AI Has Become a "Power Heavyweight"
Frontier models are trained on thousands to tens of thousands of GPUs running continuously for weeks to around 100 days. A training run is a single enormous consumption event, and on top of it the electricity for inference — the compute spent every time someone uses the model — accumulates every day.
Numbers for Scale
The figures below have been updated to the current published estimates. All of them carry wide error bars, so read them for order of magnitude rather than precision.
| Item | Scale |
|---|---|
| One training run of a frontier model | Tens of GWh (comparable to the annual use of a few thousand households) |
| Large AI cluster (top projects under construction) | Several hundred MW to 1 GW (approaching a single large power plant) |
| One ChatGPT query | About 0.34 Wh, OpenAI's published average for a short exchange |
| World data-centre electricity (2025) | Roughly 415-485 TWh, about 1.5% of world electricity (IEA) |
| 2030 projection (IEA) | About 945 TWh, roughly 3% of world electricity, with AI-specific demand around tripling |
The widely repeated claim that one ChatGPT query costs ten times a web search compares a 2023 estimate of roughly 3 Wh against a 2009 figure for search, and ignores the efficiency gains since. Treat query energy as a range instead: a short question is cheap, while heavy reasoning can exceed 20 Wh in some research.
Location Competition
US South/Midwest
Texas, Georgia, Virginia, and Ohio are growing fast. Cheap gas and nuclear power plus vast land draw concentrated hyperscaler construction.
Nordics
Iceland, Sweden, Finland, Norway. Low-carbon geothermal and hydro power, and cooling efficiency from the cold climate.
Middle East
The UAE and Saudi Arabia are directing resource money into AI infrastructure, with large projects such as Stargate UAE (OpenAI and G42).
Japan
Commercial electricity runs around 29 yen/kWh (September 2025), among the highest in the world, which is a disadvantage for very large clusters. Hokkaido still attracts interest for training workloads that do not need low latency, thanks to stable ground, abundant water, and a cool climate.
Securing Power
Nuclear: restarts and SMRs
- Three Mile Island (Pennsylvania): Constellation is restarting Unit 1, and Microsoft takes the output under a 20-year, 835 MW power purchase agreement announced in September 2024. The site was renamed the Crane Clean Energy Center, with operation expected in 2027-2028. Microsoft does not operate the plant; it secures the electricity through a long-term PPA.
- Amazon: acquired a data-centre site adjacent to the Susquehanna nuclear plant from Talen Energy for about $650 million, plus an additional power purchase agreement.
- Google: contracted with small modular reactor developer Kairos Power for multiple units totalling about 500 MW toward the 2030s.
- Japan: restarts at Kashiwazaki-Kariwa and SMR studies are under way.
On the nuclear path, the projects that deliver electricity soonest are restarts of existing reactors rather than new builds. New construction and SMRs face long review and build times, so supply starts years out.
Renewable self-generation
Solar and wind paired with batteries lower the unit price, but output varies with weather, so it is hard to point directly at round-the-clock AI training. The realistic role is as a complement to grid, gas, or nuclear supply.
Water
From the Nvidia Blackwell generation onward, per-rack power passed 100 kW and air cooling stopped keeping up, pushing the shift to liquid cooling. The water this consumes creates friction with local communities in dry regions such as Arizona, Texas, and Georgia, leading to opposition and revised siting plans. The US EPA has estimated direct data-centre water use at about 17.4 billion gallons in 2023, rising to 38-73 billion gallons by 2028. Closed-loop cooling, which recirculates water in isolation from outside air, is spreading as a mitigation.
Grid Constraints
Even with land and capital, interconnection queues of several years are common, so a site can be ready while the power lines are not. New generation faces the same queue, and large AI loads end up unable to draw the electricity their plans assumed. Grid reinforcement and faster interconnection processes are now central to AI infrastructure.
Reducing Consumption
- Model efficiency: mixture of experts, quantization, and distillation hold quality while cutting the compute required
- Dedicated inference chips (ASICs) to raise power efficiency beyond general-purpose GPUs
- Edge and on-device inference with small local models, reducing data-centre and network load
- Geographically and temporally distributed training to avoid local power peaks
Summary
The next wall for AI scaling is not clever code but physical infrastructure. Building clusters of several hundred MW to 1 GW runs straight into power, water, transmission, and land at the same time, which makes siting and policy decisions a decisive variable. Announced figures often mix estimates with caps and plans, so hold on to the order of magnitude while separating confirmed facts from projects that are still on paper.
