AI Electricity Demand 2030: IEA vs Goldman Sachs vs EPRI, Compared Honestly

The Pulse

Three of the most credible forecasters in the energy industry, the International Energy Agency, Goldman Sachs Research, and EPRI, all agree that AI will roughly double global data center electricity demand by 2030. They disagree by as much as 400 terawatt hours on what that doubling actually means in absolute terms.

The IEA’s base case puts global data center electricity consumption at around 945 terawatt hours by 2030. Goldman Sachs frames its forecast in capacity terms instead, a 165% increase in power demand compared to 2023.

EPRI, the US-focused Electric Power Research Institute, takes yet another approach entirely, publishing four distinct scenarios rather than a single number, with annual growth rates ranging from 3.7% to 15% depending on assumptions about AI adoption and efficiency gains.

Core significance

Why it matters:

• The IEA projects data center electricity consumption will roughly double to 945 terawatt hours by 2030, growing 15% annually:  That growth rate is more than four times faster than total electricity consumption growth from every other sector combined.[IEA Energy and AI energy demand from AI report]

Electricity use from accelerated servers, the GPU and AI accelerator hardware specifically, is projected to grow 30% annually in the IEA’s base case, while conventional server consumption grows at a much slower 9% per year, meaning AI hardware alone accounts for almost half of the entire net increase in global data center electricity demand.

• Goldman Sachs forecasts global data center capacity will reach 122 gigawatts by the end of 2030, a 165% increase from 2023:  The firm projects facility occupancy will peak above 95% in late 2026 before moderating as more capacity comes online from 2027 onward.[Goldman Sachs AI drive 165 percent data center power demand 2030]

Goldman’s capacity mix forecast also shows hyperscalers and wholesale operators growing from 60% of total capacity today to 70% by 2030, a meaningful concentration shift away from smaller enterprise and colocation operators.

• EPRI’s range is the widest of the three major forecasters by design, not by accident:  Its lowest growth scenario assumes limited AI adoption combined with strong efficiency gains, while its highest scenario combines rapid AI expansion with minimal efficiency improvement, placing data centers anywhere from 4.6% to 9.1% of total US electricity by 2030.[EPRI data centers consume up to 9 percent US electricity]

Deep Context: Why three credible forecasters disagree by hundreds of terawatt hours

The divergence between these forecasts is not a sign that any one organization is wrong. It reflects genuinely different modeling choices about training versus inference split, GPU efficiency curves, power usage effectiveness assumptions, and hyperscaler versus enterprise demand mix, each of which compounds significantly by 2030. [Useluminix powering AI boom grid breaks first forecasts synthesis]

EPRI’s own February 2026 update to its Powering Intelligence report raised its US data center electricity projections by 60% compared to its 2024 estimates, driven specifically by eighteen months of accelerated AI-fueled project announcements. That single revision illustrates how quickly the underlying input data can shift even within the same forecasting methodology.

EPRI’s updated low, medium, and high 2030 scenarios for US data center demand now sit at 384, 596, and 793 terawatt hours respectively, representing 9%, 13%, and 17% of roughly 4,400 terawatt hours of total US generation. That range alone, just from one forecaster’s internal scenarios, spans more terawatt hours than several entire countries consume annually.

Why the IEA’s number looks more conservative than it actually is

The IEA’s headline 945 terawatt hour figure sounds more modest than Goldman’s 165% increase framing, but the two are not actually in conflict once measured the same way. The IEA’s updated 2026 projection shows global data center electricity consumption roughly doubling from 485 terawatt hours in 2025 to 950 terawatt hours in 2030. [IEA executive summary key questions energy AI updated]

Critically, the IEA notes that electricity consumption specifically from AI-focused data centres grows much faster than the overall data center total, tripling over that same period even as the broader data center category only doubles. As covered in our training vs inference report, that AI-specific tripling is consistent with inference’s rapidly growing share of total AI compute, since inference workloads are precisely the category expanding fastest inside the broader data center electricity total.

Data Insights

By the numbers:

All figures from official IEA, Goldman Sachs Research, EPRI, and Deloitte projections cited inline.

• 130% increase: the IEA’s specific projection for US data center energy demand by 2030:  Deloitte’s independent analysis lands close to the same trajectory, projecting global data center consumption reaching 1,065 terawatt hours by 2030, modestly above the IEA’s 945 terawatt hour base case. [Brookings global energy demands AI regulatory landscape Deloitte]

The US Department of Energy’s Lawrence Berkeley National Laboratory separately projects an increase from 176 terawatt hours in 2023 to a range of 325 to 580 terawatt hours by 2028 alone, a band wide enough on its own to contain most of the disagreement between the major global forecasters.

• EPRI’s full comparison chart shows lower-bound 2030 estimates near 200 to 400 terawatt hours and mid-range scenarios near 400 to 600 terawatt hours for the US specifically:  The chart synthesizes external estimates from BCG, BloombergNEF, McKinsey, S&P, and Jefferies alongside EPRI’s own scenarios, mapped against the historical Lawrence Berkeley National Laboratory baseline. [EPRI Powering Intelligence 2026 summary future projections]

US historical demand rose gradually from approximately 70 terawatt hours in 2014 to around 150 terawatt hours by 2023, meaning even the most conservative 2030 projections still represent a faster decade of growth than anything the data center sector has previously experienced.

