A gigawatt AI data center costs billions a year to run — but how much of that is really the electricity bill?

This explores how much of a large AI data center's budget actually goes to electricity, as opposed to everything else it pays for. The corpus gives a surprising answer: electricity is a fairly small line item. Epoch AI modeled a one-gigawatt AI data center, which is about the size of a large power plant's output. It puts the total annualized cost at around $8.5 billion a year, and energy at only about $0.6 billion of that, roughly 7% What actually drives AI data center costs?. The servers cost about $5 billion a year, or 60%. That figure is mostly the cost of paying back expensive hardware that wears out and goes obsolete within a few years. Put simply, the chips cost far more than the power they use.

Some rough arithmetic helps here. A gigawatt running all year uses about 8.8 terawatt-hours. At $0.6 billion, that works out to something like 7 cents per kilowatt-hour, which is an ordinary industrial electricity price and not a premium one. That same 8.8 TWh is close to 2% of all data center electricity used worldwide in 2025. The International Energy Agency estimates that total at 485 TWh and expects it to nearly double to 950 TWh by 2030, with AI-focused facilities tripling How fast will data center electricity demand actually grow?. So electricity can be cheap for any single operator and still be a very large problem for the grid. Those are two separate questions, and news coverage often mixes them up.

The same pattern shows up when you look at the hardware and at AI companies' budgets. Inside the chips, high-bandwidth memory grew from 52% to 63% of component spending in under two years, driven by supply shortages Is memory becoming the dominant cost in AI chips?. At the company level, compute takes 54–62% of operating expenses at labs like Anthropic, while staff stay under 25% Do compute costs really dominate AI company budgets?. Across all these sources, the money goes to buying and renting silicon, and not to keeping it switched on.

That could change, though. The IEA describes total AI energy use as the result of three forces pulling against each other. The energy needed per task falls sharply every year. Adoption is growing fast. And the work itself is shifting toward energy-hungry tasks like video generation and long reasoning chains How do efficiency gains and new AI capabilities balance each other?. If chip prices fall faster than electricity prices, power's 7% share will grow. Power's share also grows when the limiting factor is getting a grid connection at all, since that matters more than the price per kilowatt-hour.

The corpus has gaps here. It has no material on how electricity prices differ by region, on cooling costs as their own line item, or on how much it costs to get connected to the grid. So the 7% figure comes from one model, not from measurements across many real facilities.