AI data centers push up utility bills
- NERC issued a Level 3 alert on May 4 after sudden swings from giant computational loads threatened grid stability, putting AI-era data centers under sharper scrutiny. - Georgia regulators already moved the other way on costs — requiring new 100-megawatt-plus customers to cover upstream grid spending and backstopping 9,885 megawatts of new supply. - The fight now is simple: who pays first for power, wires, substations, and water when AI demand shows up fast?
Electric bills are where the AI boom stops feeling abstract. Data centers need huge amounts of power, and the grid was not built for thousands of megawatts of new demand to show up all at once. For a while, the fight was mostly theoretical — would ordinary customers end up paying for the substations, wires, and power plants built for hyperscale computing? This month, that debate got more concrete when NERC, the continent’s grid reliability watchdog, issued a Level 3 alert over risks from large computational loads. ### What actually changed this month? On May 4, NERC said it had seen customer-driven large-load drops and oscillations happening in seconds — fast enough to threaten bulk-power reliability before operators can react. That is not a generic “growth is coming” warning. It is a highest-tier operational alert tied to behavior already showing up on the grid, with required responses due by August 3. (nerc.com) ### Why does that matter for your bill? Because reliability problems and bill problems are really the same infrastructure problem seen from two angles. If utilities have to build more generation, transmission, and substation capacity for big new loads, somebody has to finance it. Bain estimated in late 2024 that serving data-center growth could add about 1% a year to customer bills through 2032, and that data centers could drive 44% of U.S. load growth from 2023 to 2028. (nerc.com) ### How big is the power demand now? Bigger than even last year’s scary forecasts. EPRI now says U.S. data centers could use 9% to 17% of national electricity by 2030, up from roughly 4% to 5% today. EPRI also says those 2030 scenarios are about 60% higher than its 2024 projection — basically, the buildout accelerated faster than expected. ### So do ordinary customers always get stuck with the cost? (utilitydive.com) Not always — and Georgia is the clearest test case. In January 2025, the Georgia PSC approved rules letting Georgia Power put special terms on new customers above 100 MW. Those customers can be charged not just for site-specific hookups, but also for upstream generation, transmission, and distribution costs as construction progresses. The whole point was to stop residential and small-business customers from subsidizing giant new loads. (restservice.epri.com) ### What did Georgia do after that? In December 2025, the PSC approved a deal for 9,885 MW of new generation, most of it aimed at serving large new customers such as data centers. The important bit is the backstop — if the expected data-center contracts do not materialize, Georgia Power itself has to absorb the risk rather than leaving stranded assets on everyone else’s bill. State regulators framed that as protection for existing customers, and Georgia Power says the arrangement could put about $102 a year of downward pressure on a typical household bill starting in 2029. (psc.ga.gov) ### Does that solve the national problem? No — it just shows one model. In PJM, which covers the country’s biggest data-center corridor, planners have had to get stricter about what counts as “firm” data-center demand. The 2026 long-term forecast imposed assumptions like a minimum 36-month ramp and a 70% utilization rate unless better evidence exists. That is a sign planners do not fully trust every announced project to show up exactly as promised. (psc.ga.gov) ### Where does water come in? Power is the first bottleneck, but water is the second. Microsoft said in January that communities are increasingly worried not just about higher power bills but also tighter water supplies around new AI infrastructure. Separate academic work has been warning that data-center cooling demand is becoming a serious local water-planning issue, especially in stressed regions. (pjm.com) ### Bottom line? The real story is not “AI uses lots of electricity.” Everybody knew that. The new story is that regulators and grid operators are finally deciding who carries the risk when AI demand arrives faster than the grid can safely and affordably expand. Georgia is trying to push that risk onto the data centers. NERC’s alert shows why that fight is no longer optional. (psc.ga.gov) (blogs.microsoft.com)