U.S. data centers consumed 176 terawatt-hours of electricity in 2023, triple the 58 TWh they used in 2014 and equal to roughly 4.4 percent of all electricity generated in the country. Goldman Sachs now estimates that data-center electricity demand could climb 160 percent by 2030, a pace that would outstrip the federal government’s own projections and force grid operators, utilities, and ratepayers to confront the fastest sustained rise in a single demand category in decades.
Why accelerating AI loads are outrunning federal forecasts
Between 2010 and 2018, researchers documented a striking pattern: global data-center compute output surged while electricity consumption barely budged. Eric Masanet and colleagues at Lawrence Berkeley National Laboratory showed in a 2020 Science analysis that hardware efficiency gains, virtualization, and the shift to hyperscale facilities had temporarily decoupled computing growth from energy use. That decoupling gave policymakers reason to treat data centers as a manageable share of the grid.
The arrival of large-scale AI training clusters has changed the arithmetic. A single AI-optimized facility can draw as much power as hundreds of thousands of households, according to the International Energy Agency’s analysis of energy and AI. Training runs for frontier models now last weeks on thousands of specialized chips, each consuming far more power per operation than the general-purpose servers that dominated the earlier era. The efficiency cushion that Masanet and his co-authors identified is being compressed faster than most official scenarios anticipated.
The Department of Energy’s most recent assessment projects that U.S. data-center electricity use will reach 325 to 580 TWh by 2028, which would represent about 6.7 to 12 percent of national electricity supply. Even the upper bound of that range, however, stops two years short of the Goldman Sachs 2030 horizon. If AI training clusters continue to scale at current rates, the 580 TWh ceiling could be breached before the decade ends, turning what the DOE treats as a high case into a baseline.
For households and businesses that share the same grid, the consequences are direct. Utilities planning generation and transmission investments over 10- to 15-year horizons need accurate demand signals. Underestimating data-center load means either scrambling for expensive peaking capacity later or delaying retirements of aging fossil-fuel plants that were expected to wind down. Either outcome raises electricity costs for everyone connected to the system.
DOE and IEA data anchoring the 160 percent projection
The Goldman Sachs estimate does not appear in any federal or intergovernmental dataset. No DOE, IEA, or Lawrence Berkeley National Laboratory publication contains the specific 160 percent figure or discloses the bank’s underlying methodology. What the primary data do confirm is a trajectory steep enough to make such a projection plausible.
The DOE report documents a threefold increase in U.S. data-center electricity use over nine years, from 58 TWh in 2014 to 176 TWh in 2023. Extrapolating that growth rate forward, and layering on the additional demand from AI-specific hardware, produces numbers that align with or exceed Goldman Sachs’s headline figure. The IEA’s overview places global consumption at roughly 460 TWh in 2024, with the United States accounting for the largest national share. Ireland, by contrast, already sees data centers consuming a striking proportion of its national electricity, a preview of what concentrated buildouts can do to smaller grids.
The gap between the DOE’s 2028 projection window and Goldman Sachs’s 2030 target year matters. Official federal scenarios were designed around a planning horizon that predates the current wave of announced hyperscale campuses. Several major cloud and AI companies have disclosed multi-billion-dollar construction programs that will bring new capacity online between 2028 and 2030, precisely the period the DOE range does not cover. Goldman Sachs appears to be filling that gap with its own modeling, but without published assumptions about chip efficiency trends, facility power-usage effectiveness, or renewable procurement rates, outside analysts cannot fully evaluate the estimate.
Still, both the DOE and IEA data point in the same direction: data centers are moving from a niche slice of electricity demand to a core driver. That shift has implications not only for generation capacity but also for where and when power is needed. AI-heavy campuses are clustering near cheap land, fiber backbones, and, increasingly, large sources of low-carbon power. The result is a patchwork of regional impacts that national averages can obscure.
Open questions for grid planners and electricity ratepayers
Three unresolved issues will determine whether the Goldman Sachs projection proves conservative or overstated. The first is chip-level efficiency. Each new generation of AI accelerators delivers more computation per watt, but training workloads are growing faster than hardware efficiency is improving. If chipmakers cannot bend that curve, total electricity demand will track closer to the high end of any forecast.
The second is the speed of grid interconnection. Data-center developers have filed requests for gigawatts of new grid connections across multiple U.S. regions, but interconnection queues at regional transmission organizations already stretch years. Bottlenecks in permitting, transformer supply, and transmission construction could slow actual buildout, effectively capping demand growth below what financial models assume.
The third is how much of the new load will be met with additional clean generation rather than existing fossil plants. Many hyperscale operators are signing long-term power purchase agreements for wind, solar, and, in some cases, nuclear output. Yet the timing mismatch between data-center construction and new generation coming online is acute. If facilities are energized before contracted clean projects are built, grids may lean on gas and coal plants in the interim, raising emissions even if the long-run plan is greener.
For ratepayers, the stakes are twofold. On one hand, large industrial customers can help justify new transmission lines and generation projects that ultimately improve reliability for everyone. On the other, if utilities socialize the cost of serving data centers across their entire customer base, residential and small-business bills could rise to subsidize corporate computing growth. State regulators are only beginning to grapple with how to allocate those costs fairly.
Regulatory responses are emerging but uneven. Some states are considering higher demand charges or bespoke tariffs for very large loads, aiming to align prices with the cost of new infrastructure. Others are exploring requirements that new data centers procure a certain share of their energy from additional renewable resources, rather than simply buying existing green credits. These tools could moderate the impact of AI-driven growth on the broader grid, but they also risk pushing investment to jurisdictions with looser rules.
Meanwhile, communities near proposed campuses are weighing the trade-offs. Data centers bring construction jobs, tax revenue, and long-term operations roles, but they also require water for cooling in many designs, occupy large tracts of land, and can strain local distribution networks. As the sector scales, siting battles that once focused on power plants and transmission lines are increasingly encompassing server farms as well.
Whether the United States ultimately sees a 160 percent jump in data-center electricity demand by 2030 or something lower, the direction of travel is clear. AI and cloud computing are anchoring a new era of electricity consumption, one that is growing faster than legacy planning models assumed. For grid operators, regulators, and customers, the question is no longer whether data centers will reshape the power system, but how quickly they will do so-and who will bear the cost of keeping the lights, and the servers, on.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
