The US Nuclear Energy Value Chain: Where the Durable Profits Concentrate

The demand case is not in dispute here. US reactors ran at roughly a 91% fleet capacity factor in 2025 — they produced 91% of what they would have made running flat out every hour of the year — and supplied about 17% of the country's electricity from a fleet whose average unit is 44 years old; the EIA counted 96 operating reactors and 98,441 MW of net summer capacity in March 2026. EPRI's Powering Intelligence 2026 scenarios put data centres at 9% to 17% of national consumption by 2030 against roughly 4–5% in 2024, or 383–793 TWh, using low/medium/high project-realisation assumptions. Firm, carbon-free, already-interconnected megawatts are scarce. That much the consensus has right.

What the consensus has wrong is where the scarcity converts into profit. Scarcity of a product is not the same thing as scarcity of a producer. Uranium sits at an eighteen-year high and every US-listed uranium miner lost money in 2025. Enrichment is the genuine bottleneck — 77% of the separative work US reactors bought in 2025 — enrichment is sold as effort, priced and purchased separately from the uranium it is performed on — came from foreign suppliers, and Russia alone supplied 26% of it — yet the only listed pure-play enricher runs a business whose 2025 revenue was mostly resold Russian output under a quota that expires in 2028, and whose headline "$2.9 billion backlog" — work it says it has been contracted to do — is $2.3 billion contingent on a plant it has not yet financed. The reactor-design layer, where the market has concentrated its enthusiasm, is where a century of cost overruns actually lives: the first BWRX-300 at Darlington is budgeted at roughly C$25,700 per kilowatt including common works against a design marketed near $2,250, and Georgia Power's share of Vogtle 3 and 4 settled at $10.67 billion for 1,106 megawatts.

Follow the capital rather than the narrative and the map inverts.

First, the merchant nuclear fleet is the only layer in the chain that has converted the AI-demand story into cash without spending capital to do it. A merchant reactor is one whose owner sells its power into the open wholesale market rather than at a price a state regulator sets. Constellation averaged $3.0 billion of operating profit across 2023–2025 on operating invested capital — the money actually tied up in running the business — that moved from $44.1 billion to $45.7 billion — the fleet grew its earnings without being given more capital to do it, because the plants were paid for in the 1970s and 1980s. (The year-to-year path is violent — $1.61bn, $4.35bn, $3.09bn — because GAAP operating income here carries unrealised derivative marks — paper gains and losses on hedging contracts that have not yet settled; the average is the honest number and the single-pair "incremental return" is not.) Talen has now contracted about 85% of its share of Susquehanna to Amazon through 2042; Vistra has signed twenty-year nuclear agreements totalling roughly 3.8 gigawatts. The profit here is an annuity written against an asset base that cannot be replicated inside two decades, and that is a different animal from a windfall on a tight power market.

Second, the pick-and-shovel with a real moat is the naval one, not the commercial one. A pick-and-shovel business supplies everyone digging rather than betting on any one prospector. BWX Technologies states plainly in its own 10-K — the annual report every US-listed company must file with the SEC — that its Nuclear Fuel Services subsidiary "is the sole provider of nuclear fuel for the U.S. Navy." Sixty-eight per cent of 2025 revenue came from the US Government, backlog rose from $4.84 billion to $7.26 billion in a single year — and that backlog is the accounting remaining-performance-obligation figure — revenue under signed contracts the auditors have verified, not an aspiration. The company earned a 12% return on invested capital in 2025 — profit measured against the money tied up to produce it. That is a modest number next to a story stock's chart and an extraordinary one next to the rest of this chain.

Third — and this is the finding that should trouble anyone long the developer layer — the party best placed to judge small-modular-reactor economics has already left. Fluor, NuScale's controlling shareholder and its engineering partner, disclosed in an April 2026 Form 8-K that it realised $2.43 billion of open-market proceeds on a $570 million investment across sales beginning in September 2025; the February and April 2026 tranches contributed $1.82 billion of that total and completed the exit. NuScale reported $565,000 of revenue in the first quarter of 2026 and its weighted diluted share count went from 128 million to 320 million in a year. Oklo had no revenue line and a $33 million quarterly loss. Nano Nuclear describes itself as a “pre-revenue company.” X-energy is the more credible exception: it owns an integrated TRISO-fuel path — fuel baked into ceramic-coated pellets, each grain carrying its own containment — and receives cost-shared government revenue, but its April 2026 IPO financed development rather than proving reactor-unit economics. The layer remains capital formation in search of a delivered cost curve.

Two things qualify all of it. The enrichment chokepoint is real but its listed expression is thin and its economics are leased rather than owned: Centrus earned a 56% return on invested capital in 2025 because the Department of Energy owns the plant, and that arithmetic reverses the moment Centrus finances a plant of its own. And the merchant annuity is contracted, which means it is also capped — Cameco realised US$67.79 a pound in the second quarter of 2026 while spot — the price of a single immediate cargo, as against the long-term contracts almost all utilities actually buy under — printed $85.60, and the same contracting discipline that makes a cash flow durable is what stops it from being a windfall. Durability and upside are usually sold separately. This brief is about the former.

A nuclear kilowatt-hour comes out the end of a nine-step industrial chain whose economics barely track its physical importance. Ore is dug, milled into U₃O₈ — yellowcake — converted to uranium hexafluorideThe gaseous form of uranium, written UF₆ — and a gas is what a centrifuge needs in order to spin the heavier atoms away from the lighter ones. Everything between the mill and the fuel plant happens to uranium in this form., enriched from 0.7% to roughly 5% fissileCapable of splitting and sustaining a chain reaction. Natural uranium is only 0.7% fissile U-235; almost all the rest is U-238, which mostly just sits there. Enrichment is the business of raising that percentage — nothing is added, the useless isotope is separated out. U-235 (or to 19.75% for the high-assay low-enriched fuel that most advanced designs need), fabricated into assemblies, loaded into a reactor, burned for eighteen to twenty-four months, and eventually cooled, stored and decommissioned. Value, in the sense of physical necessity, is created at every step. Profit is captured at almost none of them.

The steps differ in one respect that decides everything else: how long a determined competitor with capital needs to replicate them. A uranium deposit takes seven to fifteen years to permit and build, but there are many deposits and the product is fungible; nobody pays a premium for a particular company's yellowcake. A fuel-fabrication line is specialised but there are several. An enrichment cascadeA long chain of centrifuges plumbed in series, each one nudging the uranium slightly richer in U-235 before passing it to the next. A single machine achieves almost nothing; the cascade is the plant. Building one is a decade-long project under international arms-control inspection. takes roughly a decade from decision to first delivery and requires nuclear-weapons-adjacent technology under international safeguards. The four suppliers who hold over 95% of that market are all wholly or substantially state-owned, so the marginal ton of capacity gets built on political rather than commercial timetables. And an operating reactor with a licence, an interconnectionThe approved physical and contractual connection between a power plant and the grid. Getting one for a new plant now means joining a study queue that runs five to seven years in most US markets — which is why a plant that already has one is worth so much more than a plant that merely could be built. and a trained crew cannot be replicated at all inside fifteen years at any price.

