The race to build more artificial intelligence capacity is no longer solely dependent on acquiring GPUs. Electricity, networks, and data centers are becoming an equally important part of the equation, and consumption forecasts explain why Microsoft, Google, and Meta have begun signing long-term nuclear agreements. The International Energy Agency (IEA) estimates that data centers consumed around 485 TWh in 2025 and could approach 950 TWh by 2030.
The essentials of nuclear for data centers in 30 seconds
- The IEA predicts that global electricity consumption for data centers will nearly double between 2025 and 2030.
- Microsoft has supported the recovery of 835 MW of nuclear capacity in Pennsylvania through a 20-year power purchase agreement.
- Google is collaborating with Kairos Power on advanced reactors and considers up to 500 MW by 2035.
- Meta has announced agreements backing up to 6.6 GW of existing and future nuclear capacity.
- Nuclear provides continuous generation, but new reactors still face uncertainties regarding costs, timelines, and large-scale buildability.
The leap in generative AI is also introducing a challenge that wasn’t as pronounced during previous cloud expansion cycles: each new generation of accelerators enables much more electrical power to be concentrated in less space.
Building a building filled with servers is not enough. Hundreds of megawatts must be delivered to it, substations and transformers must be in place, equipment cooled, and a sufficiently stable electrical supply maintained for years.
That’s where nuclear energy again becomes relevant.
Gigawatts start to matter just as much as GPUs
AI infrastructure is becoming an extraordinarily power-intensive industry.
The IEA estimates that data center energy consumption grew by 17% during 2025 and expects rapid growth to continue through the decade. Facilities specifically dedicated to AI are projected to grow even faster.
The concentration of consumption is especially significant. A traditional factory can distribute activity across multiple locations. An AI training cluster needs thousands of accelerators connected via very high-speed networks working in coordination.
That also concentrates the electrical demand.
The IEA calculates that in 2027, an advanced rack could reach a maximum demand roughly equivalent to 65 homes. Multiplied by thousands of racks, this helps explain why energy planning is increasingly influencing where to build the next generation of data centers.
The limitation may shift from semiconductors to energy availability. A company can purchase thousands of GPUs, but those GPUs hold little utility while waiting several years to secure enough electrical connection.
Nuclear offers something difficult to achieve solely through variable sources: continuous and predictable generation for decades.
That doesn’t mean replacing renewables. The IEA expects about half of the global increase in electricity for data centers through 2030 to come from renewables. AI infrastructure is trending toward a mix of renewables, nuclear, storage, gas, and enhanced grid systems.
Microsoft recovers 835 MW and Google invests in new reactors
One of the most notable moves comes from Microsoft.
Constellation announced in 2024 a 20-year electricity purchase agreement supporting the recovery of the old Three Mile Island Unit 1, now called Crane Clean Energy Center.
The facility could return approximately 835 MW of capacity to the grid.
There’s an important distinction here: it’s not the reactor involved in the 1979 nuclear accident at Three Mile Island. That was Unit 2. Unit 1 operated independently for decades and closed in 2019 due to economic reasons.
This agreement illustrates one consequence of data center growth: energy assets that are no longer sufficiently profitable can regain value if a buyer is willing to commit for decades.
Google is pursuing a different approach.
Its agreement with Kairos Power doesn’t involve recovering large traditional reactors but supports deploying advanced, smaller reactors.
The first planned project is Hermes 2 in Oak Ridge, Tennessee. The agreement between Kairos Power and the Tennessee Valley Authority proposes providing 50 MW to the grid starting in 2030. This electricity will feed Google’s data centers in Tennessee and Alabama.
The relationship is more ambitious, aiming to deploy up to 500 MW of capacity by 2035.
The technological goal of the project is to see whether a new generation of reactors can avoid some of the primary economic issues historically associated with large nuclear plants.
Small modular reactors (SMRs) seek to use standardized designs and repeatable components, which should, on paper, reduce costs and construction times. However, commercial large-scale proof remains to be demonstrated.
Meta raises nuclear commitment to 6.6 GW
Meta has gone even further.
