The AI energy problem is real and getting worse

Microsoft is restarting a reactor at Three Mile Island to power AI data centers.

Read that sentence again. The most infamous nuclear accident in American history happened at Three Mile Island in 1979. For forty-five years, the name meant fear, failure, and radiation. Now it means electricity for AI.

That’s how desperate the energy situation is.

The numbers

A single large AI data center consumes 100-300 megawatts of electricity. That’s equivalent to powering 80,000-240,000 homes. A small city.

Goldman Sachs estimates that data center power consumption will increase 160% by 2030. The International Energy Agency projects that AI alone could consume as much electricity as the entire country of Japan by 2030.

NVIDIA’s latest GPUs are more efficient per computation, but the total compute demanded is growing faster than the efficiency gains. The industry is on a treadmill where the hardware gets better and the appetite grows faster.

The nuclear pivot

Three Mile Island isn’t an outlier. Constellation Energy is restarting the reactor specifically for Microsoft. Amazon is buying a nuclear-powered data center campus. Google signed agreements with nuclear operators. Every major AI company is looking at nuclear power.

The reason is simple. Solar and wind are intermittent. AI data centers need power 24/7. Nuclear is the only zero-carbon source that runs around the clock at the scale these facilities require.

The irony is thick. The technology industry, which spent two decades championing renewable energy, is turning to the energy source that environmentalists have fought for fifty years. Because the math demands it.

Why this concerns me

I write about the future with wonder. I look at AI models and humanoid robots and reusable rockets and I feel awe. But the energy cost of that future deserves honest accounting.

Every query to an AI model uses electricity. Every training run consumes megawatts for weeks. Every data center needs cooling, which needs more electricity. The physical cost of digital intelligence is real and growing.

We’re building a future of intelligence that requires a future of energy. And the energy future isn’t built yet. We’re patching it together with restarted nuclear plants and emergency grid upgrades and data centers located next to power sources that were built for other purposes.

The SemiAnalysis estimates I’ve read suggest the AI industry’s energy consumption could reach 4-5% of US electricity generation by 2030. From nearly zero percent in 2020.

That’s a problem. Not an existential one (we can build more power generation). But a real one that requires real investment, real planning, and real tradeoffs. Energy has constraints that software doesn’t. You can’t scale a power plant with a git push.

The optimistic take

Nuclear power is being rehabilitated by AI demand. That might be a good thing independent of AI. If the energy needs of data centers drive investment in next-generation nuclear (small modular reactors, advanced fuel cycles), the entire grid benefits.

AI’s hunger for electricity might be the catalyst that finally gets us past the nuclear stigma that’s blocked clean baseload power for decades. An unexpected side effect of training language models: viable nuclear energy policy.

I’m watching this tension between AI progress and energy reality. It’s one of the most important dynamics in technology right now, and it doesn’t get nearly enough attention relative to benchmark scores and product launches.

The future of intelligence and the future of energy are now the same conversation. I hope we’re ready for that conversation. I’m not sure we are.

Related thinking: