023 when the projected price reached $89/MWh before construction began. That is what happens to paper economics on contact with procurement, and it happened eighteen months before the current wave of enthusiasm. The story turned in February 2026, when Romania's Doicesti project reached a qualified final investment decision on six NuScale modules, one to be built first as a proof, targeting 2033.³⁶ Both halves of that sentence belong in the record.

Two honest readings coexist. The sceptical one: "factory-built" remains a projection, since no SMR line anywhere produces at rate, and first units are precisely where nuclear overruns live. The genuinely new one: the regulator, which this desk regards as the true bottleneck, has started moving at unfamiliar speed. A Gen-IV construction permit in roughly two years, a Canadian licensing track that kept pace with construction, the NRC's risk-informed Part 53 framework finalised in March 2026 after six years of rulemaking and NexGen's final federal approval landing fourteen business days after its hearing closed: four data points in one year on the axis that has produced none for decades.³²

The long option

Fusion belongs in this piece for one reason: the same buyers writing nuclear offtake are writing fusion offtake, and the desk needs a consistent way to price the difference.

The private industry raised $4.48bn in the twelve months to July 2026, taking cumulative funding to $14.24bn across 56 companies.²⁴ Commonwealth Fusion has SPARC's first plasma targeted for 2026, with reporting split on whether it has slipped into 2027 (the magnet assembly cadence, one of eighteen toroidal-field coils delivered by January at roughly fortnightly intervals, is consistent with the slip); a reported $1bn raise landed in July, and the company has already filed for grid interconnection in Virginia for its ARC plant, early-2030s delivery, with Google holding 200 MW of offtake. Helion holds Microsoft's 2028 delivery contract, the most aggressive date in the sector, of which Helion's own chief executive says it is "an aggressive milestone" that is "going to be hard", whilst the physicist whose field-reversed-configuration research underpins the approach argues, from outside the company, that confinement remains the unsolved problem.³⁷ ITER, the public flagship, rebaselined to 2034 for first plasma and 2039 for deuterium-tritium operations.

The discipline: no private company has demonstrated net energy gain; no fusion electricity exists on any grid; and the headline "gain" figures the sector quotes typically exclude the wall-plug energy driving the shot, so Q-plasma and Q-total differ by more than an order of magnitude. An interconnection application is regulatory groundwork, not evidence a reactor works. Price fusion as venture optionality on a different decade's market, thereby keeping it out of every supply and demand balance in this piece, which is exactly where it currently belongs.

Who is actually building

Strip the announcements and count construction, because construction is where national programmes stop being press releases.

China operates 64 reactors and is building 37 more, cleared eight further units in one State Council sitting three weeks before this piece published, and its new five-year plan targets 110 GWe by 2030.²⁵ Even that machine has a stretch target: completing every unit now under construction lands around 103 GWe, so 110 assumes the build keeps accelerating. Its Hualong One is the only Gen-III design in genuine serial production. Russia builds modestly at home and sells everywhere: Rosatom claims involvement in nearly forty projects abroad and a $206bn order book (its own figures, quoted here as claims, not audits), bundled with fuel and state finance across Turkey, Egypt, Bangladesh, India and Hungary, and its four projected foreign start-ups for 2026 have, as of publication, confirmed none. Korea is the West's outsourcing option, with Barakah delivered and the Czech Dukovany contract, roughly $18.6bn for two units targeting 2036, signed in June 2025 once the court challenge cleared. France finally connected Flamanville and now decides, with a final investment decision due before end-2026, whether the €72.8bn six-unit EPR2 programme (first power at Penly 2038) escapes the Hinkley pathology; at roughly €12bn a unit forecast against Hinkley's ~£23bn actual, the sanction price already assumes it does.³⁸ The United States leads on demand and restarts whilst its new-build record remains two reactors in thirty years; the May 2025 executive orders and statutory NRC decision deadlines are the most aggressive licensing reform attempted, and unproven, with the revival of V.C. Summer's abandoned units now a Brookfield-managed option on a 2028 investment decision rather than a build. Poland's first AP1000s at Lubiatowo-Kopalino now target 2036 commercial operation on a ~€42bn programme, which with the Czech build, the Dutch and Swedish programmes and Belgium's life-extension reversal makes Europe's eastern half the continent's real build frontier.³⁹ India targets 100 GW by 2047 from roughly 8 GW today and is legislating private entry, the SHANTI bill through the lower house at time of writing. Japan restarts. Canada builds the West's first SMR. Britain pays the bills on two megaprojects and calls it a programme.

The uncomfortable symmetry of that chart: the hyperscaler bid is concentrated in exactly the jurisdictions with the worst recent delivery records, because that is where the hyperscalers are. The world's most reliable builders are, for now, not for hire in the markets that suddenly need them most.

Where this settles

Our base case, stated so it can be marked to market later.

Demand is the easy half. The operating fleet plus units under construction commit the fuel cycle regardless of any AI narrative: the World Nuclear Association's 2025 reference case takes capacity from 398 GWe today to 746 GWe by 2040, with reactor requirements rising from roughly 69,000 tU (about 179 Mlb) towards 150,000 tU, and that demand is inelastic to price in the way only reactor fuel is (fuel is a rounding error against the capital cost of the machine it feeds).³³ The COP28 tripling pledge, at 38 signatories since China joined in March 2026, and the WNA's upper case will not survive contact with delivery; they do not need to. The decade's incremental terawatt-hours come from China, India, Korea, Russia, Japanese restarts and American life extensions, with Western large new-build arriving late in the 2030s and SMRs resolving their serial-economics question at Darlington units 2 to 4 around 2030-33. Washington has meanwhile put an instrument on the table this market has never traded against: the January 2026 critical-minerals proclamation, uranium included, directed officials to consider price floors, which is a sentence worth reading twice from a government that was until recently the market's largest source of secondary supply.³⁴

Supply is the asymmetry. Mine output responds on a ten-year lag (Rook I: approved March 2026, pounds early 2030s), the midstream on a longer one (first new Western SWU 2028-2032), and the January 2028 Russian-enrichment cliff is contractual rather than speculative. A market where demand is committed and supply is discretionary clears upward on any disappointment, and this one carries a free option: the piece's entire hyperscaler funnel sits ON TOP of a fuel-cycle balance that was already tightening without it. The trade never needed the AI story. The AI story is the convexity.

Who gets paid, in declining order of certainty. First, incumbent low-cost producers with term books, because they sell committed demand at prices set by the marginal delayed mine, and their accounts still carry the runway: Kazatomprom realised $67.88/lb in the first half against an $86 spot and $95.50 term. Second, Western enrichment and conversion, where the 2028 cliff forces re-booking into capacity that expands only from 2028-2032 and whose order books (Urenco's record EUR 21.3bn) already stretch into the 2040s. Third, owners of restartable and uprateable reactors, the Constellation model, who monetise the bid without construction risk. Fourth, and only then, SMR developers, which remain venture-binary until a unit-2 cost prints against a unit-1. Fifth, the engineering, forging and fuel-fabrication chain, which is paid on volume under every scenario that involves anyone building anything.

What would change our mind

A living document ends with its falsifiers, and this desk keeps a kill list for every standing view.

An SMR delivered on schedule and near budget would move us, and the real signal is serial: unit 2 at Darlington costing meaningfully less than unit 1, because that is the entire economic thesis of the technology and it has never been demonstrated. A regulator approving a design in materially compressed time would move us more; the Natrium construction permit and the statutory NRC deadlines are the first genuine evidence on that axis in decades, and we mark them as such. A hyperscaler offtake reaching commercial operation (Crane in the second half of 2027 is the first test the calendar allows) converts the funnel's top into its bottom for the first time. Fusion electricity on a grid, as opposed to a gain figure in a shot, re-prices the entire long end.

Against those: schedule slip at Palisades past its contracted obligations, a cost overrun landing on a hyperscaler that then walks, or a Kazakh supply response faster than guidance suggests would each cut the thesis where it stands. The next dated catalyst is immediate: Kazatomprom reports first-half results on 21 August, the day after this piece publishes, with any 2027 guidance the first hard read on whether the depletion warning binds.

Nuclear spent forty years as the technology that was always about to matter. What changed is not the reactor; it is the arrival of a buyer whose demand does not negotiate and whose balance sheet does not blink. Whether that is enough to fix a delivery machine that has failed for two generations is the live question of this market, and the honest answer is that nobody knows yet, which is precisely why the fuel cycle, where commitment is already contractual, is where we would rather be positioned than in anybody's announcement. In sum: respect the bid, discount the paper, own the chokepoints.