Energy Intelligence flags AI power surge

Electricity demand used to be a pretty boring planning problem. Utilities could look at population, weather, factories, and appliance trends, then make decent guesses years ahead. AI is breaking that rhythm. The new issue isn’t just that data centers use a lot of power. It’s that they show up in giant lumps, move faster than grid buildouts, and often stay half-hidden until a project is already competing for transformers, turbines, and interconnection slots. (gridstrategiesllc.com) ### Why is AI different from normal load growth? A new subdivision or a few more EVs add demand gradually. An AI campus can arrive like a step-function — one customer, one location, hundreds of megawatts. That changes the grid’s problem from smoothing growth to serving concentrated shocks. E3’s December 2025 paper makes the point plainly: (gridstrategiesllc.com)re well. (ethree.com) ### How big is the power draw, really? Globally, data centers used about 415 TWh in 2024, or roughly 1.5% of all electricity demand. Inside a modern facility, servers and accelerators do most of the work and consume around 60% of the power. In the U.S., EPRI’s latest range is much starker — data centers could reach 9% to 17% of national electricity use by 2030. That is why this has moved from tech-story to grid-story. (iea.org) ### Why are forecasters suddenly nervous? Because the official numbers are jumping fast, but the underlying projects are messy. Grid Strategies says five-year U.S. peak-load forecasts rose more than sixfold in three years, from 24 GW to 166 GW. It also says data centers account for about 55% of forecast demand growth over the next five years. But the same report argues the data-center piec(iea.org) which projects are real, when they will arrive, or how large they will actually be. (gridstrategiesllc.com) ### What makes the forecasts so fuzzy? The catch is that “data center” is too broad a bucket. Some facilities are cloud-heavy, some are enterprise, some are crypto, and AI load often is not broken out at all in public utility forecasts. Grid Strategies says AI load is not categorically tracked in any publicly available utility forecast. S(gridstrategiesllc.com)he same time. (gridstrategiesllc.com) ### Doesn’t efficiency solve this? Partly — but not fast enough to make the problem go away. Energy Intelligence’s broader AI work argues AI will only moderately raise global electricity demand through 2050 because hardware and software get more efficient, and AI can help the power system run better too. That is an important brake on the most extreme doom scenarios. But efficiency helps over time. Grid planners still have to survive the near-term buildout wave. (energyintel.com) ### Why does this spill into the energy transition? Because every big AI load forces a practical question: what power source can show up on time? If renewables, transmission, storage, or firm clean power lag, developers fall back to gas, diesel backup, or whatever can be procured fastest. So AI doesn’t just raise demand. It can also distort the generation mix and crowd scarce equipment. That is why this story sits right on top of decarbonization, not beside it. (powering-intelligence.epri.com) ### So what’s the real takeaway? Basically, the AI power story is not “the grid is doomed” and it is not “efficiency will fix everything.” It is that electricity demand has become less smooth, less transparent, and more concentrated. Old forecasting models were built for broad trends. AI infrastructure behaves more like a series of giant private bets landing on local grids. (powering-intelligence.epri.com)buted — power, not chips, may be the real bottleneck. (gridstrategiesllc.com)