That is what this brief means by durable profit, and it is worth being precise about it. Ordinary profit is what a business earns until someone else notices and copies it. Durable profit — economists call it economic rent — is what survives the copying, because something the competitor cannot buy stands in the way: a licence, a qualified plant, a customer who is not allowed to shop elsewhere. It is not the same thing as a big margin, and it is not the same thing as growth.

When the market talks about nuclear scarcity it usually means uranium. The scarce thing is not uranium. The scarce things are permission and time.

The three small pools that still have to be counted

Conversion is scarce, but the listed access is embedded. Natural uranium cannot enter a centrifuge until U₃O₈ is converted to UF₆. The Department of Energy's advanced-nuclear Liftoff report identifies ConverDyn's Metropolis Works as the only US facility capable of that step and uses roughly 7,000 metric tonnes a year of current capacity. Cameco says its Port Hope facility holds about 18% of world primary conversion capacity. That is physical scarcity; it is not a clean public pure play. The utility ultimately sets the contract price against a very short supplier list, while restart and expansion at existing licensed sites are the commoditising force. Value is captured inside Cameco and private ConverDyn, not by a separately listed US converter.

Fuel fabrication is qualification-gated recurring service, not a design royalty. Westinghouse's three fabrication sites make bespoke light-water-reactorA reactor in which ordinary water both cools the core and moderates the reaction — the conventional, long-licensed design. Proven and understood, which is precisely why the advanced designs that depart from it face a longer regulatory road. assemblies; Cameco's 2025 filing says the Westinghouse core business produced about US$4.3 billion of revenue and that most core cash flow is secured under contracts lasting three to more than ten years. X-energy's TX-1 plant adds a different proposition: vertically integrated TRISO fuel for a reactor fleet that does not yet exist, with government cost share. Fabrication keeps rent where qualification, customer-specific design and an installed base overlap. It loses rent when a second qualified line exists or when the reactor order never arrives. Cameco's 49% Westinghouse stake is the proven listed access; XE is the option on a new fuel standard.

Decommissioning is funded work without a durable listed toll. The NRC requires power-reactor licensees to report trust-fund sufficiency at least biennially and cites a $280–612 million reference range per plant. That creates a real, long-duration service pool. It does not create a sole source: owners competitively procure dismantling, remediation, packaging and storage, while the trust—not a vendor franchise—holds the scarce capital. The regulator controls release of funds and licence termination; the owner controls the bid. With no clean US-listed pure play in the mandated set, the layer is neutral as an industry pool and non-investable here, not silently omitted.

The stack inverts the sector's own narrative. The layers with genuine, defended, cash-generating positions today — the installed fleet and the enrichment/naval-fuel nodes — are the layers where little new competitive capacity can arrive quickly. The layers absorbing the most capital and attention — reactor design and greenfield mining — are the two where the historical record of return on capital is worst.

That is not fashion. It follows from how the chain's incentives are arranged, and most of the conclusions in Part II fall out of that arrangement.

The hyperscaler wave follows straight from that map. A technology buyer facing a five-to-seven-year interconnection queue and a one-to-three-year data-centre build cycle will pay a large premium for megawatts that already exist and are already connected, and will be extremely reluctant to underwrite first-of-a-kind nuclear construction risk. Both halves of that sentence show up in the deal record. Carnegie's June 2026 survey counts roughly 13 gigawatts of announced hyperscaler nuclear agreements, of which about 6.9 gigawatts sit in power-purchase agreements with the big four — and the large, firm, dated ones (Microsoft–Constellation at Crane, Amazon–Talen at Susquehanna, Meta–Vistra across PJMThe wholesale electricity market across much of the eastern United States, and where most of the merchant nuclear fleet in this brief sells its power. Its Texas counterpart, ERCOT, appears later. Neither owns any generation; they run the auctions that set the price everyone else receives.) are all against reactors that were built decades ago. The agreements tied to new designs are development partnerships, not deliveries.

Enrichment is the one node in the Western nuclear chain where demand cannot be met, cannot quickly be met, and where the incumbents are all governments.

A chokepoint, in the sense this brief uses the word, is a step in a long industrial chain that everything upstream and downstream must pass through, and that very few parties can perform. Ore can be dug in many countries and a fuel assembly can be pressed in several factories; the enrichment in between is performed by four suppliers. That is what makes the word worth using — not that the step is difficult, but that going around it is not an option, and the queue to join the list of people who can do it is a decade long.

Centrus's own 10-K puts the global commercial enrichment market at roughly 50 million SWU a year, with four suppliers holding over 95% of it: Rosatom/TENEX at about 27 million SWU of capacity, Urenco at about 17 million, China's CNEIC at about 11 million and Orano at about 8 million. Every one of those is state-owned or state-controlled. Centrus's own share of the LEULow-enriched uranium — uranium raised to roughly 5% U-235, the fuel conventional power reactors run on. The ordinary product of the enrichment industry, as distinct from the high-assay grade the new designs want. market is, in its words, "less than 5%" — and that share is largely brokered rather than produced, because the company's operating enrichment asset is a single demonstration cascade in Piketon, Ohio, built for the Department of Energy and licensed to produce roughly 900 kilograms of high-assayEnriched further than normal fuel — to 19.75% U-235 rather than about 5%. Assay is simply the percentage. The jump sounds small and is not: reaching that concentration consumes several times the enrichment effort per kilogram, so a fleet of advanced reactors would strain the same constrained plants far harder than its size suggests. Hence HALEU, high-assay low-enriched uranium. fuel a year.

The American position is where the investable question lives. The Energy Information Administration's 2025 Uranium Marketing Annual Report — published 29 July 2026, three days before this brief's evidence cutoff — records that US civilian reactor operators purchased 13 million SWU from four sellers in 2025 at an average of $108.70 per SWU, up 11% from $97.66 the prior year. Of that, US-origin SWU was 23%. Foreign-origin was 77%, of which Russia supplied 26%, France 18%, the United Kingdom 14% and the Netherlands 8%.

Read that again. Two years after the Prohibiting Russian Uranium Imports Act took effect in August 2024, Russia remained the single largest source of enrichment services consumed by American reactors. The ban permits waivers, the Russian Suspension AgreementA US trade settlement that, instead of imposing anti-dumping duties on Russian uranium, allows a capped quantity into the country each year. Its quotas — and the carve-out written into them for Centrus — are what keep Russian enrichment legally flowing to American reactors despite the import ban. They expire in 2028.'s quotas run to 2028, and — as Centrus's 10-K discloses in unusual detail — the 2020 amendment to that agreement "explicitly sets aside sufficient quota in 2021 through 2028 for Centrus." The chokepoint is not a future risk. It is a present dependency with a legislated expiry date.

