Goldman: data-center power to 66GW

Goldman Sachs on May 20 put a hard number on a constraint the AI industry has been talking around for months: electricity. The bank’s commodities research team said U.S. data-center power demand is expected to climb from 31 gigawatts in 2025 to 41 gigawatts in 2026 and 66 gigawatts in 2027, based on data-center development schedules from Aterio and a 70% capacity-utilization assumption. (goldmansachs.com) That forecast matters because it shifts the bottleneck from chips alone to the grid that feeds them. (goldmansachs.com) Goldman said U.S. data-center capacity is expected to reach roughly 95 gigawatts by the end of 2027, more than double the level at the end of 2025. Scheduled capacity additions are projected at 13.6 gigawatts in 2026 and 36.3 gigawatts in 2027, versus realized additions of 6.4 gigawatts in 2024 and 8.5 gigawatts in 2025. ### Where does the 66-gigawatt figure come from? Goldman Sachs said the 66-gigawatt figure is a demand forecast, not a nameplate count of every server hall under construction. The bank’s researchers — Hongcen Wei, Daan Struyven and Samantha Dart — tied the estimate to a roughly 95-gigawatt U.S. data-center capacity base by end-2027 and assumed 70% utilization. (goldmansachs.com) Aterio’s development schedules are central to that model. Goldman said those schedules show a sharp acceleration in annual additions after 2025, even with delays and cancellations still present in the market. ### Why does power show up before the next wave of servers does? (goldmansachs.com) Electricity infrastructure moves on a different timetable than AI deployments. Goldman said the rise in demand has implications for regional electricity markets and grid reliability, a sign that transmission, interconnection and local generation are becoming part of the data-center buildout equation. The pressure is concentrated rather than evenly spread. (goldmansachs.com) Dense compute clusters can require large increments of power in specific regions, which raises the cost of getting new capacity online and can lengthen project timelines, according to Goldman’s research framing. ### How much of a data center’s electricity bill goes to cooling? New Atlas reported this week that data centers consumed 485 terawatt-hours of electricity in 2025 and that about 30% of that total went to cooling. That makes thermal management one of the largest non-compute loads in the system. Cooling is also where one of the more concrete mitigation ideas surfaced. (goldmansachs.com) A paper published two weeks ago in *Cell Reports Physical Science* described a cold-plate design workflow that combines topology optimization with electrochemical additive manufacturing to directly print high-resolution pure-copper coolers. ### What exactly did the researchers build? (goldmansachs.com) The research team said the design uses 3D-printed pure-copper cold plates with complex internal geometries that conventional manufacturing struggles to produce. New Atlas said the approach could cut cooling energy use by more than 90%, reducing cooling’s share of total data-center electricity sharply if deployed at scale. (newatlas.com) The paper itself is narrower than the headline version. It reports a cold-plate development method and performance gains at the device-cooling level, which means the data-center-wide savings figures are an extrapolation from component performance rather than a report from a live hyperscale deployment. (sciencedirect.com) ### What does this change for operators and suppliers? Goldman’s forecast points to a market where securing power, moving workloads and cutting wasted energy become operating priorities alongside buying accelerators. That favors companies selling liquid-cooling hardware, power-management systems and software that can schedule compute around capacity constraints, according to the bank’s framing of the demand surge and grid implications. (newatlas.com) The next data points are already scheduled. Goldman’s forecast runs through the end of 2027, and the cold-plate results are now in the academic record in *Cell Reports Physical Science*, where operators and equipment makers can track whether the design moves from lab performance into commercial systems. (sciencedirect.com) (goldmansachs.com)