Table 1: Global data center electricity demand forecasts by 2030

Forecaster	2030 estimate	Measurement basis	Key assumption	Confidence framing
IEA, base case	945 TWh globally	Direct consumption	15% annual growth, AI servers triple	Single base case plus scenarios
Goldman Sachs	165% increase from 2023	Capacity and demand growth rate	122 GW global capacity by 2030	Single central forecast
Deloitte	1,065 TWh globally	Direct consumption	Consistent with IEA trajectory	Single projection
EPRI, US only	384 to 793 TWh	Direct consumption, US specific	3.7% to 15% annual growth range	Four explicit scenarios

Table 2: Where the forecasts actually agree versus diverge

Dimension	Where forecasters agree	Where they diverge
Direction of demand	All project significant growth through 2030	Magnitude of growth, 2x to 6x range across studies
AI specific share	AI workloads grow faster than total data center demand	Whether AI triples, or grows even faster relative to baseline
Methodology	Training and inference both factor into models	PUE assumptions, hyperscaler mix, efficiency gain pace

The Business Case: Which forecast should enterprises actually plan around

For enterprises making multi-year AI infrastructure commitments, the practical answer is not to pick a single forecaster’s number and plan around it precisely. The practical answer is to plan around the range, and specifically around EPRI’s explicit scenario structure, since it is the only major forecast built deliberately to show low, medium, and high outcomes rather than a single point estimate.

EPRI’s medium scenario of 596 terawatt hours for US data centers by 2030 sits close to the midpoint of where IEA, Goldman, and Deloitte’s global figures imply the US share should land once adjusted for the country’s roughly 45% share of global data center electricity consumption. Treating that medium scenario as the working planning assumption, while building contingency for the high scenario given how often EPRI itself has revised upward, is the most defensible approach available.

As covered in our AI data center power consumption pillar, the practical infrastructure implications of even the lower-bound forecasts are already reshaping where enterprises can realistically site new compute, regardless of which exact terawatt hour figure ultimately proves correct.

Expert Nuance: The real lesson is that every forecast keeps getting revised upward

The single most informative data point across all of these competing forecasts is not any individual 2030 number. It is the direction every major forecaster has moved every time they have updated their models over the past two years. [ArXiv concentrated siting AI data centers regional power stress]

EPRI raised its own projections by 60% in a single February 2026 revision. The IEA’s updated 2026 report kept its central trajectory roughly consistent with its 2025 figures, but explicitly noted that supply chain and construction bottlenecks are the main factor preventing even more aggressive near-term scenarios from materializing faster, not a lack of underlying demand.

That pattern, demand forecasts revised upward repeatedly while physical bottlenecks rather than demand ceilings explain why actual deployment lags the most aggressive projections, is the most reliable signal in this entire body of research. Every forecaster studied here has been more likely to underestimate AI electricity demand at the time of original publication than to overestimate it.

Strategic Outlook

1. Watch for EPRI’s next scenario revision as the most useful real-time signal:  Because EPRI updates its model based on actual state-level commercial development pipelines rather than equipment shipment projections, its revisions tend to reflect genuine on-the-ground project activity faster than the IEA’s or Goldman’s global modeling approaches.
2. The gap between nameplate capacity announcements and realized peak load will keep narrowing the range of legitimate uncertainty:  As more announced data center projects either reach construction or get cancelled outright over the next 18 months, today’s widest-range scenarios will collapse toward whichever forecaster’s assumptions about project realization rates proves most accurate.
3. US specific demand will likely continue outpacing the global average growth rate:  With roughly 45% of global data center electricity consumption already concentrated in the US, and IEA projecting 130% US specific growth by 2030 versus a global doubling, the US grid will absorb a disproportionate share of whichever global scenario ultimately plays out.

Key Question Answered

How much will AI increase electricity demand by 2030, according to IEA, Goldman Sachs, and EPRI?

All three major forecasters project global data center electricity demand will roughly double by 2030, though they disagree meaningfully on the exact magnitude.

The IEA’s base case projects 945 terawatt hours globally, Goldman Sachs frames its forecast as a 165% increase in power demand compared to 2023 reaching 122 gigawatts of capacity, and Deloitte’s independent estimate lands close to the IEA at 1,065 terawatt hours. EPRI, focused specifically on the United States, publishes four scenarios ranging from 384 to 793 terawatt hours by 2030, representing 9% to 17% of total US electricity generation. The forecasts diverge primarily due to different assumptions about power usage effectiveness, the pace of GPU efficiency gains, and the mix of hyperscaler versus enterprise demand, but every major forecaster has revised its projections upward at least once since 2024, suggesting these figures are more likely to prove conservative than excessive.

The Takeaway

The disagreement between the IEA, Goldman Sachs, and EPRI is not really a disagreement about whether AI will dramatically increase electricity demand by 2030. All three agree it will. The disagreement is about exactly how dramatic, and that gap, hundreds of terawatt hours wide, is itself the most important planning signal in this entire body of research.

A forecasting range this wide, produced by three of the most credible energy research organizations in the world, is not a reason to dismiss any single number. It is a reason to plan for the upper half of every published range, given how consistently these same organizations have revised their own estimates upward as actual AI adoption data has come in faster than their prior models assumed.

For utilities, regulators, and enterprises all making decisions today based on a 2030 horizon still four years away, the most defensible position is not betting on the lowest number in the range, nor assuming the highest number is alarmist. It is recognizing that every forecaster studied here has been wrong in the same direction, underestimating demand, and planning infrastructure commitments accordingly.