The Pulse
Goldman Sachs Raises Data-Center Power Forecast on AI Token Growth
Goldman Sachs now expects global data-center power demand to rise 170% by 2030 from 2025 levels, up from a 117% estimate made in March. Analysts cite higher hyperscaler spending, more AI-server shipments and expanding token consumption desp
AI.info Team ·
Goldman Sachs has raised its forecast for global data-center power demand after hyperscaler spending, AI-server shipments and token consumption all exceeded the pace analysts expected earlier this year.
The bank now expects data-center power demand worldwide to increase 170% by 2030 from 2025 levels. Goldman’s previous estimate, discussed in March, called for a 117% increase over the same period. The new forecast arrives as the physical systems supplying AI workloads face longer connection queues, shortages of turbines and transformers, limited skilled labor and growing resistance from communities and regulators.
“The binding constraint has migrated away from megawatts and towards what I think of as all the Ts,” Allison Nathan, co-host of the Goldman Sachs Exchanges podcast, said in the episode, which Goldman recorded on August 12 and published on September 1. She listed “turbines, transformers, transmission, and tradespeople” as the constraints now shaping how quickly AI infrastructure can expand.
Goldman’s analysts do not argue that every announced data center will be built on schedule. Their forecast instead reflects a larger project pipeline, higher use of existing facilities and a rise in the amount of computing that customers are willing to buy as AI services become cheaper and more capable.
Goldman’s 108-Gigawatt Revision
Carly Davenport, a senior analyst on Goldman Sachs’ Americas utility team, said the bank raised its overall U.S. electricity-demand forecast through 2030 to a 3.5% compound annual growth rate from about 3.2% previously. The primary change came from a revised data-center outlook based on data from 451 Research.
Goldman now expects U.S. data-center power demand to reach about 108 gigawatts in 2030, up from an earlier estimate of roughly 83 gigawatts. Davenport said the increase reflects new projects entering development queues and higher utilization of facilities that are already operating.
Vacancy rates across major U.S. data-center markets have fallen to about 1% to 2%, according to Davenport, compared with a range of roughly 2% to 7% over the past several years. Goldman forecasts average U.S. vacancy of about 3% in 2030, suggesting that new construction will add capacity while existing buildings continue to run closer to full use.
Hyperscaler spending assumptions have also moved sharply. Brian Singer, global head of GS Sustain at Goldman Sachs, said the bank’s analysts had estimated hyperscaler spending at about $1.2 trillion for 2027 in March. That figure now stands at $1.7 trillion. The estimate for 2029 has increased from $1.5 trillion to $2.1 trillion.
“It’s hard to ignore when you have all of those factors to not further raise forecasts, even though it’s only been a handful of months,” Singer said. “And so, that’s what we’ve done.”
Goldman also expects higher shipments of servers designed for AI workloads. Those systems consume more power than conventional enterprise servers, particularly when deployed in large clusters used for model training and inference. The bank’s forecast treats the increase in AI hardware as part of a wider demand cycle rather than as a temporary construction surge.
Efficiency Has Not Reduced the Appetite for Compute
The unusual feature of the current forecast is that efficiency improvements have not produced a corresponding fall in spending or electricity demand. Models are becoming more capable per unit of computation, and companies are paying closer attention to how many tokens they use. Yet those savings are helping customers run more tasks rather than abandon AI workloads.
Goldman’s analysts describe tokens as the basic units of text, code or other data processed by large AI models. Lower costs per token make it economical to apply models to a wider range of enterprise work, including tasks that previously could not justify the expense of extensive computation.
“All those expected or anticipated efficiency gains are actually happening,” George Lee, co-head of the Goldman Sachs Global Institute, said during the discussion. “And yet, the demand profile is both recovering those and expanding at the rate you described.”
Lee said greater token efficiency expands the number of enterprise tasks that can be addressed economically. The result is a familiar rebound effect: each unit of computation becomes more productive, but the lower price encourages customers to consume more units overall.
Goldman has not seen productivity gains reduce technology research-and-development budgets, Singer said. Customers are still spending more on computing and consuming more tokens rather than holding output constant while cutting their AI budgets.
Companies are beginning to monitor token usage more carefully. Singer said customers are examining where and how they consume tokens in an effort to reduce intensity without reducing output. That may improve efficiency at the application level, but Goldman does not see evidence that aggregate demand has peaked.
The International Energy Agency has reached a similar conclusion from a broader energy perspective. Its updated analysis of energy and AI says global data-center electricity consumption rose 17% in 2025, while electricity use from AI-focused data centers increased 50%. The agency projects total data-center consumption to nearly double from 485 terawatt-hours in 2025 to 950 terawatt-hours in 2030.
