S&P Global projects that electricity demand from data centers will roughly triple by 2030, a pace that far exceeds the global baseline the International Energy Agency set when it estimated consumption would more than double over the same period. The forecast lands as utilities, grid operators, and state regulators across the United States confront a surge in interconnection requests concentrated in a handful of states. With data centers already drawing approximately 415 TWh worldwide in 2024, or about 1.5 percent of total electricity use, the gap between planned generation and incoming load is widening fast enough to force hard choices about transmission investment, permitting timelines, and fuel mix well before the end of the decade.
Why a tripling forecast changes the grid math right now
The tension behind S&P Global’s projection is not simply about bigger numbers on a national chart. It is about where those numbers land. Lawrence Berkeley National Laboratory has tracked U.S. data-center electricity consumption back to 2014 using server shipment data and facility-level surveys compiled in its latest report to Congress. That historical series shows steady growth, but nothing close to the step-change now anticipated. The LBNL dataset, built from granular shipment records and earlier academic studies, reveals that a small cluster of states, led by Virginia, Texas, and a few others with established data-center corridors, absorbed the bulk of new capacity over the past decade.
If S&P Global’s tripling estimate holds, those same states will bear a disproportionate share of new load. Transmission planning cycles in most U.S. regions run three to five years from study to energization. That timeline means grid upgrades approved today would barely reach service by 2028 or 2029, precisely when the steepest demand ramp is expected. Utilities filing integrated resource plans in 2026 face a concrete problem: they must commit capital to generation and wires based on forecasts that have shifted dramatically in less than two years, with little room for error if demand materializes faster than construction schedules allow.
The IEA’s own base-case estimate, which projects global data-center consumption reaching roughly 945 TWh by 2030, already represents a doubling from the 2024 level of approximately 415 TWh. In its Energy and AI analysis, the agency ties that growth to artificial-intelligence training workloads, cloud migration, and expanding digital services. S&P Global’s tripling scenario implies U.S. growth alone could push well beyond the share assumed in that global aggregate. The practical result: even if worldwide totals stay within the IEA’s range, regional grids in the states hosting the largest campuses could hit binding constraints years ahead of any national shortfall.
What LBNL shipment data and IEA baselines actually show
Two primary datasets anchor the current debate. The IEA’s global figures establish that data centers consumed about 415 TWh in 2024, accounting for roughly 1.5 percent of world electricity use, and then model an increase to about 945 TWh by 2030. Those numbers function as a floor and a reference trajectory for every forward projection, including S&P Global’s more aggressive view.
Separately, LBNL’s 2024 report constructed a U.S.-specific consumption timeline stretching back a decade, drawing on server and storage shipment volumes, facility power-use-effectiveness ratios, and peer-reviewed research cataloged through federal repositories. The report does not extend to a formal 2030 scenario, but its historical trend line makes clear that recent annual growth rates in the United States already exceeded earlier projections. That acceleration is what gives S&P Global’s tripling forecast its plausibility: the gap between LBNL’s measured trajectory and the new forward estimate is large, yet the direction of the data supports a sharper curve than analysts assumed even three years ago.
Another point emerging from the LBNL work is the role of efficiency. Over the last decade, better cooling systems, higher server utilization, and improved power-use effectiveness helped moderate the energy impact of rising compute demand. Yet those gains appear to be slowing just as AI and high-performance computing workloads arrive. If each new generation of hardware no longer delivers proportionate efficiency improvements at the facility level, then electricity demand will track more closely with raw compute growth, tightening the margin for error in resource planning.
No publicly available S&P Global document details the exact methodology or regional assumptions behind the tripling figure. The projection has been cited in utility filings and industry presentations, but the underlying model, including how it allocates growth across states or accounts for efficiency gains from newer chip architectures, has not been released in a form that outside analysts can fully audit. That matters because the difference between a national tripling and a regional tripling is the difference between a manageable build-out and a grid emergency in specific corridors.
Gaps in the forecast and what grid planners should watch next
Several questions remain open. First, the IEA’s global base case and S&P Global’s tripling estimate use different scopes and likely different definitions of what counts as a data center. Colocation facilities, enterprise server rooms, and hyperscale campuses each carry distinct load profiles. Without a transparent methodology from S&P Global, it is difficult to know whether the two forecasts are measuring the same universe of facilities or whether definitional differences account for part of the gap.
Second, LBNL’s historical series ends before the current wave of AI-driven construction announcements. The report captures a period when most new capacity served cloud computing and enterprise workloads, not the dense GPU clusters that AI training demands. GPU racks can draw five to ten times more power per rack than traditional CPU-based servers, and they concentrate that load in smaller footprints. That combination stresses both on-site distribution equipment and upstream substations, even when total megawatts at a campus look similar on paper to earlier generations of facilities.
Third, neither the IEA baseline nor the LBNL history fully resolves how much demand might be offset by emerging efficiency measures such as liquid cooling, workload shifting, or more aggressive use of demand response. These tools can reshape hourly load profiles and reduce peak stress on local feeders, but they do not eliminate the need for additional generation and transmission if annual energy consumption climbs as rapidly as S&P Global suggests.
For grid planners, the most important near-term indicators will not be national forecasts but project-level signals. Interconnection queues in a handful of regions already show clusters of very large data-center requests arriving within a few years of one another. Local land purchases, substation upgrade applications, and long-term power contracts can all reveal whether operators are preparing for sustained multi-gigawatt build-outs or a shorter burst of activity. Regulators and utilities that track these granular markers will be better positioned to distinguish between transient hype cycles and durable structural demand.
At the same time, planning processes need to adapt to the speed of change implied by a tripling scenario. Traditional three- to five-year transmission studies and decade-long integrated resource plans were built for slower-moving demand trends. If data-center growth continues to surprise on the upside, states may need more iterative planning frameworks that can incorporate new information annually, along with contingency pathways that outline how to respond if specific clusters of projects advance faster than expected.
The core policy dilemma is straightforward: underbuild and risk reliability problems and economic losses if demand arrives, or overbuild and saddle customers with the costs of underutilized assets if forecasts prove too aggressive. The IEA baseline and LBNL history provide guardrails, but they do not settle that trade-off. Until S&P Global or other forecasters disclose more detail about their assumptions, utilities and regulators will have to navigate between these benchmarks, using local data to refine the picture. What is clear already is that the geography of data-center growth will matter as much as the global totals, and that decisions made over the next few years will shape how resilient-or how strained-the grid will be when 2030 arrives.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