The obvious rebuttal is that capacity is being added, and it is. The question is how fast. Urenco USA's National Enrichment Facility at Eunice, New Mexico — in Urenco's own words, "the United States' only commercial uranium enrichment facility" — currently runs 4.3 million SWU a year, which the company describes as approximately one third of current US demand. An expansion adding 700,000 SWU completes in 2027. A second, larger expansion of 2.1 million SWU was announced on 2 June 2026: construction begins in 2029, the first cascades produce in 2032, and full production arrives in 2036. Total installed capacity reaches "more than 7 million SWU over the next decade."

Seven million SWU in 2036, against thirteen million consumed in 2025 — before a single new reactor comes online, before a single restart, and before the high-assay demand that advanced designs will add. The commodity that is supposed to compete away the enrichment rent takes ten years to arrive and, on the incumbent's own published schedule, does not close the gap.

So the chokepoint is real, and unusually well documented. The harder question, the one that decides whether it is investable, is who captures the rent it generates.

Here the answer is uncomfortable for the obvious trade. Urenco and Orano are private and state-controlled; they are not purchasable, and their role in this analysis is to define the ceiling rather than to be owned. The listed vehicle, Centrus Energy, holds a first-mover position in American-owned enrichment and a genuine technology asset — but it also holds a set of constraints that the market's framing tends to skip past. The Piketon plant is leased from the Department of Energy. The centrifuge intellectual property is subject to a 2002 agreement under which DOE holds an irrevocable licence it may exercise commercially if Centrus misses a milestone or is judged to have abandoned deployment, with a royalty back to Centrus capped at $665 million. And the July 2026 award that transformed the story is a fixed-price $900 million contract to deliver, by March 2032, one metric ton of HALEU as uranium hexafluoride plus a specified amount of deployed capacity — with options for up to $170 million of purchases bringing the total to $1.07 billion. Fixed-price government construction contracts transfer cost overrun risk to the contractor.

A reactor in the United States earns money in one of two entirely different ways, and confusing them is the fastest route to a wrong conclusion about this sector.

A regulated reactor — Vogtle, Summer's surviving units, Duke's Carolinas fleet, Dominion's Surry and North Anna — is a rate-base asset. Its owner earns an authorised return on the equity portion of its undepreciated capital cost, set by a state commission. Georgia Power's retail return on equity is fixed at 10.50% with an equity ratio of 56%, evaluated against a band of 9.50% to 11.90%, above which two-fifths of the excess is refunded to customers and only one-fifth retained. Virginia Power's authorised ROE was set at 9.70% in the 2023 biennial review. Those numbers are the ceiling and, in practice, close to the floor. A regulated reactor cannot earn a windfall from tight power markets because its output is not sold into one; it also cannot lose money on a cold winter. It is a bond with an operating risk attached.

A merchant reactor — Constellation's fleet, Talen's Susquehanna, Vistra's Comanche Peak and its PJM units, Dominion's Millstone — sells into a wholesale market. Dominion's own 10-K states the position with unusual candour: Millstone "does not have a predetermined rate structure that provides for an ROIC," operates within a regional transmission organisation, and "primarily competes on the basis of price." The merchant asset's return is whatever the market and its contracting skill produce.

Which is why the last two years matter so much. The hyperscalers have not increased the number of nuclear megawatts. They have converted merchant megawatts into contracted ones at prices that reflect scarcity rather than marginal cost, and in doing so re-rated an entire asset class from cyclical to annuity.

Three structures, three different economics, and they are not interchangeable.

Constellation–Microsoft at Crane. A twenty-year power purchase agreement signed in the third quarter of 2024 to restart the retired Three Mile Island Unit 1 as the Crane Clean Energy Center, about 835 megawatts, with financing support from a Federal Financing Bank loan maturing in October 2055 at 37.5 basis points over comparable Treasuries. This is the purest expression of the thesis: a plant that was closed for economics in 2019 is now worth restarting because a single counterparty will underwrite twenty years of it. The asset existed; the offtake created the value.

Talen–Amazon at Susquehanna. The most instructive of the three, because it was tested. Two words carry the whole dispute. A load sited behind the meter takes its power directly from the plant without the electricity ever touching the public grid — and therefore, its sponsors argued, without paying for the grid. A load sited front of the meter is an ordinary customer: the plant sells into the network, the network sells to the customer, and the wires owner is paid along the way. The physical arrangement can be identical in both cases; what differs is who gets a bill. The original structure put the data centre behind the meter, drawing directly from the plant; FERC rejected the amended interconnection agreement in a 2–1 order on 1 November 2024, finding the parties had not justified the expanded behind-the-meter arrangement. The parties restructured. Under the amended June 2025 agreement, Susquehanna delivers to the PJM grid, Talen acts as retail supplier to Amazon, PPL handles transmission — a "front-of-the-meter" arrangement that, as Talen's own 10-K notes, is not subject to the same approval. The volume is 1,920 megawatts through 2042, ramping to full quantity no later than 2032. Against Talen's 90% share of Susquehanna's 2,494 megawatts — 2,245 megawatts — that is roughly 85% of the asset, contracted for sixteen years.

Vistra–Meta and Vistra–Amazon. Twenty-year agreements totalling about 3,800 megawatts: 1,200 megawatts from Comanche Peak to AWS with deliveries beginning in the fourth quarter of 2027 and ramping to full capacity by 2032, and more than 2,600 megawatts of energy, capacity and upratesRegulatory permission to run an existing reactor at higher output than its original licence allowed, usually after equipment upgrades. It is the cheapest megawatt in the industry — new capacity without a new plant — but it requires NRC approval, so an announced uprate is a request, not yet a megawatt. with Meta across the PJM nuclear units. Carnegie's survey breaks the Meta transaction into 2,176 megawatts of operating generation and 433 megawatts of uprates, contingent on NRC approval of the licence extensions.

In all three the buyer is paying for certainty and speed, and the seller is paying for that premium with its optionality. A merchant plant fully contracted for twenty years has traded the right to a spike for the certainty of a coupon. Whether that is a good trade depends entirely on what happens to power prices — and the honest answer is that nobody knows, which is precisely why both sides signed.

The same instinct, applied to nearer years, is what a generator means by hedging. To hedge is simply to sell output forward at a known price before the year arrives, so the revenue is fixed whatever the market later does. A fleet reported as "84% hedged for 2027" has already priced 84% of what it expects to generate; only the unsold remainder actually rides on the market. Read that way, the hedge table is not a measure of caution — it is a map of where a merchant generator's earnings are still genuinely uncertain, and when.

The merchant metric block: what the filings disclose—and what they do not

Sources: Constellation FY2025 10-K; Vistra FY2025 10-K; Talen FY2025 10-K. “N/D” is a finding, not a data gap filled by estimation.

