The largest cloud provider on earth just committed two decades of spend to electrons that, in part, will not exist for another four to six years.
On September 30, 2026, Amazon signed a 20-year agreement with Constellation for 690 megawatts (MW) of power from the Calvert Cliffs nuclear plant in Maryland, including a 190 MW uprate that is not expected to come online until somewhere between 2030 and 2032. Read that date range again. The largest cloud provider on earth just committed two decades of spend to electrons that, in part, will not exist for another four to six years.
That is not an energy story. It is a capacity planning story, and it lands directly on anyone buying GPU hours, signing multi-year cloud commitments, or deciding where their training data should live. Nuclear has gone from a sustainability talking point to the power source infrastructure leaders say they want most, and the gap between what they want and when they can get it is now the single biggest variable in AI infrastructure pricing.
This article breaks down what the reactor deals actually buy, why the survey enthusiasm and the delivery timelines do not line up, and the five decisions cloud buyers should make differently because of it.
The appetite is real and it is broad. In the BCS 2025 industry report, which surveyed more than 3,000 data center professionals across 41 countries, 75% said they support nuclear power as part of the data center energy mix. BCS is a data center consultancy with an interest in a growing build market, so treat the enthusiasm as directional. The independent numbers point the same way: an IEEE Power and Energy Society poll of more than 1,200 power engineers, run between December 2025 and February 2026, found 80% consider small modular reactors (SMRs) important to clean energy goals, and 58% named meeting data center demand as the most significant challenge of the next decade.
The demand side explains why. Deloitte's 2025 AI Infrastructure Survey of 120 US data center and power company executives found 72% rate power and grid capacity as very or extremely challenging, the top barrier on the list. Deloitte projects US AI data center power demand rising from 4 gigawatts (GW) in 2024 to 123 GW by 2035, a thirtyfold increase, and notes that some grid connection requests already face a seven-year wait.
The money has followed. On January 9, 2026, Meta announced agreements with Vistra, TerraPower and Oklo that could finance up to 6.6 GW of new and existing nuclear capacity by 2035. Forbes reported in July 2026 that every major AI company has now signed at least one nuclear deal this year, together approaching 10 GW of capacity.
For three years, the AI infrastructure conversation was about chips. In 2026 the conversation is about megawatts. Uptime Institute's Five Data Center Predictions for 2026 opens with a blunt verdict: developers will not outrun the power shortage. Uptime projects 75 to 125 GW of growth in global data center power demand through 2030, and the timing problem is structural.
According to Uptime's analysis, as reported by The Register in January 2026, a large AI data center can be built in about three years. A large solar farm takes about five. A wind farm or gas turbine site takes around six. A nuclear plant takes at least a decade. The building is the fast part; the power behind it is the slow part.
Nuclear wins the long game because of what AI workloads need: firm, 24/7 power at high capacity factors. Training runs do not pause when the wind drops. That is why hyperscalers keep choosing reactors even though they are the slowest option on the list. They are not buying power for 2026. They are buying certainty for 2032.
Look closely at the major agreements and a pattern appears. Most near-term megawatts come from reactors that already exist. Most new-build megawatts arrive in the 2030s.
Meta's 1,121 MW agreement with Constellation in Illinois starts in June 2027 and keeps an existing plant running. Its Vistra deal covers roughly 2.6 GW from three existing PJM plants, plus uprates. The new-build portion is where the timelines stretch: Oklo's 1.2 GW campus in Pike County, Ohio is targeted to come online as early as 2030 and reach full capacity by 2034, and TerraPower's Natrium units for Meta account for up to 2.8 GW, all of it 2030s capacity. Amazon's Calvert Cliffs deal follows the same shape: 500 MW of existing output now, 190 MW of new uprate capacity later.
