Nuclear for Data Centers Runs on Two Clocks

Hyperscalers have signed roughly 8.5 GW from existing and restarting U.S. reactors. Conditional SMR deals top 45 GW. Only one data-center-tied new reactor is under construction. Do not confuse the two clocks.

Every few weeks another headline says AI will be powered by small modular reactors. Some of those headlines are about real contracts. Most are about options on machines that do not exist yet. Sorting one from the other is the only way to use nuclear in a power plan that has to hit 2027 and 2028.

A detailed public tally of signed deals puts U.S. hyperscaler contracts with operating, restarting and uprated reactors at roughly 8.5 GW. The International Energy Agency counts about 45 GW of conditional data-center agreements with SMR projects. Of new reactors tied to data centers, only Kairos Power’s 50 MW Hermes 2 plant in Oak Ridge — which will feed the TVA grid serving Google load from 2030 — is under construction (Nuclear News Network).

Those three numbers should never appear in the same sentence without a label.

Clock one: existing fleet, restarts, uprates

This is the clock that can move this decade. Amazon’s Susquehanna deal with Talen runs up to 1,920 MW through 2042 and is already delivering under a front-of-the-meter structure after FERC blocked a larger behind-the-meter arrangement in 2024. Microsoft’s 20-year PPA with Constellation covers the 835 MW Crane Clean Energy Center restart — formerly Three Mile Island Unit 1 — targeted for 2027. Meta’s Clinton deal with Constellation starts deliveries in June 2027. Amazon’s Comanche Peak PPA with Vistra targets Q4 2027. Google is backing the Duane Arnold restart in Iowa toward Q1 2029 after a $1.9 billion DOE loan closed in September 2026 (Nuclear News Network).

Restarts are not free options. Reuters reported that delayed transmission projects were blocking a speedy Crane restart, and that full transmission upgrades could run past 2030 even if FERC waivers help with interconnection rights. Constellation still targets 2027 for returning megawatts to the PJM grid under the Microsoft PPA (Reuters; Reuters).

Uprates matter for additionality. Meta’s Vistra package and Amazon’s Calvert Cliffs deal with Constellation bundle uprates into the contract so the buyer is not only reallocating existing clean megawatts. A plain PPA on an operating plant moves accounting. An uprate or restart adds capacity. Underwriters should ask which one they bought.

Clock two: SMRs and first-of-a-kind new build

This clock mostly rings in the 2030s. Google’s Kairos framework covers up to 500 MW by 2035, with Hermes 2 as the pathfinder. Meta’s TerraPower and Oklo agreements support a large portfolio by 2035 on paper. Amazon’s X-energy partnership talks about more than 5 GW of new nuclear by 2039. Switch’s Oklo master agreement floated 12 GW by 2044 and is explicitly non-binding. The IEA’s 45 GW of conditional SMR offtake is not a construction tally (Nuclear News Network).

Co-location looked like a shortcut until regulators intervened. FERC rejected an expanded behind-the-meter hookup at Susquehanna in November 2024. Talen and Amazon moved to front-of-the-meter. FERC later ordered PJM to create transmission services for co-located load so campuses that sit next to reactors still pay for the grid services they use. The cost advantage of plugging in next door narrowed. The permitting and emergency-planning overlay of a reactor on a data-center campus never went away.

What operators are doing while they wait

They are buying gas, fuel cells and credit. Google’s Wyoming campus pairs a 702 MW Bloom fuel-cell plant and CCS with Tallgrass and Black Hills gas generation — a firm-power bridge while nuclear options mature (Google). That is not hypocrisy. It is calendar math. A data hall can be built in a few years. A first-of-a-kind reactor takes most of a decade. Blue Energy and Crusoe have even described gas-first, nuclear-later sequencing for a Texas project for that reason.

Gallup data cited in the nuclear tally found Americans more opposed to a data center in their area (71%) than to a nuclear plant (53%). The political problem is often the campus, not the reactor. Putting both on one fence line stacks two hard sells.

How to put nuclear in a 2027–2030 model

For capacity that must energize before 2030, restrict the nuclear line items to signed PPAs on operating plants, funded restarts with NRC milestones in hand, and contracted uprates with owner capex committed. Everything else — SMR frameworks, non-binding master agreements, MOUs to “explore” — belongs in a separate optionality schedule with probability weights, not in the base-case MW stack.

For carbon accounting, ask whether the deal adds electrons (restart, uprate, new build) or reallocates existing ones. For grid risk, ask whether delivery is front-of-meter or co-located, and what transmission services the campus must buy after FERC’s PJM orders. For political risk, remember Gallup’s finding that the data center is often less popular next door than the reactor. Stacking both on one site doubles the hearing calendar even when the engineering is clean.

What to watch

Crane’s 2027 return to service and whether transmission limits trim delivered MW.
First deliveries under Meta–Clinton and Amazon–Comanche Peak in 2027.
Hermes 2 construction milestones toward 2030.
Site announcements that turn Meta–TerraPower and Amazon–X-energy frameworks into licensed projects.
Any revival of co-location deals under PJM’s new transmission services.

Why it matters for builders, buyers and investors

If your 2028 power stack depends on SMRs, you do not have a power stack. You have a press release. Use existing-fleet PPAs, restarts and uprates for near-term firm carbon-free supply. Treat SMR frameworks as optionality with license and FOAK cost risk. For $CEG, $VST and peers, the equity story that clears this decade is the operating fleet and restarts; the SMR story is a longer-duration call option.

Not investment advice.

Sources

Report a problem with this page ↗