The regulatory overlay changed twice. FERC's 18 December 2025 order in Docket EL25-49 found PJM's large-load and behind-the-meter rules no longer just and reasonable and required three service paths: network integration service, firm contract demand and non-firm contract demand. The reportable event at this brief's cutoff is the 18 June 2026 final order on rehearing, compliance and paper hearing. FERC set aside parts of December, accepted and rejected parts of PJM's compliance filing, established rates and terms for the new services, and ordered further filings within 60 days.

What changed is who bears the charge. Network service can still be charged on gross demand, but a qualifying co-located load may choose firm or non-firm contract-demand service and be charged on its reserved contract demand irrespective of usage. That is not a free behind-the-meter bypass and it is not full freight on gross load. It is a regulated toll sized to the grid service reserved. The Talen–Amazon front-of-meter restructuring remains the worked example: the PPA survived, PPL took transmission and delivery, and Talen kept a premium-priced generation contract. The final order therefore narrows the original thesis. Wires capture a bounded slice when service is reserved; they do not automatically confiscate the co-location premium.

Exelon, which owns no generation, is still a second-order beneficiary where large load requires new transmission investment approved into rate base. But the benefit is conditional on topology, contract demand and commission-approved capital—not a permanent percentage of every nuclear-data-centre deal. That is lower-return and lower-risk exposure, and now stated at the scale the final rule supports.

The capacity market supplies a separate cash-flow floor. PJM's 2026/27 Base Residual AuctionPJM’s forward capacity auction, run ahead of the delivery year it covers. Generators bid to be paid for standing ready; the price clears where supply meets the grid’s reliability requirement. It is the mechanism that puts a second, availability-based revenue line under every existing plant. cleared 134,310.8 MW of UCAPUnforced capacity — a megawatt discounted for how often the unit actually fails to show up. It is the honest unit of account in a capacity market: a reactor that almost never stops contributes far more UCAP per nameplate megawatt than a plant that does. at the $329.17/MW-day RTO cap; including fixed-resource-requirement supply, the system finished only 139 MW above its reliability requirement. That does not disclose a nuclear plant's realised price, but it shows why an existing accredited megawatt has a second scarcity value before any hyperscaler PPA is signed.

The PTC and the state zero-emission-creditState subsidies paid to nuclear plants for generating carbon-free power, introduced to stop reactors closing on economics. Like the federal credit, a ZEC props up the bottom of the revenue range rather than adding anything to the top. programmes put a floor under the existing fleet's cash flow; the hyperscaler contracts put a premium above it. A merchant nuclear operator today owns a bounded-downside, contracted-upside cash stream on a fully depreciated asset. There is not another power-generation asset class in the United States that can say that sentence.

Every generation of nuclear technology has been sold on the promise that this time the units will be standardised, factory-built and cheap. The claim is not absurd — it is how aircraft, ships and gas turbines got cheap — and it may eventually be right. It has never yet been right on the schedule promised, and the current cycle is producing its evidence in real time.

The regulatory picture has genuinely improved. On 4 March 2026 the NRC authorised a construction permit for TerraPower's Kemmerer Power Station Unit 1 in Wyoming — the first commercial reactor construction approval in nearly a decade and the first for a non-light-water design in more than forty years, with the technical review completed in under eighteen months. The ADVANCE Act also cut the qualifying advanced-reactor applicant rate to $148 an hour from the $318 full-cost rate from October 2025. Faster reviews and lower fees remove friction. They do not finance concrete, long-lead equipment or fuel.

The construction bottleneck is not. Ontario Power Generation's Darlington New Nuclear Project — the first grid-scale BWRX-300, the design GE Vernova is offering into the US market and the one the Tennessee Valley Authority has applied to build at Clinch River — is budgeted at C$6.1 billion for the first unit plus C$1.6 billion of systems and services common to all four, with a total four-unit budget of C$20.9 billion for roughly 1,200 megawatts. Ontario approved the final investment decision on 8 May 2025 with the first unit targeted in service in 2030.

The BWRX-300 was introduced to the market at a price point of roughly $2,250 per kilowatt. The first unit plus its common works comes to about C$25,700 per kilowatt; the full four-unit programme, on OPG's own budget and including interest and contingency, to about C$17,400 per kilowatt. Even granting every reasonable adjustment — Canadian dollars, financing costs included where the marketing number excluded them, first-of-a-kind learning that later units will not repeat — the gap is not a rounding error. It is roughly an order of magnitude on the first unit and something close to seven-fold across the programme.

And Darlington is the good case. It has a government owner with a balance sheet, an experienced nuclear operator, an existing licensed site, an integrated project-delivery contract structured specifically to control overruns, and a regulator that granted its construction licence. Georgia Power's share of Vogtle 3 and 4 — a project widely treated as the cautionary tale — landed at $10.670 billion of final net investment for 1,106 megawatts of its ownership share, about $9,650 per kilowatt excluding $440 million of capitalised financing costs. The much-derided large reactor came in cheaper per kilowatt than the celebrated small one is budgeted at.

That inversion has a straightforward explanation. Small modular reactors are cheaper per unit and more expensive per kilowatt until volume arrives, because the fixed costs of a nuclear island — containment, safety systems, licensing, security, staffing — do not scale down proportionally with power output. The economics only work at the fourth or tenth or twentieth unit. Which means the entire SMR investment case rests not on the first plant working, but on someone ordering enough of them, fast enough, that the learning curve bites before the capital runs out.

Against that requirement, the order book is thin. NuScale's own first-quarter 2026 disclosure describes its exclusive partner ENTRA1 "continuing its work" with TVA to "progress planning" for up to six gigawatts, and Nuclearelectrica shareholders approving "proceeding with the next phase" of the RoPower project at Doiceşti. Those are real milestones and they are not orders. The Fluor FEEDFront-end engineering and design — the paid study phase in which a contractor works a concept up into a buildable design with a credible cost estimate. It is real revenue and a genuine signal of intent, but it ends before a shovel moves; a completed FEED is the step before anyone decides whether to build. Phase 2 engineering work on RoPower completed in late 2025 with, in NuScale's words, "no comparable activity in 2026" — which is why revenue fell to $565,000 in the quarter. Oklo has signed a DOE Other Transaction AgreementA flexible federal contracting instrument that sits outside the ordinary procurement rules, used to move fast on research and demonstration work. Here it matters for a specific reason: it lets a reactor be built on a federal site under the Department of Energy’s own authority, bypassing NRC licensing until commercial operation. for its Aurora powerhouse at Idaho National Laboratory under the Reactor Pilot Program, had its Nuclear Safety Design Agreement approved by DOE's Idaho Operations Office, and had its NRC Principal Design Criteria topical report approved on an accelerated schedule. That is a faster path, executed under DOE authorisation rather than NRC licensing, with NRC licensing "subsequently" pursued for commercial operation. Nano Nuclear is developing the KRONOS microreactor at the University of Illinois and describes itself as pre-revenue.

