# BNEF’s 194-GW Data Center Forecast Shows AI’s Power Pipeline Outrunning the Grid > Source: > Published: 2026-07-23 14:53:53.157627+00:00 # BNEF’s 194-GW Data Center Forecast Shows AI’s Power Pipeline Outrunning the Grid - BloombergNEF now projects US data center power capacity will reach 194 GW in 2035, an 83% increase from its December forecast and roughly 2.5 times its April 2025 estimate. [[1]](https://www.tomshardware.com/tech-industry/data-centers/bnef-nearly-doubles-its-us-data-center-power-forecast-to-194gw) - Nearly half of the forecast capacity would support AI training and inference, while data centers would consume about 20% of US electricity by 2035, according to reporting on BNEF’s outlook. [[2]](https://heatmap.news/sparks/bnef-data-centers-electricity) - The forecast is derived from BNEF’s asset-level project pipeline and infrastructure assumptions, not 194 GW of facilities already under construction. One of the report’s authors said its completion assumptions may prove too generous. [[3]](https://www.latitudemedia.com/news/bnef-nearly-doubled-its-forecast-for-us-data-center-power-demand/)[[4]](https://about.bnef.com/insights/category/data-centers/) - PJM and ERCOT confirm that large-load requests are reshaping grid planning, but both are revising their methods because duplicate, speculative and early-stage projects can distort forecasts. [[5]](https://www.pjm.com/-/media/DotCom/library/reports-notices/load-forecast/2026-load-report.pdf)[[6]](https://www.ercot.com/files/docs/2026/04/01/ERCOT_LargeLoad_Update_April2026_B-C_-Hearing.pdf) BloombergNEF has raised its forecast for US data center power capacity in 2035 to 194 gigawatts, up 83% from the 106 GW it projected in December 2025. That earlier figure was already 36% above BNEF’s April 2025 estimate of 78 GW, meaning the outlook has increased roughly 2.5-fold in 15 months.[[1]](https://www.tomshardware.com/tech-industry/data-centers/bnef-nearly-doubles-its-us-data-center-power-forecast-to-194gw) The revision is evidence of how quickly planned AI infrastructure has expanded, rather than proof that every proposed campus will be built. More than 100 GW entered BNEF’s development pipeline during the first part of 2026, and nearly half of its projected capacity would run AI training and inference. If developers installed all the high-powered AI chips BNEF expects to be available through 2035, the requirement could reach 229 GW.[[2]](https://heatmap.news/sparks/bnef-data-centers-electricity) BNEF maintains an asset-level database tracking data center capacity, ownership and project status, then models demand using pipeline visibility and infrastructure buildout. Mark Daly, the firm’s head of technology and innovation and a co-author of the latest report, said its use of historical completion rates could prove too generous because an increasing share of proposals comes from first-time developers facing shortages of equipment, labor and grid connections, as well as local opposition.[[3]](https://www.latitudemedia.com/news/bnef-nearly-doubled-its-forecast-for-us-data-center-power-demand/)[[4]](https://about.bnef.com/insights/category/data-centers/) ## A forecast driven by projects, chips and AI workloads The latest estimate extends a broad upward shift among power-demand forecasters. The Electric Power Research Institute’s February 2026 outlook places data centers at 9% to 17% of US electricity consumption by 2030. Its low case assumes projects under construction and one-quarter of those in advanced planning are operational by then; its high case includes all projects under construction or in advanced planning and 30% of those in early planning.[[7]](https://powering-intelligence.epri.com/executive-summary.html) Lawrence Berkeley National Laboratory’s earlier bottom-up model estimated US data center consumption of 325 to 580 terawatt-hours in 2028, equal to 6.7% to 12% of national electricity use. Unlike EPRI’s commercial-project approach, the LBNL scenarios were built partly from projected shipments of servers, storage and networking equipment.[[7]](https://powering-intelligence.epri.com/executive-summary.html)[[8]](https://eta-publications.lbl.gov/publications/2024-lbnl-data-center-energy-usage-report) S&P Global’s 451 Research estimated in April that the full-load grid power requirement of hyperscale, leased and cryptocurrency data centers could reach 183.2 GW in 2030, up from its June 2025 forecast of 117.1 GW. It put the comparable 2025 figure at 64.4 GW.