AI Has Become "Power Heavyweight"
AI models up to 2022 could be trained on hundreds of GPUs, but from GPT-4 class onward, the scale is thousands to tens of thousands of GPUs running 100 days continuously. One training run uses power on the level of one small Japanese city.
Numbers for Scale
| Item | Scale |
|---|---|
| GPT-4-class training power | ~50 GWh (equiv. monthly use of 100k households) |
| Frontier cluster (2026) | 100MW-1GW (equiv. 0.5-1 nuclear reactor) |
| 1 ChatGPT query | ~3 Wh (5-10x a Google search) |
| 2026 world AI power | 2-3% of all world power forecast (IEA) |
| 2030 forecast | 5-10% of world power (high uncertainty) |
Location Competition
US South/Midwest
Texas, Georgia, Virginia, Ohio growing fast. Cheap power (natural gas, nuclear) and vast land are attractions.
Nordics
Iceland, Sweden, Finland, Norway. Strength in geothermal/hydro and cooling efficiency from cold climate.
Middle East
UAE, Saudi Arabia putting oil money into AI infrastructure. Large-scale investment like the Stargate concept (OpenAI + G42).
Japan
Power unit price world's highest (25-35 yen/kWh) is disadvantageous for large AI clusters. Meanwhile, expectations for land with low earthquake risk, abundant water, cooling suitability (Hokkaido).
Power-Supply Measures
Nuclear Restart/New Builds
- Microsoft's Three Mile Island restart contract (2024.9)
- Amazon's nuclear-adjacent data-center contract
- Google's investment in an SMR startup
- Japan also: Kashiwazaki-Kariwa restart, SMR consideration
Renewable Self-Generation
Lower power unit price with solar/wind + batteries. The problem is variability; hard to use directly for full AI workloads but effective as a complement.
Water-Resource Problem
From the Blackwell generation, liquid cooling becomes mandatory. Water use for data-center cooling becomes a regional issue. In US Arizona and Texas, water disputes have surfaced, and resident lawsuits have occurred.
Grid Constraints
Even if you want to build a new data center, cases of years-long waits because power lines can't be laid frequently occur. The US Federal Energy Regulatory Commission (FERC) upgraded AI-grid development to a top-priority issue in 2025.
Moves on the Solution Side
- Model efficiency: MoE, quantization, distillation for equal performance at 1/10 power
- Dedicated ASICs to improve power efficiency
- Edge inference (Apple FoundationModels, on-device LLMs)
- Distributed training to avoid local power peaks
Summary
AI scaling is no longer computer science but power engineering. Building 100MW-1GW-class clusters is directly hit by physical constraints of power, water, transmission lines, and land, making location competition and political judgment key elements of AI competition.