X-energy is now part of the public comparison. The company completed its IPO on 27 April 2026 and filed its first 10-Q as Nasdaq-listed XE. The Q1 filing reports $42.4 million of revenue and grant income, 92% from the US Government; $67.3 million of operating cash use; $43.0 million of capital expenditure before $28.8 million of grant reimbursement; and $944.0 million of cash plus short- and long-term investments at 31 March, before about $1.1 billion of net IPO proceeds. That is materially more runway and industrial substance than the three pre-revenue peers. It is not a reactor order. X-energy says Amazon is not obligated to purchase its priority queue slots, customers may delay, reduce or terminate commitments, and it has not yet entered a technology-fee agreement. DOE and Dow cost sharing make the first plant more financeable; they do not yet show that the fourth unit earns its cost of capital.

The buyer incentives say the same thing. Google's Kairos agreement targets up to 500 MW, first power by 2030 and later deployments through 2035. Amazon receives priority manufacturing slots, rights of first refusal and most-favoured pricing from X-energy; X-energy warns those terms may compress margins. Hyperscalers have bought optionality and queue position around new designs while buying firm output from existing reactors. That asymmetry is the evidence, not a criticism of the technology.

The distinction that matters for capital allocation is not between good and bad companies. It is between a firm, financed, dated order and everything else. Backlog in this layer is a promise to consider building; backlog at BWX Technologies is an accounting obligation the customer has already committed funds against. Both get called "backlog." They are not the same instrument.

It is worth knowing exactly where the difference lives, because it decides several grades in this brief. "Backlog" is a word a company chooses for itself, and it can mean anything from a signed and funded order to a letter of intent nobody has priced. The remaining performance obligation is not: under the accounting standard, a company may only report revenue it is contractually entitled to for work it has not yet done, and its auditors have to agree. When a company discloses that its backlog equals its remaining performance obligations, it has quietly said the two numbers survive an audit. When it discloses them separately, the gap between them is the part that depends on something that has not happened yet.

The market has correctly identified that firm carbon-free power is scarce and incorrectly identified who gets paid for the scarcity — it has bid up the layer that sells the idea of new reactors and left comparatively cheap the layers that own the old ones and hold the fuel-cycle and defence chokepoints.

Three claims carry the brief, each stated at the confidence the evidence supports.

Two further conclusions run against the sector's own consensus rather than against the market's price, and both are inferences rather than observed facts — labelled accordingly.

The enrichment chokepoint is more durable than the bears think and less capturable than the bulls think. Urenco's published schedule — first new cascades in 2032, full production 2036, total capacity above 7 million SWU against 13 million consumed in 2025 — means the physical shortage persists well into the 2030s. But the listed vehicle's spectacular 2025 return on capital (56%) is an artefact of not owning the plant it operates. As Centrus finances a plant of its own, its capital base grows by billions and that return converges toward a regulated-utility number. The chokepoint's rent is real; the equity's claim on it is being diluted by the very expansion that justifies the equity. Inference — mechanism disclosed, magnitude not

The wires companies can capture a durable, unlevered slice of the datacentre-nuclear trade while owning no nuclear at all. FERC's June 2026 final order charges co-located large load on reserved contract demand under the service it elects; the Talen–Amazon restructuring shows a front-of-meter form, with PPL taking transmission and delivery. Southern's traditional operating companies have contracted roughly 16 gigawatts of large load since 2023 with minimum-bill provisionsA contract term obliging the customer to pay for a floor quantity of power whether or not it uses it. It is the utility’s answer to the risk that a data centre is announced, built into the rate base, and then never fully switched on — the customer, not the ratepayer, carries the empty capacity. and, on about 13 gigawatts, minimum duration, termination payments and posted collateral. That is contracted volume growth at an allowed return — modest, but conditional on network configuration and approved investment. Inference

The whole question resolves, layer by layer, into a single test: can a competitor with capital and patience replicate this position inside a decade? Where the answer is no, excess returns persist. Where the answer is yes, activity expands and margin does not.

Three layers earn an unqualified yes, and they have nothing in common technologically. What they share is a licence, a physical plant, or a customer relationship that cannot be duplicated at any speed a return-seeking investor would tolerate. That one criterion explains the sector's oddest pricing: the layers where new capital is most eager to go — mining and reactor design — are precisely the layers where new capital can go, which is why nothing there stays scarce.

The regulated fleet is marked "capped, not competed," and the distinction is real but it cuts both ways. A capped return is a low return; it is also an extraordinarily reliable one, and in a chain where three-quarters of the listed names lost money in 2025, reliability has a value the ROIC column understates.

A moat is an answer to one question: what happens when the thing that makes you scarce stops being scarce? Applied to this chain the question takes six specific forms, and each has a dated, sourced answer as of this brief's cutoff.

What if Western enrichment additions erase the HALEU shortage? They do not, on the incumbents' own schedules, before the mid-2030s. Urenco USA goes from 4.3 to about 5.0 million SWU by 2027 and to "more than 7 million" only by 2036, against 13 million SWU of US demand in 2025. Centrus's own initial build-out is 12 metric tons a year of HALEU capacity with first new capacity "expected to come online by 2029." The shortage narrows; it does not close. Known Known — both schedules published by the operators

What if an SMR design hits its cost target at the fourth unit? Then the design layer becomes genuinely valuable — and the enrichment and fuel layers become more valuable, not less, because a standardised fleet multiplies HALEU demand against the same constrained supply. This is the sector's most important second-order effect and Section 10 develops it. Note the timing: OPG's fourth Darlington unit is targeted for the mid-2030s. Nobody will know before then. Known Unknown

What if uranium mean-reverts? Very little happens to the durable-profit map, because almost nobody in it is levered to spot. US utilities bought 87% of their 2025 volume under long-term contracts at $55.91 a pound while spot averaged far higher; Cameco realised $67.79 in the second quarter of 2026 against an $85.60 spot print. A uranium bear market would hurt the junior miners badly — they are already lossmaking at an eighteen-year-high price — and would barely touch Constellation, BWXT or the regulated utilities. Known Known

What if the merchant PPAs re-contract lower? This is the genuine risk to the annuity, and it is a 2040s question, not a 2020s one. Talen's Amazon contract runs to 2042; Constellation's Microsoft agreement is twenty years from 2024; Vistra's are twenty years with extension options. What re-contracts sooner is the uncontracted remainder, and there the disclosure is specific: Vistra had hedged about 100% of expected 2026 generation, 84% of 2027 and 58% of 2028 as of February 2026. The exposure is real from 2028. Known Known

What if a regulator caps a utility's nuclear returns? They already have — that is what an allowed ROE is. Georgia Power at 10.50% with sharing above 11.90%; Virginia Power at 9.70%. The regulated question is not whether returns get capped but whether the capital programme gets approved, and on that the evidence is currently permissive: Southern reported weather-adjusted commercial sales up 10.9% at Georgia Power in the second quarter of 2026, "largely driven by data centers." Known Known

What if FERC closes the behind-the-meter path? It largely has, and the sector adapted in about seven months. The relevant conclusion is not that co-location is dead but that its premium is now shared with the transmission owner. Estimated

Every grade below rests on that company's own opened filing or release. Grades are structural — advantaged, neutral or exposed on competitive economics — and carry no view on price. The return-on-capital arithmetic behind them is set out once, in full, at the end of the section.

