The AI race is really a power race
The chips are only half the story. A chip without electricity is a paperweight, and America can't build power plants, transformers and power lines as fast as tech companies want to build data centers. China can.
Plain-English explainer by Joe Miller, a nuclear engineer. Not investment advice. Figures are the latest public numbers we could find and change fast; sources are at the bottom of each page.
How much power AI needs
U.S. data centers used about 4.4% of the country's electricity in 2023. The Energy Department's Lawrence Berkeley National Laboratory projects 6.7% to 12% by 2028, meaning their use could double or triple in five years. A single large AI campus now draws 1 to 5 gigawatts, as much as one to five nuclear reactors or a city of a million homes.
China vs. America
- China ended 2025 with about 3,890 gigawatts of power plants, up 16% in one year, including 1,200 GW of solar and 640 GW of wind. It is also building 41 nuclear reactors (see China builds power plants fast, members).
- China makes about twice as much electricity as the U.S. and has spare capacity in many regions, so a new data center there can plug in quickly.
- In the U.S., many regions are short: PJM, the grid from Chicago to Virginia, saw its capacity prices jump more than tenfold in two years (see your electric bill).
The bottlenecks
| What's short | How bad | Why |
|---|---|---|
| Gas turbines | GE Vernova had a 116 GW backlog by mid-2026; slots are sold out for years, into the 2030s | Only three big makers (GE Vernova, Siemens Energy, Mitsubishi) |
| Big transformers | 3 to 5 years for the largest; U.S. average about 128 weeks | Few U.S. factories; demand for some types up 274% since 2019 |
| Grid connections | Waits of 4 years or more in many markets; Texas had 474 GW of large-load requests when it paused approvals in August 2026 | Utilities must study every request and build lines and substations |
| Power lines | A major interstate line can take 10 years or more to permit and build | Many states and landowners, lawsuits |
| New nuclear | Restarts 2 to 4 years; new reactors 2030s | Licensing, supply chain, cost |
The workarounds
- Bring your own power: building gas turbines or fuel cells next to the data center, like xAI in Memphis (see Elon Musk and Colossus).
- Buy existing plants' output: tech companies signing deals with running nuclear plants and restarting closed ones.
- Flexible data centers: agreeing to cut power during grid emergencies in exchange for faster hookups.
- Going where the power is: West Texas, the Dakotas, Louisiana and even overseas. See where America's AI is being built.
What to expect
- Electricity, not chips, will be the main limit on U.S. AI for the rest of this decade.
- More gas in the short run, because it's what can be built fastest, then more nuclear in the 2030s.
- Higher bills and more local fights unless data centers pay their full share of grid upgrades.
Sources
Checked November 16, 2026.
- Lawrence Berkeley National Laboratory: 2024 U.S. Data Center Energy Usage Report (Dec. 2024)
- CarbonCredits.com: China adds power 8x more than the U.S. in 2025 (Feb. 4, 2026)
- Turbomachinery: GE Vernova gas turbine backlog hits 116 GW (July 23, 2026)
- The Next Web: U.S. power firms scramble for transformers and turbines (July 9, 2026)
- LLM Pulse: U.S. data center moratoriums and blocked projects (2026)