The International Energy Agency says conditional off-take agreements between data-center operators and small modular reactor projects have climbed to 45 GW, up from 25 GW at the end of 2024. That is not built capacity. It is the part of the market where AI power demand stops being a forecast and starts looking like a procurement problem.
Key facts
- Conditional off-take agreements between data-center operators and SMR projects grew from 25 GW at the end of 2024 to 45 GW today.
- Capital expenditure by five large technology companies topped more than $400 billion in 2025, and the IEA says it is set to rise another 75% in 2026.
- Electricity demand from data centers rose 17% in 2025, while global electricity demand grew 3%.
- The IEA says electricity consumption from data centers is set to double by 2030, and power use from AI-focused data centers is poised to triple.
- Tech companies accounted for around 40% of corporate renewable power purchase agreements signed in 2025.
What happened
The IEA's new report, Key Questions on Energy and AI, says the AI buildout has moved from a power planning problem to a market-formation problem. Five large technology companies spent more than $400 billion on capital expenditure in 2025, and the agency says that figure is set to rise another 75% in 2026. At the same time, electricity demand from data centers rose 17% last year, more than five times the pace of global electricity demand, which grew 3%. The point is not subtle: AI demand is rising faster than the grid around it, and the money is already moving.
The most interesting line in the report is not the one about demand growth. It is the one about contracts. Conditional off-take agreements between data-center operators and SMR projects have grown from 25 GW at the end of 2024 to 45 GW today, according to the IEA. That is a sharp move in less than a year. It does not mean 45 GW is under construction or fully financed. It does mean the conversation has shifted from "could nuclear help?" to "which project, on which site, under what terms?"
The agency also says power consumption per AI task is falling fast, with efficiency improving at a pace it calls unprecedented in energy history. That is not the end of the story, though. More people are using AI, AI agents are becoming more energy-intensive, and the overall demand curve still points up. The report's forecast says electricity use from data centers should double by 2030, and power use from AI-focused data centers should triple.
There is another detail worth keeping in view. Tech companies accounted for around 40% of all corporate renewable power purchase agreements signed in 2025, according to the IEA. In other words, the same buyers who have already spent years locking in solar and wind are now widening the menu. Nuclear is not replacing renewables in these deals. It is being layered on top of them, usually because buyers want firm power at a scale and clock speed that intermittent generation alone cannot always provide.
Why it matters
This is where the IEA report stops being a macro story and starts looking like the early shape of a market. The data-center boom is not only pulling on transmission, transformers, turbines, and grid interconnections. It is also pulling nuclear back into procurement conversations that, a few years ago, would have sounded speculative.
NNN has already been tracking that shift from the vendor side. Our coverage of GridMarket and Deployable Energy's data-center nuclear pipeline showed how suppliers are trying to package nuclear as a buyable product, not just a design. Our SMR explainer lays out why the serial-build case matters so much. And our license-path analysis makes the same point from the regulatory angle: execution is the product now.
The IEA report gives that trend a harder edge. It suggests that the AI boom is not just increasing electricity demand. It is also reshaping the kinds of contracts buyers are willing to sign. Once a data-center operator starts treating firm power as a strategic input rather than a utility bill, nuclear starts to look less exotic. Maybe not easy. Not cheap, certainly. But legible.
That matters because SMRs live or die on repeatable demand. One-off enthusiasm does not build a fleet. Off-take agreements do. If 45 GW of conditional deals are real, then the market is no longer hunting for a story. It is hunting for projects that can survive siting, licensing, interconnection, and financing long enough to turn paper into megawatts.
Background
The IEA's report is careful not to oversell the efficiency side of AI. It says per-task power use is falling, but also that usage is spreading and new use cases, especially agents, are power-hungry. That is the tension running through the whole report. You can make each prompt cheaper and still burn more electricity overall if there are enough prompts.
The grid side is just as messy. The IEA says supply chains for gas turbines, transformers, advanced chips, and IT components have tightened over the past year, while planning and regulatory systems are struggling to keep up with the project pipeline. That is why so many data-center developers are turning to onsite generation, battery storage, and long-term off-take structures. They are trying to buy certainty, not just electrons.
Nuclear fits that logic because it promises firm output over long horizons. But it also brings its own queue of problems: siting, licensing, construction risk, and the simple fact that every project has to work in the real world, not the slide deck. The reason the data-center story matters is that it gives those projects a customer base with unusually strong incentives to move. Hyperscalers do not need a philosophy of energy. They need power that shows up on time.
That is why the IEA's 45 GW figure lands. It suggests the market is starting to treat firm low-carbon power as a procurement category. The exact mix will vary by region and by company. Some buyers will stick with renewables and storage. Some will lean on gas. Some will try nuclear. But the direction of travel is hard to miss: the people who need the most power are no longer content to wait for the grid to sort itself out.
What's next
Watch for actual signed PPAs, site announcements, and financing structures that turn the IEA's conditional pipeline into named projects. Also watch the agency's promised new government-industry platform on energy and AI; if it produces better data, the conversation around data-center power will get sharper fast.
For now, the useful takeaway is simple. AI is still a load problem. It is also becoming a procurement problem. And once power demand starts showing up in contracts, nuclear stops being a background option and starts competing for real business.
Questions
- Is 45 GW of off-take the same as 45 GW under construction?
- No. It is a conditional pipeline, not a build total. It shows that data-center buyers and SMR vendors are negotiating around future supply, but it does not mean the projects are financed or started.
- Why are data centers showing up in a nuclear story?
- Because AI workloads need large, steady power loads. The IEA says data-center electricity demand rose 17% in 2025 and is set to keep climbing, which pushes buyers toward firm power options.
- What does this mean for SMRs?
- It suggests SMRs are moving from concept to procurement. The bigger the off-take pipeline, the easier it is for developers to argue that a project has a real customer behind it.
Sources
- Data centre electricity use surged in 2025, even with tightening bottlenecks driving a scramble for solutions — International Energy Agency
- Key Questions on Energy and AI — International Energy Agency
About Nuclear News Network
Nuclear News Network (NNN) is an independent publication covering the global nuclear energy sector — reactor construction, SMRs, fuel supply, policy, operations and fusion. NNN publishes a daily brief, same-day analysis of major developments, and reference guides used across the industry. Articles are produced by the NNN Newsroom, an editorial automation system with human oversight, under the publication's editorial standards.