US Data Center Power Demand Set to Double to 66 GW by 2027

A forecast that changes the grid conversation

Goldman Sachs Commodities Research expects US data center power demand to more than double in two years. Its forecast has demand rising from 31 gigawatts in 2025 to 41 GW in 2026 and 66 GW in 2027, driven mainly by the buildout of AI infrastructure.1 The analysts behind the forecast, Hongcen Wei, Daan Struyven and Samantha Dart, base it on an expectation that total US data center capacity will reach about 95 GW by the end of 2027, more than twice the end-2025 level, with facilities running at an assumed 70% utilization.13

The more telling number is the share of peak load. Goldman projects that data centers will account for 8.5% of total US peak summer power demand in 2027, up from 4.1% in 2025 and 5.3% in 2026.1 When one class of customer doubles its share of peak demand in two years, the strain does not stay with that customer. Goldman says the tightening will affect electricity prices and grid stability across the country.3

The forecast came out in May. A follow-up note in early October suggests the bank still expects rapid growth, but the details have shifted, and those shifts are covered below.10

The pipeline is bigger than what will get built

The Goldman research stands out for how much it discounts its own data. The forecast relies on facility-level schedules from Aterio, which tracks location, permitting progress, construction announcements and satellite imagery.1 Taken at face value, those schedules show capacity additions of 13.6 GW in 2026 and 36.3 GW in 2027. Actual additions were 6.4 GW in 2024 and 8.5 GW in 2025.14

Goldman does not take the schedules at face value. In the past, only about 72% of data centers scheduled to start up within the next four quarters did so on time. After adjusting for that, the analysts expect roughly 60% of next year's scheduled capacity to arrive on time, falling to about 50% over two years.1 Even with that discount, they still forecast 11.5 GW of new capacity in the last three quarters of 2026, after 2.2 GW came online in the first quarter.3

The reasons for the gap say a lot about how the industry works. Developers often file applications in several regions at once and then build only at the best site, which inflates the pipeline. Supply chain and labor shortages are the most common causes of delay. Once permits are in hand, a typical facility takes 18 to 24 months to build.1 One summary of the client report calls the multi-region filing practice "site gaming" and says waits for electrical equipment alone can add months.8

The coverage does not agree on the size of the raw pipeline. Goldman's public article gives scheduled additions of 13.6 GW and 36.3 GW.1 A summary of the underlying client report, published May 5, gives much larger planned figures of 19 GW for 2026 and 69 GW for 2027.8 The likeliest explanation is that the two sets of numbers measure the pipeline at different stages or with different filters. Either way, the 66 GW forecast is a heavily discounted version of an even larger set of ambitions.

Not even Goldman agrees with itself

The wide spread between forecasts gets less attention than it should. The client-report summary says Goldman's own Equity Research team, working from S&P 451 Research data, projects 39 GW for 2026 and 50 GW for 2027. That is 16 GW below the commodities team's 2027 figure.8 The same summary puts outside estimates anywhere from just over 30 GW (US Department of Energy) to 67 GW (Boston Consulting Group).8

The October note adds another wrinkle. It forecasts 38% demand growth in both 2026 and 2027, which works out to about 12 GW of new demand in 2026 and 17 GW in 2027. It also lowers the end-2027 capacity estimate to 90 GW, down from roughly 95 GW in May.10 Goldman still describes its 2027 outlook as "largely unchanged". By simple addition, though, 31 GW plus 12 GW plus 17 GW comes to about 60 GW for 2027, not 66 GW. The October figures may be defined differently, since the reports do not reconcile the two notes. But the direction of the revision points slightly down, and the headline about doubling demand should be treated as the top of a plausible range rather than a firm estimate.

This does not weaken the main conclusion. Every forecast cited in the coverage still shows demand growing faster than the grid is used to handling. The spread does matter, however, for utilities deciding whether to commit billions of dollars to generation and transmission based on a load that could land almost anywhere between 50 GW and 66 GW.

Geography decides who feels the squeeze

The impact will vary sharply by region. Goldman expects the average 2027 data center additions in each of three power markets, the Mid-Atlantic, Texas and the Mid-Continent, to exceed the entire country's additions in 2025.1 Those regions roughly match the PJM, ERCOT and MISO grid operators.8

The outcomes split from there. Goldman sees elevated reliability risk in the Mid-Atlantic, the Mid-Continent and the Northwest, where planned new generation is small compared with the incoming load. It warns that those regions may eventually have to turn some data centers away.1 Texas and Georgia, by contrast, should see only marginal tightening because they have large amounts of new generation in development.4 Tennessee, New England and Florida are already so tight that new data center additions there are likely to be limited.3

PJM shows how this works in practice. Its Independent Market Monitor attributes about 7.9 GW of additional data center demand in the 2025/26 delivery year and about 12 GW in 2026/27, and connects that growth to a doubling of capacity costs across the region.13 Even so, PJM continues to attract investment, especially in Virginia and Ohio, because being close to customers can outweigh scarce power when developers choose sites.8 Goldman lists power availability and time-to-client as the main factors in location decisions.1 Where those two pull in opposite directions, households and other ratepayers end up absorbing the conflict.

The wider context: half of all new demand

The Goldman figures fit a larger pattern. As reported by Fortune, the International Energy Agency found that data centers accounted for roughly half of all new US electricity demand in 2025 and expects them to keep doing so through 2030.6 The US Energy Information Administration projects total US consumption rising from a record 4,195 billion kWh in 2025 to 4,288 billion kWh in 2026 and 4,356 billion kWh in 2027, largely because of AI and crypto data centers.9 An EIA analysis in March forecast load growth of about 1.9% in 2026 and 2.5% in 2027, with the fastest growth in ERCOT and PJM.16

The contrast between those numbers is the key point. National demand growth in the low single digits sounds manageable. The trouble is that data center load is concentrated in a few places and peaks sharply. Planners also point to a timing mismatch: a data center can be built in one to three years, while major grid upgrades can take five to fifteen, according to the IEA.17 So the binding constraint is connecting new load at specific substations, not total generation nationwide.

Forecasters outside Goldman see the gap widening. TrendForce expects global data center power demand capacity to reach 161 GW in 2026, up about 31%. It projects a US supply-demand gap of more than 170 GW by 2030.12 TrendForce adds that the headline gap leaves out behind-the-meter generation and so may overstate the real shortfall, while interconnection delays could still cause genuine shortages.18 McKinsey expects data center electricity demand to grow about 24% a year through 2030 and says power supply will trail compute demand until then.19

The likely outcome

The 66 GW figure is less important than what it shows about which constraint matters. Goldman's own adjustments assume that roughly half of scheduled capacity over two years will miss its target date.1 Supply chains, labor and grid connections, rather than demand, are already setting the pace of the buildout. Goldman also acknowledges risk in the other direction: projects not yet in the pipeline, and capital spending heavy enough to cut construction times to about a year, could push additions above schedule.3

Taking the coverage as a whole, US data center load will very likely grow somewhere between 50% and 100% by 2027. Where it lands in that range will depend less on AI demand than on how fast transformers, turbines and transmission lines arrive. Goldman's October note says local opposition has so far had little effect on growth trajectories.10 That may hold for now, but resistance is growing.6 The regions Goldman flags for reliability risk are where it is most likely to become a real constraint.1 Texas and Georgia are positioned to benefit, and the Mid-Atlantic is the region to watch over the next two summers.