AI Data Centers' Power Use Is Hard to Predict — New Study Shows How to Fix It

data centers are huge electricity consumers, and their power use can swing wildly depending on what they're working on. That makes it hard for grid operators to plan how much backup power to have ready. A new study used real power measurements from a 448-GPU AI cluster to test forecasting methods. They found that a common forecasting tool (quantile LightGBM) only achieved 63.4% coverage for a 90% confidence interval after a change in usage—meaning it missed the actual power use more often than it should. But by adding online calibration that adjusts to changing conditions, coverage improved to 87.3%. So the forecasts became much more reliable.

Even with accurate forecasts, though, how you use them matters. If you just sample independently from the forecast, you get unrealistic ramp patterns (sudden spikes or drops) and end up buying too much reserve power. The researchers showed that by reconstructing the temporal dependence (the way power use evolves over time) from historical errors, they could reduce the mismatch in ramp statistics and lower the actual cost of reserves by 7.4–10.6% across three forecasting tasks. In plain terms, they saved money by making the forecasts not just accurate on average, but realistic in how they unfold over time.

This matters because grid operators need to decide how much reserve power to procure—extra capacity that sits idle but must be paid for. Over-procurement costs money, which can be passed on to consumers. Under-procurement risks blackouts. The study shows that marginal forecast quality (how good each individual prediction is) isn't enough; you also need to consider how predictions hold together over time and how they feed into decisions. As AI data centers proliferate, getting this right could mean more stable electricity grids and lower costs for everyone.