AI Energy Strategy
Prerequisites
Microsoft has signed a deal to restart Three Mile Island. Google is funding geothermal startups. Amazon bought a nuclear-powered campus. These are not PR moves. They are the logical conclusion of a physical constraint.
Data centers run continuously. A training cluster does not pause at night because the wind stopped. The term for power that runs around the clock regardless of weather is firm power. Coal and gas are firm. Wind and solar, without storage, are not.
Tech companies have committed to clean energy, but matching renewable purchases on an annual average does not mean running on clean power every hour. Hourly matching requires clean firm power. Nuclear and geothermal are the only low-carbon sources that are both firm and carbon-free, which is why hyperscalers are suddenly interested in them.
A 500 MW data center at 90% utilization needs roughly 450 MW almost every hour. A 500 MW solar farm averages about 25% of its rated output across the year. To cover the same annual energy: 2,000 MW of solar capacity, plus storage for nights and cloudy stretches, plus backup for multi-day droughts.
At what scale does battery storage become impractical for covering a week-long wind drought?
A week of storage for 500 MW requires roughly 84,000 MWh. At $334/kWh installed (NREL's 2024 benchmark), that is about $28 billion in batteries for one facility. A 1,000 MW nuclear plant running at 90%+ capacity factor pairs directly with this load, around the clock, for 60+ years, with almost no storage required.
The hyperscaler push into nuclear and geothermal is a response to firm-power arithmetic, not exotic technology bets.
Why can't large data centers simply buy enough solar and wind to claim 100% clean energy on an hourly basis?
The issue is not generation cost but the astronomical cost of storing enough energy to cover multi-day wind and solar droughts.
The answer is DLesson complete
Back to Curriculum