Uranium's long-term indicatorA published estimate of the price at which multi-year supply contracts are currently being written, as distinct from the spot price of a single immediate cargo. Since utilities buy most of their uranium under long-term contracts, this is the number closer to what producers will eventually receive — though, as this section shows, still well above what they receive today. reached US$93.00 a pound on 31 March 2026, the highest in more than eighteen years — and three of the four US-listed producers still ran operating losses in 2025. That is not a bull market failing to arrive. That is a layer whose economics do not work at prices the sector spent a decade praying for.

The price the sector quotes is not the price anyone receives.

US utilities bought 87% of their 2025 volume under long-term contracts at $55.91 a pound. The spot print that generates the headlines applied to 13% of the market. A miner's revenue is set by a contract book negotiated years earlier, which is why a producer can watch spot rise 50% and see its realised price rise 18%. That is not a moat being harvested; it is a cyclical spread arriving slowly, and for the sub-scale producers it arrives too slowly to cover the cost of the capital they raised waiting for it.

Primary sources: Cameco FY2025 MD&A; Energy Fuels Q1 2026; UEC Q2 2026 release; Ur-Energy FY2025 10-K; Denison FY2025 annual exhibit.

All three of the facts that decide whether this position converts into durable profit are disclosed, not inferred.

The backlog is not what the headline says. Centrus reports $3.8 billion of total backlog across both segments extending to 2040, of which the LEU segment is $2.9 billion. Of that $2.9 billion, $2.3 billion is contingent on potential new Piketon capacity: $2.1 billion under definitive agreements and $0.2 billion still subject to definitive agreements. The accounting measure is smaller. At 31 March 2026 LEU-segment RPO was $0.7 billion and total company RPO was $0.8 billion, extending to 2030.

Roughly one dollar in five of the headline LEU backlog is an obligation the accountants recognise; four in five are conditional on a plant being financed. Both numbers are honestly disclosed by the company. Only one of them is a receivable.

The current business is substantially resold Russian material. Under the 2011 TENEX Supply Contract, which runs through 2028, Centrus buys separative work contained in low-enriched uranium from Russia's state supplier and delivers natural uranium hexafluoride in exchange. Its 10-K states that the amended Russian Suspension Agreement "explicitly sets aside sufficient quota in 2021 through 2028 for Centrus," and that the contract obliges the company to pay for its minimum annual purchase whether or not it orders the material. The economics that produce today's 56% return therefore rest on a sanctioned counterparty and an expiring quota — which is exactly why the company is racing to build.

The expansion is capital-intensive and the funding is not settled. Centrus signed an engineering, procurement and construction contract with Fluor Federal Services on 9 February 2026 for the Piketon expansion, "on a time and materials basis at agreed labor rates." That phrase is the whole tell. Under a fixed-price contract the builder names a number and eats any overrun; under time and materials the owner pays for the hours and the steel however many there turn out to be, and the builder cannot lose. The choice does not change what the plant costs — it decides who finds out the hard way. Fluor is not taking fixed-price risk on this, which is how contractors now price first-of-a-kind nuclear construction. Centrus's own 10-K says it expects to increase capital expenditures "by approximately several hundred million," and describes a funding framework of national-security missions, third-party prepayments, direct foreign investment and commercial contracts. Capital spending was $23.2 million in the first quarter of 2026 alone, exceeding the $19.7 million spent in all of 2025.

Urenco and Orano, the private incumbents. Neither is investable, and both bound the opportunity. Urenco's 17 million SWU of global capacity and Orano's 8 million exist and are contracted; Urenco's American expansion timetable — 2029 construction start, 2032 first cascades, 2036 full output — is the clock Centrus is racing. If the incumbents accelerate, Centrus's window narrows. If they do not, Centrus is the only American answer to a shortage that persists into the mid-2030s. That is the whole investment question in this layer, and it is genuinely open.

This is the layer the mandate anticipated would hold little durable profit. NuScale is effectively pre-revenue; Oklo and Nano are pre-revenue; X-energy earns cost-based services and grant income but no technology fee or reactor-unit return. Every name is funded by equity issuance, customer or government cost share rather than operating free cash flow. What follows is not a judgment on the technology. It is a statement about who is earning a positive return on deployed commercial-reactor capital today, and the answer is nobody.

The strongest rebuttal is a fair one: a company can sell a stake for balance-sheet reasons that say nothing about the asset. Fluor's 2025 operating loss and its history of fixed-price project charges give it every reason to raise cash. Then look at what Fluor did next. In February 2026 it signed the Piketon expansion contract with Centrus "on a time and materials basis at agreed labor rates." A contractor that has been burned by fixed-price nuclear construction sold its reactor-design equity and took cost-plus work on the fuel-cycle build. That is a coherent view of where the risk sits, expressed twice with its own money.

One correction before the grades. Exelon owns no nuclear generation. Its Q2 2026 10-Q describes it as "a utility services holding company engaged in the energy transmission and distribution businesses through ComEd, PECO, BGE, Pepco, DPL, and ACE" — the generation fleet was separated into Constellation in 2022. Exelon is graded here as what it is: a pure wires company with material indirect exposure to the same demand.

Method: “N/D” means the filing does not isolate undepreciated nuclear rate base at the parent level. The capital proxy uses total company capex and therefore measures funding intensity, not nuclear economics. That limitation is preferable to allocating rate base without the commission schedules required to do it.

The return-on-capital derivation, shown in full

Every grade above that leans on capital efficiency rests on the table below. It is computed, not disclosed, and the method is stated so a reader can disagree with it precisely.

One number does most of the work, so it is worth saying plainly what it measures. Return on invested capital asks a single question: for every dollar tied up in this business — plants, inventory, working capital, whatever the owners had to fund — how many cents of after-tax operating profit came back this year? Profit margin cannot answer that, because a business can earn a fat margin on an enormous pile of capital and still be a poor use of money. The comparison that matters is against what the capital could earn elsewhere, which is why the table sets every company against the same 8% hurdle. Clear it and the business is creating value; miss it and the business is growing, honourably and usefully, while destroying it.