[[9]](https://www.spglobal.com/market-intelligence/en/news-insights/articles/2026/5/surging-us-data-center-power-demand-tests-sustainability-targets-101100707) Those numbers should not be treated as a clean ranking. Forecasts measure different things: connected or nameplate capacity, peak demand, average power, or annual electricity consumption. They also make different assumptions about utilization, cooling overhead, project cancellations, chip efficiency and construction timelines. A June review from Columbia University identified a spread of more than 100 GW among published 2030 peak-demand estimates and attributed much of the variation to methodological differences, including the use of utility queues versus individually tracked projects.[[10]](https://www.energypolicy.columbia.edu/wp-content/uploads/2026/06/ElectricityLoad-CGEP_WhitePaper_062326.pdf) BNEF’s April 2025 outlook illustrates the distinction. It projected 78 GW of US data center power capacity in 2035 but average hourly demand of 49 GW, reflecting that facilities do not continuously draw their maximum rated power. The firm also estimated that a US data center typically took about seven years to progress from early development to full operation.[[11]](https://about.bnef.com/insights/commodities/power-for-ai-easier-said-than-built/) The latest increase therefore says as much about the quantity and scale of proposed campuses as it does about eventual electricity consumption. BNEF expects non-hyperscaler capacity to nearly quintuple over the coming decade, while hyperscaler capacity almost triples. First-time developers account for much of the incremental pipeline, increasing the probability of delays, financing failures and duplicate site requests.[[2]](https://heatmap.news/sparks/bnef-data-centers-electricity)[[3]](https://www.latitudemedia.com/news/bnef-nearly-doubled-its-forecast-for-us-data-center-power-demand/) ## The national total hides sharper regional pressure BNEF projects that data centers will consume about 12% of US electricity in 2030 and 20% in 2035, compared with 5.9% currently. The regional concentration is more severe: the firm estimates data centers could account for 34% of electricity use in PJM’s territory, while the corresponding share in the main Texas grid operated by ERCOT could reach 22% by 2035.[[1]](https://www.tomshardware.com/tech-industry/data-centers/bnef-nearly-doubles-its-us-data-center-power-forecast-to-194gw)[[12]](https://techcrunch.com/2026/07/21/data-centers-expected-to-use-4x-more-electricity-by-2035/) PJM’s own planning documents support the direction, if not the precise magnitude, of BNEF’s forecast. The grid operator expects its overall summer peak to rise to 222.1 GW in 2036, an increase of 65.7 GW over ten years, while annual net energy demand rises by about 582 terawatt-hours. PJM made explicit data center adjustments in many of its utility zones.[[5]](https://www.pjm.com/-/media/DotCom/library/reports-notices/load-forecast/2026-load-report.pdf) PJM does not simply accept every request submitted by a utility. Its 2026 process requires firmer service or construction commitments for near-term additions and discounts longer-dated projects to account for uncertainty. That tighter screening reduced several near-term forecasts from the previous year even as the longer-term growth rate remained far above historical levels.[[13]](https://insidelines.pjm.com/pjms-updated-20-year-forecast-continues-to-see-significant-long-term-load-growth/) Texas presents the queue problem in its most extreme form. ERCOT said in April that it was tracking about 410 GW of large-load interconnection requests, of which approximately 87% were data centers. That queue total is almost five times ERCOT’s standing system peak record, but it is not a forecast of completed projects.[[6]](https://www.ercot.com/files/docs/2026/04/01/ERCOT_LargeLoad_Update_April2026_B-C_-Hearing.pdf) ERCOT separately described its preliminary 367.8-GW load forecast for 2032 as a planning snapshot rather than a prediction and said it was probably higher than the demand that will ultimately materialize. The forecast combines ERCOT’s base economic model with information supplied by transmission and distribution companies working with prospective large customers.[[14]](https://www.ercot.com/news/release/04152026-ercot-releases-preliminary) The concentration of requests can create a circular planning problem. Developers submit projects in multiple locations to find the fastest available connection; utilities incorporate some of those requests into forecasts; and grid operators may then plan generation and transmission around demand that is duplicated or too early to finance. At the same time, dismissing the queues altogether would miss real campuses already securing land, equipment and power contracts. ## Connection speed has become part of the forecast BNEF estimates that the most US data center capacity connected in a single year has been 7.1 GW. Even if that record pace continued annually through 2035, its base case would still face a 19-GW gap between expected facilities and available grid or on-site supply.