U.S. Growth Moves Toward the Midwest
Goldman expects the PJM power market, which covers much of the Mid-Atlantic, to remain the largest U.S. data-center market through 2030. PJM currently accounts for about 35% of U.S. data-center power demand, supported by existing clusters, a high-voltage transmission network and merchant nuclear plants that appeal to large electricity buyers.
The more surprising shift is expected in the Midwest. Goldman forecasts that the Midcontinent Independent System Operator, or MISO, will overtake Texas, the Southeast and the Pacific Northwest to become the second-largest U.S. data-center market by 2030, with about 16% of national demand.
MISO spans a broad region from Minnesota to Louisiana. Davenport said regulated utilities in states including Iowa, Wisconsin and Louisiana are helping drive the market by developing generation, transmission and distribution capacity for large customers.
Texas is expected to remain a major center but fall to third place, with about 14% of U.S. data-center demand in 2030. ERCOT, the state’s main power market, accounts for about 15% of demand today. Goldman expects Texas to tighten as demand accelerates around 2028, although it does not currently view the market as critically constrained.
Regional conditions differ because power markets do not all have the same generation pipeline, transmission capacity or regulatory structure. Goldman expects the Mid-Atlantic, Mid-Continent and Northwest to face higher reliability risks, while Texas and Georgia have more generation under development to absorb new load.
Grid Queues Are Turning Into a Business Constraint
Data-center developers often want facilities online before the grid can provide a connection. Goldman says the wait for a U.S. grid connection can range from two to seven years depending on the regional market, while large customers typically want to begin operations much sooner.
Behind-the-meter generation has emerged as a bridge. Under that model, a data center receives power from an on-site plant that operates separately from the broader grid. Goldman expects about 30 gigawatts of behind-the-meter gas capacity by 2030, representing roughly 20% of its projected data-center demand at that time.
Davenport described the arrangement as a temporary measure rather than a permanent replacement for grid power. Large data centers generally prefer grid connections because the grid offers a cheaper and more dependable source at multi-gigawatt scale. Some projects may begin with on-site generation and later connect to the grid when transmission upgrades are complete.
Goldman expects natural gas to supply about 60% of data-center demand and renewables to supply roughly 40%. Nuclear-plant restarts may provide additional capacity in specific markets, but the bank’s forecast assumes an “all of the above” approach rather than a single technology carrying the expansion.
Gartner’s June forecast points in the same direction. The research firm expects worldwide data-center electricity consumption to reach 565 terawatt-hours in 2026, up 26% from 447 terawatt-hours in 2025. Gartner estimates that AI-optimized servers will account for 31% of data-center power consumption this year and consume more electricity than conventional servers in 2027.
Transformers, Turbines and Cooling Systems Set the Pace
Goldman’s supply analysis has expanded from six constraints to seven. The list includes the pervasiveness of AI, productivity improvements, power prices, policy, parts, people and the physical environment surrounding data centers.
Parts shortages are already affecting project schedules. Singer said one major turbine manufacturer indicated during a recent earnings call that it expected to be halfway sold out for 2031 by the end of 2026. Turbine availability does not eliminate demand for electricity, but it limits the ways developers can supply new facilities.
Skilled labor presents a slower-moving problem. Goldman expects demand for electricians and high-voltage welders to rise, but training and certification can take years. The U.S. spent much of the past decade with little growth in electricity demand, reducing the incentive for workers to enter some of those trades.
Cooling adds another layer of pressure. Goldman says more than half of data centers are located in areas exposed to elevated physical risks such as extreme heat, humidity or drought. Operators may need to choose between using more water and using more electricity, depending on local conditions and the cooling technology available.
The IEA has also warned that AI servers are changing the physical design requirements of data centers. Its analysis says the power density of AI servers increased elevenfold between 2020 and 2025 and could rise another fourfold by 2027. A single advanced server rack could then have peak power demand equivalent to about 65 households.
Those figures explain why Goldman’s forecast is not simply a projection of more buildings. It is a forecast of larger electrical connections, denser equipment, new generation, expanded transmission and a growing need for workers who can install and maintain high-voltage systems.
Goldman does not forecast a nationwide U.S. power shortage. Davenport said certain markets, especially PJM, are becoming critically tight while ERCOT is likely to face greater pressure later in the decade. The forecast puts the central number plainly: global data-center power demand in 2030 is now expected to stand 170% above 2025 levels, even after efficiency gains have reduced the amount of energy required for each token.