Method. NOPATNet operating profit after tax — what the business earned from operations, taxed, but before any effect of how it is financed. Using it means a company loaded with debt and a company with none are compared on the same footing: on what the assets earn, not on who lent the money. = operating income × (1 − 21%); Cameco uses its disclosed 26.9% effective rate and remains in Canadian dollars throughout. Invested capital = total assets − cash and equivalents − current liabilities, averaged across opening and closing balance sheets. ROIC spread is measured against a transparent 8% comparative hurdle, not a security-specific WACCWeighted average cost of capital — the blended return a specific company must earn to satisfy both its lenders and its shareholders. It is the theoretically correct hurdle, and it is different for every issuer, which is exactly why this brief uses one flat 8% instead: a common yardstick a reader can argue with beats a set of bespoke ones a reader cannot check.; it is a deliberately common yardstick, informed by the 2026 risk-free-rate environment but not presented as an estimate of any issuer's exact cost of capital. Incremental ROIC = change in NOPAT ÷ change in invested capital from 2024 to 2025; “n/m” means the capital denominator contracted, was immaterial, or the business remained pre-profit. FCF margin = (operating cash flow − total capital expenditure) ÷ revenue. Total capex is a conservative proxy because sustaining capex is not separately disclosed consistently; this is not labelled sustaining FCF. Source frames are each issuer's filed XBRL data and Cameco's 2025 results. Comparability limits: derivative marks make Constellation, Vistra and Talen incremental returns economically noisy; GEV's 45.2% is a whole-company recovery metric, not a nuclear-segment return; Centrus leases rather than owns its plant; utilities' tax credits and construction programmes depress this uniform screen. Denison, Nano Nuclear, X-energy and Sprott lack a meaningful 2025 operating-return base and are assessed on funding and contract quality instead.

The horizontal axis below is explicitly qualitative: it ranks how dependent the current equity story is on undelivered nuclear, uncontracted power or commodity upside. It does not purport to measure fair value without a consistent live valuation dataset. The vertical axis applies the durability test consistently: can a funded competitor replicate the position within a decade, and does the company earn a positive return on the capital it employs?

The strongest argument against everything above is not that the merchant fleet is bad or that BWXT is overrated. It is that this brief has mistaken the top of a cycle for a structure.

Every generation of this industry has been sold a renaissance. One study of 180 nuclear projects found that 175 exceeded their initial budgets and timelines.

The historical base rate, as cited in the July 2026 bear case

The bear case. Start with what actually happened the last time. The Energy Policy Act of 2005 launched a nuclear renaissance underwritten by federal loan guarantees. US utilities proposed more than thirty reactors. Four entered construction. Two were abandoned mid-project after more than $9 billion had been spent, with South Carolina ratepayers still paying for electricity they never received. The two that finished — Vogtle 3 and 4 — cost more than $36 billion against roughly $14 billion estimated at construction start. That was the renaissance. It fizzled through a specific mechanism: loan guarantees transfer risk to the public, which, as the Congressional Budget Office warned in 2008, "cause recipients to invest in excessively risky projects because they do not bear all the cost of a project's failure."

Now apply that to today. The economics have not improved. Each unit of energy from a new US reactor costs roughly three times the equivalent from solar or wind, and the trend lines diverge: nuclear has become more expensive over time while renewables have become cheaper. Small modular reactors are worse, not better, on a per-unit-energy basis — the Idaho NuScale project was cancelled in 2023 because $9.3 billion for 462 megawatts was too expensive for the utilities that would have bought the power, and TerraPower's Natrium is estimated at $9.4 billion for 345 megawatts. The only recent SMR experience outside the West, Russia's KLT-40S, took thirteen years from construction start to generation against an expected three.

Meanwhile the demand the whole thesis rests on may not need nuclear at all. A single SMR is under 300 megawatts. Stargate's Abilene campus wants up to five gigawatts; Meta is planning five gigawatts in Louisiana; in January 2026 a Texas air permit was granted for 7.65 gigawatts of gas-fired generation, the largest in the United States. Gas is faster, cheaper and buildable now. If the hyperscalers were serious about nuclear as a primary answer rather than a procurement hedge and a public-relations position, the deals would be firm orders rather than development agreements — and after two years of announcements, the firm orders are still, overwhelmingly, contracts for reactors that already exist.

And the merchant "annuity" is a bet on a single customer class in a single capex cycle. The contracted premium exists because four companies decided simultaneously that they needed power in 2027 rather than 2032. If AI capital spending digests — and it has digested before — those contracts remain, but the incremental deal flow stops, the capacity-market prices that support the uncontracted tail normalise, and a merchant generator trading at a premium multiple on scarcity re-rates back to what it has always been: a levered call on power prices.

Where the bear case is right

On the reactor-design layer, it is simply correct, and this brief's grades already reflect it. The Darlington and Vogtle cost data, the Fluor exit and the $565,000 revenue quarter are not counter-evidence to the bear case; they are the bear case, sourced from the companies themselves. Nothing here argues that SMRs will be cheap or that new build will be profitable. The 2005 analogy is apt, and the base rate on cost overruns is the single most reliable regularity in this industry.

It is also right that gas, not nuclear, is meeting the marginal datacentre load. EPRI's own scenario work concludes that under current policy natural gas dominates near-term incremental supply, with build rates in the high scenario more than double the recent average. Constellation buying Calpine and Talen buying Freedom and Guernsey are two nuclear operators voting with $10 billion that the near-term answer is a gas turbine.

Where it fails

It fails on a category error: the bear case is an argument about new nuclear construction, and the durable profit identified in this brief is not in new nuclear construction. Every one of its strongest facts — 175 of 180 projects over budget, Vogtle at three times its estimate, SMRs at three times the cost of wind — is a statement about the cost of building a reactor. None of them is a statement about the value of one that already exists. In fact they are the opposite: the harder and costlier it is to build a reactor, the more valuable the 96 already operating in the United States become, because replacement cost is the ceiling on what a competitor could ever charge and the floor under what an incumbent can. A bear case built on construction cost overruns is, read carefully, the strongest available bull case for the installed base.

It fails a second time on the enrichment node, where the bear case has nothing to say at all. Whether SMRs are built or not, the 96 existing reactors need 13 million SWU a year, 77% of which came from abroad in 2025 and 26% from Russia, under a quota expiring in 2028, against a domestic expansion that reaches full output in 2036. That shortage is indifferent to reactor construction economics.

And it fails a third time on the naval layer, which is not a commercial market at all. The US Navy's reactor programme is funded by appropriation, not by the levelised cost of electricityThe all-in cost of a megawatt-hour over a plant’s whole life — construction, fuel, operation and financing — spread across every unit it will ever produce. It is the standard yardstick for comparing a reactor against a gas turbine or a solar farm, and the one on which new nuclear currently loses. It is also irrelevant to a navy, which is the point being made here., and BWXT's position in it does not depend on whether a single commercial SMR is ever built.

What survives is an important qualification rather than a refutation. The bear case correctly identifies that the merchant re-rating is partly cyclical — a windfall from one capex cycle's urgency — and this brief does not claim otherwise. What it claims is narrower and, it argues, more defensible: that the portion of the merchant cash flow now written into twenty-year contracts with investment-grade counterparties has been converted from cyclical to structural, and that the conversion is irreversible for the contract's duration. Talen's 85%, Vistra's 3.8 gigawatts and Constellation's Crane agreement are not forecasts. They are signed.

