Artificial intelligence power demand in US could top 50 GW by 2030: EPRI

Power demand for artificial intelligence in the US could increase tenfold by 2030, according to the Electric Power Research Institute.

Total AI power demand is currently estimated at 5 GW but could reach more than 50 GW by 2030, according to an Aug. 11 report by the institute, known as EPRI, done in conjunction with AI benchmarking research organization Epoch AI. Rapid advances, particularly the training of large-scale "frontier models," are driving this electricity demand growth.

AI is the dominant near-term driver of data center growth, the report said. The analysis focused on the technical drivers of AI power consumption, modeling both demand trajectories for individual AI training sites and broader AI needs. By 2030, forecasts suggest that AI, including both training runs as well as AI inference, could consume over 5% of US generation capacity, according to the report, "Scaling Intelligence: The Exponential Growth of AI's Power Needs."

"Power demand from individual training runs may rival the output of major power plants, requiring new approaches to grid planning, permitting, and infrastructure investment," EPRI said.

Frontier models at issue

Frontier AI training models, which refer to the computationally intensive process of training large, advanced AI models, currently consume about 100 MW to 150 MW each. The analysis estimated that the power required for the largest frontier AI training runs will grow between two and three times annually, reaching 1 GW to 2 GW by 2028, and 4 GW to 16 GW by 2030.

"This demand would be highly significant, with the high end for a single model approaching 1% of total US power capacity," EPRI said.

AI power demand growth can be estimated using different approaches, such as looking at AI chip production projections, investment plans by leading AI companies, or assessments by datacenter and industry specialists. The EPRI analysis reviewed a number of such estimates, including its own forecast of datacenter growth, forecasts from the International Energy Agency, and semiconductor shipment projections.

The report included a projected need for more electricity to train the largest frontier AI models. In its modeling, the forecast peak power demand for Meta Platforms' planned Louisiana data center is 2 GW or more by 2030. One project, known as Stargate and sponsored by a group of companies that may include OpenAI and Microsoft, was forecast to scale up to 5 GW by 2030, but EPRI said this project may have been canceled. The report still includes a 1.2-GW Stargate data center campus in Abilene, Texas, by 2026.

Capacity constraints seen

EPRI cautioned that the individual projections did not provide robust enough evidence about future growth, especially at the end of the decade, because it would require a high growth rate in AI investments. But if the current trends continue, "AI will be a significant part of US energy sector by 2030."

Whether the projected power demand can be met remains unclear, according to EPRI. While hyperscaler investment suggests rapid growth, constraints in building transmission and new generation could hinder the growth of data centers, the report said. The analysis suggested that power demand for AI will likely reach gigawatt-scale training runs by 2028, but beyond that, scaling is less certain.

The rapid increase in power demand will have societal implications, EPRI said, with power growth posing challenges for tech companies that have already committed to using clean energy. Power constraints due to AI demand could also unlock a greater push for energy growth, disrupting traditional planning processes and possibly having environmental consequences.