How Reactor Restarts, Uprates and Life Extensions Affect Uranium Demand
60-second answer: Nuclear growth forecasts lump together three mechanisms that do different things to uranium demand. A restart returns a shut reactor to service, restoring demand that had left the market: Palisades (800 MW) and Crane (835 MW) together represent roughly 1.6 GW of it. An uprate raises an operating reactor's licensed power, adding genuinely new demand, usually a few percent per plant. A life extension keeps an operating reactor running past a retirement date, preserving demand against a baseline where it would have vanished. And a power contract for an already-operating plant, the most common headline of the AI era, does none of the three by itself: it changes who buys the electricity, not how much uranium the reactor consumes. Counting each mechanism once, against the right baseline, is most of the work in an honest demand model.
The stakes are concrete: our nuclear-for-AI deal book tracks 33 GW of announced deals, and classifying each megawatt by mechanism is the difference between 14.7 and 12.4 million pounds a year of modeled demand.
Three mechanisms, three baselines
| Mechanism | What changes physically | Demand effect | Against what baseline |
|---|---|---|---|
| Restart | A shut unit returns to service | Restores its full fuel consumption | The shut state (zero) |
| Uprate | An operating unit's licensed power rises | Adds the increment only | The unit's prior licensed power |
| Life extension | An operating unit keeps running past a planned retirement | Preserves existing consumption | The retirement scenario |
| PPA with an operating plant | Nothing; output is reallocated to a new buyer | None, unless it funds one of the above | n/a |
The baseline column is where models go wrong. A restart is incremental against today's fleet; a life extension is incremental only against a forecast that assumed the retirement; an uprate bundled inside a PPA adds its megawatts and nothing more.
Case study: Palisades, a restart in progress
Palisades in Michigan, shut in 2022, is the first US attempt to restart a decommissioning-track reactor. The sequence so far, from the NRC's plant record and dated reporting: the NRC moved the plant back to an operating license in 2025, fuel arrived in October 2025 (ANS), the original October 2025 restart target slipped to early 2026 and then further as repairs stretched (Michigan Public), and as of mid-2026 the fuel is loaded, major projects are closed out, and the reactor has not yet gone critical, with Holtec targeting operation ahead of a March 2027 supply contract (ANS, July 2026).
Two lessons for demand modeling. Restart fuel demand arrives before restart power: the first core was bought, fabricated and loaded while the grid date kept moving. And restart schedules carry engineering risk that announcement dates hide; the demand is real, its timing is soft.
Case study: Crane and Duane Arnold, restarts with buyers
The Crane Clean Energy Center (the former Three Mile Island Unit 1, 835 MW, not to be confused with the TMI-2 unit that suffered the 1979 accident) is under restart construction with Microsoft as a 20-year buyer, targeting 2028; Duane Arnold in Iowa (615 MW) follows with Google, targeting Q1 2029 (NRC's Crane record; both deals are source-linked in the deal book). These pair a restart with a PPA, and the pairing is one demand event, not two: the restart creates the incremental fuel consumption, the PPA decides who buys the output. A tracker that counts "restart MW" and "PPA MW" as separate line items counts the same reactors twice, which is why our deal model carries plant identities and counts each physical unit once.
Uprates: small numbers, genuinely new
An uprate amends the reactor's license to raise its maximum power, and requires NRC approval in one of three classes: measurement-uncertainty recapture (under 2%), stretch (a few percent), and extended uprates that can reach 20% with major equipment replacement (NRC power uprates). The megawatts are small next to restart headlines, and they are the cleanest form of new demand: no new site, no new license class, just more thermal output through an existing core, consuming proportionally more fuel.
Uprates hide inside bigger announcements. Meta's 20-year Vistra agreement covers 2,609 MW, of which 2,176 MW is existing output and 433 MW is uprates across three plants; its Clinton agreement includes a 30 MW uprate inside 1,121 MW. In both, the uprate megawatts are the only new uranium demand. Modeled at our screening intensity of 170 tU per GWe-year, a 100 MW uprate is about 17 tonnes of uranium (roughly 44,000 lb of U₃O₈) a year: real, but a rounding error next to the misread headline capacity (how that coefficient works).
Life extensions: demand that does not disappear
License renewals and subsequent renewals keep the operating fleet operating. Nothing new is consumed relative to today; everything is preserved relative to forecasts that assumed retirements. That makes life extensions a modeling instruction: fix the retirement baseline before speaking of increments. A forecast built on "licenses as of 2020" sees large future demand from extensions; one built on "extensions will keep passing" sees none. Both can be defensible, and mixing them double counts. Our demand model treats the operating fleet as the baseline, which is why extensions and operating-plant PPAs contribute zero incremental pounds there.
Preventing double counts in the AI era
The 2025–2026 deal wave stacked contracts on top of physical events, and the honest accounting rules fall out of everything above:
- One physical unit, one count, however many agreements reference it.
- An existing-plant PPA contributes its uprate share and nothing else.
- A restart plus its PPA is the restart's megawatts, once.
- Gross deal-associated capacity and incremental capacity are different totals, reported separately, which is exactly how the deal book now presents them.
- First-core purchases for restarts land before grid dates, so fuel-demand timing leads power-demand timing (first cores explained).
The wider market context sits in the supply and demand guide, and the deal-level detail in the AI-datacenter guide.
Frequently asked questions
Do reactor restarts increase uranium demand? Yes, against the current fleet: a restarted reactor resumes buying fuel that a shut one did not, and its first core is purchased before it produces anything. Palisades (800 MW) and Crane (835 MW) each represent roughly 0.3–0.4 million pounds of U₃O₈ equivalent a year at screening intensity once operating.
Does a data-center PPA with a nuclear plant add uranium demand? Not by itself. A contract with an already-operating plant reallocates existing output to a new buyer; the reactor burns the same fuel. It adds demand only to the extent it funds a restart, an uprate or new construction, and only those megawatts count.
How much extra uranium does a power uprate need? Roughly in proportion to the power increase. At a screening intensity of 170 tU per GWe-year, a 100 MW uprate consumes about 17 additional tonnes of uranium a year, around 44,000 lb of U₃O₈. NRC-approved uprates range from under 2% to as much as 20% of a plant's power.
What is the difference between Crane Clean Energy Center and Three Mile Island? Crane is the renamed TMI Unit 1, which operated safely until 2019 and is being restarted for Microsoft. The 1979 accident happened at the separate Unit 2, which never operated again and is not part of the restart.
Are license extensions new uranium demand? Only against a baseline that assumed the plant would retire. Extensions preserve existing consumption rather than adding to it, so whether they show up as "new demand" in a forecast depends entirely on the retirement assumptions the forecast started from.
This article is for informational purposes only, not investment advice.