In January 2026, they announced agreements with Vistra, TerraPower, and Oklo that, along with their previous deal with Constellation, support up to 6.6 GW of new and existing nuclear capacity by 2035.
This figure requires context.
It doesn’t mean Meta will immediately have 6.6 GW from new reactors. Part of it comes from operational plants, and other parts depend on projects that will need regulatory approval, financing, construction, and startup.
TerraPower initially plans for two Natrium units with up to 690 MW, with the possibility to build six more later.
Oklo aims to develop a nuclear campus in Ohio with up to 1.2 GW.
These projects are particularly relevant for the data center industry because they suggest a different relationship between power generation and large consumers.
Traditionally, a power plant supplies electricity to a grid serving a variety of customers. The growth of AI is creating individual buyers who may need amounts of power comparable to small towns.
And, most importantly, buyers capable of signing long-term contracts.
For a nuclear developer, knowing in advance who will buy the electricity for 20 years can significantly influence project financing.
AI could reshape reactor economics
Nuclear energy has faced a paradox for decades.
It contains vast amounts of energy, reactors can operate for long periods, and their production has low direct carbon emissions. But building new reactors in many Western markets has become a costly and lengthy process.
An analysis by Tomas Pueyo, Why Nuclear Is the Best Energy, compiles many arguments used to advocate for this technology—from energy density and fuel availability to land use, waste, and safety compared to other sources.
The article takes a pro-nuclear stance, and some interpretations of radiation, historical accidents, waste, and regulation are debatable. However, it raises a key question that’s especially relevant given the expansion of AI: to what extent are nuclear’s economic problems due to physical technological limitations versus how plants are built and regulated.
Upcoming advanced reactor projects will help determine this.
If SMRs can repeat designs, manufacture components industrially, and cut construction times, they might find their first significant markets in data centers.
Conversely, if delays and cost overruns similar to recent large nuclear projects occur, the tech industry will need to seek large capacities from other sources.
From GPU shortages to megawatt shortages
In the early years of generative AI’s explosion, attention focused on NVIDIA and GPU availability.
The next phase is significantly amplifying the problem.
An AI cluster requires accelerators, HBM memory, high-speed networks, storage, cooling, and an electrical infrastructure capable of delivering enormous energy loads. It also needs transformers, substations, and grid connections not originally designed for hundreds of megawatts in data centers.
That’s why companies whose core business isn’t electricity generation are now directly involved in energy supply.
They’re not only building data centers; they’re trying to ensure the energy that allows those data centers to operate.
The IEA forecasts that global data center energy consumption could reach 950 TWh by 2030, roughly double what it was five years prior. If this projection holds true, energy infrastructure will increasingly determine where new AI capacity can be deployed.
And this could be the most significant shift of all.
For years, data center location depended on connectivity, latency, taxes, land availability, and proximity to users. The availability of hundreds of megawatts is now rising high on that list.
In short, the AI race is no longer confined to semiconductor factories and model labs—it’s moving into power plants, transmission lines, and substations.
Nuclear has returned to the radar of tech giants precisely because it can supply one of the resources that the next AI generation will need at enormous scale: electricity available 24/7 for decades.
Frequently Asked Questions
Why do AI data centers require so much electricity?
Advanced models use thousands of accelerators running simultaneously. Added to this are memory, networks, storage, electrical systems, and cooling, substantially increasing the total power needs of new data centers.
Is Microsoft reopening Three Mile Island?
Microsoft’s agreement with Constellation supports the financial recovery of the old Three Mile Island Unit 1, now called Crane Clean Energy Center. It is not related to the Unit 2 reactor involved in the 1979 accident.
Will Google use small nuclear reactors for its data centers?
Google maintains a partnership with Kairos Power to support deploying advanced reactors. The program aims for up to 500 MW by 2035, although that capacity still needs to be built and brought online.
Will nuclear energy replace renewables in powering AI?
Forecasts suggest a complementary approach. The IEA expects renewables to supply a significant part of the new electricity for data centers, while nuclear, gas, storage, and upgraded grids will serve other needs depending on each market.