| Deal | Capacity | Type | First power |
|---|---|---|---|
| Meta and Constellation (Illinois) | 1,121 MW | Existing | June 2027 |
| Meta and Vistra (PJM) | About 2.6 GW | Existing plus uprates | Phased |
| Meta and Oklo (Pike County, OH) | Up to 1.2 GW | New build | From 2030, full 2034 |
| Meta and TerraPower (Natrium) | Up to 2.8 GW | New build | 2030s |
| Amazon and Constellation (Calvert Cliffs) | 690 MW incl. 190 MW uprate | Existing plus uprate | Uprate 2030 to 2032 |
Here is the detail most coverage skips. Constellation stated that all electricity from Calvert Cliffs will keep flowing into the PJM regional grid as usual, rather than being physically dedicated to Amazon facilities. The agreement is a long-term purchase contract that gives Constellation the revenue certainty to relicense the plant for 20 more years. That is valuable, but it is not a reactor wired to a GPU hall.
This matters to buyers for the same reason we tell readers to ask every neocloud whether it owns its GPUs. When a provider markets "nuclear-powered" capacity, the follow-up question is: do you own the megawatt, or do you own a contract on a grid? The answer determines how exposed that capacity is to the same regional grid congestion as everyone else.
A contrarian read is warranted, and the evidence for it is strong. Deloitte's 2025 analysis estimated that new nuclear capacity could meet only about 10% of the projected increase in data center power demand over the next decade. The Bulletin of the Atomic Scientists argued in July 2026 that nuclear and SMRs will be unable to fill data center needs in the next several years, precisely when AI demand grows fastest. Its arithmetic is simple: an SMR is defined as producing less than 300 MW, while single campuses such as Stargate in Abilene and Meta's Louisiana project are planned at up to 5 GW each.
Execution risk is not theoretical either. Forbes noted in July 2026 that the SMR designs tech companies are funding have not been completed commercially in the US, and that the last large US reactors, Plant Vogtle Units 3 and 4, finished seven years late at more than double their original budget. Even the survey respondents who back nuclear are skeptical about timing: in the BCS 2025 report, 70% said SMRs will not be ready this decade and 60% worried public opinion could slow progress.
Public opinion is a genuine variable. A Utah State University survey of nearly 900 residents in August 2026 found that nuclear power increased the likelihood of opposition to a local data center for 41.4% of respondents. The industry is betting on reactors; the neighbors have not signed off yet.
Our read: nuclear is the right long-term bet and the wrong near-term plan. The gap between those two statements is where cloud pricing will be decided for the rest of the decade.
If power is the constraint and firm power arrives late, three consequences follow for anyone buying infrastructure rather than building it.
GPU capacity will concentrate in regions with existing firm generation and available interconnection, and those regions will price accordingly. Expect regional spreads in GPU hour pricing to persist rather than converge. Tracking those spreads across providers, not just across instance types, becomes a procurement skill.
Twenty-year purchase agreements with relicensing and uprate investment attached are not cheap power. Deloitte's 2025 analysis found residential power prices rose faster than the national average in eight of the nine top data center markets. That cost does not stay on the utility bill. It shows up in colocation rates, in reserved instance pricing and in the unit economics of every GPU hour. This is the same pattern we track on the token side: a cheaper token is not a cheaper bill, and a nuclear-backed region is not a cheaper region.
This is the part almost nobody budgets for. When GPU availability is decided by megawatts, the best capacity in 2027 may sit in a different region, or with a different provider, than your training data does today. Sunny Smith, founder and CTO of Massed Compute, made the point on Episode 7 of the DataStorage.com Podcast: storage has to sit next to the GPUs, and moving a petabyte between data centers is prohibitive. Teams that let their datasets become landlocked in one hyperscaler region will pay egress every time power availability forces them to move. Zero-egress, S3-compatible storage from providers such as Backblaze B2, Wasabi or Cloudflare R2 keeps that move a network decision rather than a billing event.
Practical application, ordered by how much money each one protects.
Hyperscalers are buying reactors for 2032. Your job is to make sure your data, your contracts and your budget are still flexible when that power finally arrives.
Compare AWS, Google Cloud, Azure, and alternatives like Backblaze B2 Discover how much you could save in seconds
