For years, the limiting factor in AI was chips: who could get enough GPUs. In 2026 a different constraint is taking over. The question is no longer just can we build it? but can we power it?
The demand curve
According to the International Energy Agency (IEA), data centres worldwide used around 415 terawatt-hours (TWh) of electricity in 2024. By 2030 the IEA projects roughly 945 TWh, more than double, with AI as the main driver.
To put that in perspective, 945 TWh is more than many large industrial countries use in a year.
Hitting the grid wall
Building an AI data centre takes a year or two. Getting a big new grid connection can take much longer. In several regions, new data-centre projects are waiting years in connection queues, and utilities are warning about strain on local supply.
The result: data-centre location now follows power, not just land, tax breaks or fibre. Companies are building next to power plants, signing deals directly with generators, and even building their own generation on site.
The nuclear bet
AI data centres need something renewables alone struggle to provide: huge amounts of steady, 24/7, low-carbon power. That’s why big tech has turned to nuclear.
- Microsoft, Google, Amazon and Meta have all signed nuclear agreements, adding up to roughly 10 gigawatts of committed capacity across more than a dozen deals.
- Microsoft signed a 20-year agreement with Constellation Energy to restart Unit 1 at Three Mile Island (now the Crane Clean Energy Center), an 835 MW reactor expected back online around 2027.
- Small modular reactors (SMRs) are moving from paper to hardware. In July 2026 several advanced reactor companies, including Aalo, achieved criticality in test reactors under the US Department of Energy’s pilot programme.
Reality check: signed isn’t switched on
Here’s the part the press releases skip. Industry trackers estimate that only around a fifth of the nuclear capacity big tech has committed to is actually delivering power today. Restarts take years; new reactors take longer; SMRs at commercial scale are largely a late-2020s to 2030s story.
Critics, including writers at the Bulletin of the Atomic Scientists, warn against treating next-generation nuclear as a near-term fix. In the meantime, much of the new demand is being met by gas.
Efficiency is the quiet hero
While the big power deals grab headlines, a lot of progress comes from using less:
- Liquid cooling is replacing air cooling in AI data centres, cutting energy used for cooling.
- Smaller, specialised models can do many jobs as well as giant general models, at a fraction of the energy.
- On-device AI (see our NPU post) moves work off the grid and onto efficient laptop chips.
- Smarter scheduling runs flexible workloads when and where clean power is plentiful.
Why this matters to you
AI’s energy appetite will shape electricity prices, climate targets and where the next wave of tech jobs lands. It’s also a reminder that “the cloud” is a very physical thing: concrete, cables, cooling and power stations.
Next in the series: The Password Is Finally Dying: Passkeys Explained
Sources: IEA via TechTarget: Powering AI’s future · IAEA Bulletin: Data centres eye advanced nuclear · Forbes: The AI boom is making nuclear bankable again · Bulletin of the Atomic Scientists: Don’t fall for the hype · Axis Intelligence: Nuclear energy for data centers 2026