Three consequences follow from the structure above that are not currently priced into the way the sector is discussed, and each is falsifiable.

1. If SMRs succeed, the enrichment chokepoint tightens rather than loosens — and the fuel layer captures more of the value, not the design layer. Most advanced designs need high-assay low-enriched uranium at 19.75% rather than the roughly 5% conventional reactors use, and enriching to that assay consumes several times the separative work per kilogram of fuel. A standardised, successful SMR fleet therefore multiplies HALEU demand against a supply base whose entire American expansion is a single leased cascade in Ohio and a Urenco plant that reaches full output in 2036. The design layer's success is the fuel layer's pricing power. The natural corollary — that a reactor developer's rational move is to integrate backwards into fuel — is already visible in DOE's HALEU allocation rounds, where TRISO-X, TerraPower, Kairos, Radiant and Westinghouse have received conditional commitments precisely because they cannot source the fuel commercially. Estimated — mechanism disclosed, magnitude not

2. FERC's co-location framework makes the wires share a function of reserved demand, not a permanent toll on gross datacentre load. The behind-the-meter structure sought to avoid transmission charges. FERC's June 2026 final order preserved three service paths but charged firm contract-demand and non-firm contract-demand customers on the capacity they reserve. The Talen–Amazon restructuring shows one front-of-meter outcome, with PPL handling transmission and delivery; it does not prove every project will take that form. Generators may retain more of the premium when reserved demand is materially below gross load, while wires owners earn only on service and approved network capital actually used. Inference — final rule, implementation pending

3. The contracting that makes merchant nuclear durable is the same act that transfers the option value to the buyer — and the market is valuing the contracts as if the seller kept both. A twenty-year fixed-price PPA on an existing reactor converts a volatile cash flow into a bond. That is worth a great deal, and it is why this brief grades the merchant operators advantaged. But it also means that if power prices in PJM and ERCOT rise as much as the demand forecasts imply, the incremental value accrues to Microsoft, Amazon and Meta, not to Constellation, Talen and Vistra. The generators have sold the very upside the demand thesis promises. Cameco's Q2 2026 result is the same phenomenon one node upstream: US$67.79 realised against $85.60 spot, because a good contract book is a cap as well as a floor. Discipline and upside are not the same asset, and this sector is currently being valued as though buying one gets you the other. Known Known — the contract terms and realised prices are disclosed

A fourth effect is an open question rather than a conclusion, because the evidence does not yet support a verdict. If the 2028 expiry of the Russian Suspension Agreement quota arrives on schedule with Western capacity still short, the adjustment has to come from somewhere: higher SWU prices, waivers, reactor underfeeding and overfeedingThe dial every enricher can turn. Fuel of a given richness can be made from less uranium and more enrichment effort (underfeeding) or more uranium and less effort (overfeeding). Which way the industry turns the dial depends on the relative price of the two — so a shortage at one node quietly transfers demand, and value, to the other. trade-offs that consume more uranium to save enrichment, or utilities drawing down inventories. Each of those redistributes value between the uranium and enrichment nodes in a different direction. The 2025 SWU price of $108.70 was already up 11% year on year. Which mechanism dominates is genuinely unknown, and it is the single most consequential open question in the fuel cycle. Known Unknown — where to dig next

A verdict without a falsifier is an opinion. Each grade below is paired with the single observable development that would most directly invalidate it.

Each is stated so that it can be scored right or wrong against a specific published source, with a confidence tag sized to the evidence rather than to conviction.

Every source below was opened and read during the preparation of this brief. Tier 1 = issuer filings and regulator or government publications. Tier 2 = reputable trade, wire or research publication. No Tier 3 source carries a conclusion.

Financial data for the return-on-capital table was drawn from each company's XBRL financial data as filed with the SEC and reconciled to the filings listed above. Uranium spot ($85.60/lb, 23 July 2026) and long-term ($93.00/lb, 31 March 2026) price indicators are market-service quotations reported in trade press and are used only as market context — no grade rests on them.

The following could not be confirmed against a primary source opened during this work. None of them carries a conclusion, a grade, or a figure anywhere in this brief.

• Sprott Physical Uranium Trust's current pound holdings and premium or discount to net asset value. The trust's product page was opened but the live holdings and NAV figures did not render in retrievable form. The SRUUF grade rests only on the instrument's structure, which is disclosed and unambiguous, not on its size or its current discount.
• Orano's US enrichment plans. Reporting suggests an Oak Ridge project, but no Orano primary was opened. Orano appears in this brief only through Centrus's 10-K characterisation of its ~8 million SWU French capacity. No conclusion about US supply additions rests on Orano.
• Any Q2 2026 results for Constellation, Vistra, Talen, Centrus, BWXT, NuScale, Oklo or Nano Nuclear. Constellation's Q2 2026 call is scheduled for 6 August 2026, after this brief's evidence cutoff of 1 August 2026. Every figure for these companies is Q1 2026 or FY2025 as filed. Cameco, GE Vernova, Southern, Dominion and Exelon had reported by the cutoff and their Q2 data is used.
• Westinghouse's standalone financial statements. Westinghouse is private. Its economics enter this brief only through Cameco's equity-accounted disclosure and the publicly announced government partnership terms. No margin, return-on-capital or cash-flow claim is made about Westinghouse itself.
• Urenco's and Orano's profitability, contract prices, or order-book economics. Neither is listed and neither publishes at the granularity required. They are used solely to bound the competitive ceiling on capacity and timing, from Urenco's own published announcement and Centrus's 10-K.
• The precise per-kilowatt cost basis comparability between Darlington and Vogtle. The Darlington budget is in 2024 Canadian dollars and explicitly includes interest and contingency; the Vogtle figure is Georgia Power's net investment in US dollars excluding ~$440m of capitalised AFUDC. The comparison is directionally sound and is presented only as such — the gap is large enough to survive the adjustment, but the two numbers are not on an identical basis and no conclusion turns on their exact ratio.
• TerraPower's own financial disclosure. TerraPower remains private and enters through the NRC's construction-permit release and a clearly tiered secondary cost estimate. No grade or return claim is made about it. X-energy is now public and is graded separately from its opened Q1 2026 filing.
• NuScale's and Oklo's forward order pipelines beyond what their own filings state. ENTRA1's arrangements with TVA and Oklo's customer letters of intent are characterised only as their filings characterise them — planning, phases and agreements, not orders.
• Current market capitalisations and valuation multiples for any company here. This brief takes no view on price and quotes none. The scorecard therefore measures qualitative story burden—dependence on undelivered reactors, uncontracted power or commodity upside—not what a security's market price embeds.

This is an internal structural assessment of industry economics and competitive position. It is not investment advice, not a recommendation, and not a solicitation. It contains no price targets, entry or exit levels, or position sizing. Evidence cutoff: 1 August 2026.