U.S. data centers will consume roughly one-fifth of all electricity generated nationally by 2035, about four times the current level, according to a new BloombergNEF report published July 21, 2026. The consultancy projects that AI-driven computing will push U.S. data center capacity to nearly 200 gigawatts over the next decade, with nearly half of that capacity devoted to AI training and inference.

For context, a gigawatt is roughly the output of a large nuclear reactor or a major coal plant. Training refers to the computationally intensive process of building an AI model from scratch; inference is what happens when that trained model is actually put to work answering queries or generating output. Both are power-hungry, but training in particular demands enormous clusters of processors running for weeks or months at a stretch.

The revised BNEF estimate for 2035 electricity demand is 83% higher than what the consultancy predicted just seven months earlier in December 2025. It also continues a pattern of upward revisions across the energy forecasting community. EPRI more than doubled its 2024 estimate for data center electricity demand, and S&P raised its forecast by more than a third between October and April. The U.S. Department of Energy now states data centers could consume up to 9% of U.S. electricity generation annually by 2030, up from 4% of total load in 2023, while EPRI projects a range of 9% to 17% of national electricity by that same year EPRI.

BNEF's April 2025 baseline forecast had U.S. data-center power demand rising from almost 35 gigawatts in 2024 to 78 gigawatts by 2035. By December 2025, the consultancy issued a more aggressive scenario in which demand could reach 106 GW. The current report pushes the projection to nearly 200 GW, reflecting what BNEF describes as an aggressive AI adoption trajectory.

The regional grid implications are concrete. The PJM Interconnection, a regional transmission organization spanning Virginia to Illinois, will see 34% of its electricity go to data centers in the coming decade. ERCOT, which manages the grid covering most of Texas, will need to devote 22% of its generating capacity to data centers. By 2033, the U.S. will host 64% of global AI chips by power demand.

Globally, BNEF's aggressive scenario projects data centers will create 1,935 terawatt-hours of new electricity demand by 2033, nearly as much as India uses annually. The IEA reports that global electricity generation supplying data centers was 460 TWh in 2024. Its Base Case projects that figure growing to over 1,000 TWh in 2030 and reaching 1,300 TWh in 2035, with data centre electricity consumption roughly doubling from 485 TWh in 2025 to 950 TWh in 2030, accounting for around 3% of global electricity consumption. Under the IEA's high case, 2035 demand could exceed 1,700 TWh, roughly 45% above the Base Case. A worst-case scenario reviewed by the IEA 4E Technology Collaboration Programme projects consumption nearing 8,000 TWh by 2030.

The scale of individual facilities is also shifting. Bloom Energy's 2026 Data Center Power Report projects that by 2030, about one in five data center campuses will exceed gigawatt scale, rising to about one in three by 2035 Bloom Energy. A study hosted on SSRN projects data center electricity consumption rising from 23 TWh in 2025 to 371.8 TWh by 2035 under a high-growth scenario.

The velocity of these forecast revisions is itself the story. Multiple independent consultancies and agencies, each using different methodologies, have all moved their numbers sharply upward within the past 18 months. BNEF's 83% upward revision in a single reporting cycle is not a marginal adjustment.

The broader context here is that this reflects a structural reassessment of how much computing capacity AI workloads will pull into the built environment, and how quickly grid planners must respond to that pull. The PJM and ERCOT figures warrant particular attention. When a single load category absorbs over a third of a regional grid's electricity, it ceases to be a demand-side variable and becomes a structural feature of grid planning. Transmission upgrades, generation dispatch, and resource adequacy decisions are typically made on decadal timelines. The data center buildout is operating on a shorter cycle than that, which raises genuine questions about whether grid expansion can keep pace with rack-level power demand, regardless of where the electricity ultimately comes from.

The concentration of AI chip power demand in the U.S. also has geopolitical dimensions. Hosting 64% of global AI compute by power demand means the U.S. grid bears a disproportionate share of the energy cost of the AI transition. That concentration creates exposure, both to grid reliability events and to the political risk that other jurisdictions may impose data sovereignty or carbon-border requirements that fragment the global compute landscape.

There is reason for tempered optimism on the supply side. The same buildout driving demand is also accelerating investment in new generation, including nuclear, solar plus storage, and behind-the-meter resources, meaning power generated on-site or under private contracts that don't flow through the public grid. Bloom Energy's projection of gigawatt-scale campuses suggests that the industry is already moving toward architectures, on-site generation, and dedicated power purchase agreements that bypass traditional grid constraints. The 200 GW question is not whether the demand materializes. It is whether the generation and transmission buildout can land in the right places at the right times to meet it.