[[1]](https://www.tomshardware.com/tech-industry/data-centers/bnef-nearly-doubles-its-us-data-center-power-forecast-to-194gw) That constraint changes which projects are likely to reach operation. Sites near existing substations, power plants or underused transmission infrastructure can advance faster than campuses that require major network construction. Some developers are pursuing behind-the-meter generation or flexible service arrangements rather than waiting for an unrestricted grid connection.[[3]](https://www.latitudemedia.com/news/bnef-nearly-doubled-its-forecast-for-us-data-center-power-demand/)[[15]](https://www.ferc.gov/news-events/news/ferc-launches-aggressive-targeted-action-speed-large-load-integration) Federal regulators are now treating large-load connections as a distinct category of grid planning. In June, the Federal Energy Regulatory Commission ordered six regional grid operators to justify or revise their rules for connecting data centers and other large customers. The proceedings cover application and study processes, cost shifting, behind-the-meter generation, flexible transmission service and the joint study of large loads with nearby generation.[[15]](https://www.ferc.gov/news-events/news/ferc-launches-aggressive-targeted-action-speed-large-load-integration) FERC Commissioner David Rosner said speculative requests can inflate forecasts through duplication and projects that never materialize. He described escalating readiness requirements intended to deter such requests and cost-recovery agreements designed to keep other customers from paying for infrastructure built for a data center that fails to appear.[[16]](https://www.ferc.gov/news-events/news/commissioner-rosners-remarks-large-load-show-cause-orders-e-7-e-12-june-18-2026) These rules matter to the BNEF estimate because interconnection is no longer a final administrative step after a data center has been designed. Available power increasingly determines the site, construction schedule, generation technology and commercial viability of the entire project. A forecast based primarily on announced computing demand will run higher than one constrained by substations, turbines, transformers, permits and transmission construction. ## The revision is a planning range, not a settled outcome The strongest conclusion from the new forecast is that utilities must prepare for a much wider range of demand than they considered plausible in early 2025. BNEF’s 194-GW base case may overshoot if first-time developers fail, AI hardware efficiency rises faster than expected or training demand slows. It may undershoot if inference spreads rapidly through consumer and enterprise products and developers overcome connection bottlenecks with on-site generation. Efficiency alone does not guarantee lower electricity use. More efficient chips reduce the power required for a given computation, but they can also make additional training and inference economical. The resulting demand depends on model development, AI adoption, chip shipments and the fraction of available hardware developers can actually energize. For grid planners, the costly errors sit on both sides. Underbuilding generation and transmission can delay viable data centers and tighten reliability margins. Overbuilding around duplicate or speculative requests can leave other customers paying for unused infrastructure. PJM’s project derating, ERCOT’s effort to batch large-load studies and FERC’s proposed readiness and cost-recovery safeguards are attempts to narrow that range before capital is committed.[[6]](https://www.ercot.com/files/docs/2026/04/01/ERCOT_LargeLoad_Update_April2026_B-C_-Hearing.pdf)[[13]](https://insidelines.pjm.com/pjms-updated-20-year-forecast-continues-to-see-significant-long-term-load-growth/)[[16]](https://www.ferc.gov/news-events/news/commissioner-rosners-remarks-large-load-show-cause-orders-e-7-e-12-june-18-2026) The 194-GW figure is best read as a measure of the AI industry’s current appetite for power under BNEF’s project-completion assumptions. Whether the United States reaches it will depend less on the number of campus announcements than on how many projects secure credible customers, equipment, generation and a firm place on the grid. ## Companies mentioned ## Further sources [[1] Tom's Hardware, “Data centers forecast to use 20% of US power by 2035,” July 22… ↗](https://www.tomshardware.com/tech-industry/data-centers/bnef-nearly-doubles-its-us-data-center-power-forecast-to-194gw) [[2] Heatmap News, “Data Centers Will Use Enough Electricity to Power Every U.S. Hou… ↗](https://heatmap.news/sparks/bnef-data-centers-electricity) [[3] Latitude Media, “BNEF nearly doubled its forecast for US data center power dema… ↗](https://www.latitudemedia.com/news/bnef-nearly-doubled-its-forecast-for-us-data-center-power-demand/) [[4] BloombergNEF, Data Centers research and methodology overview. ↗](https://about.bnef.com/insights/category/data-centers/) [[5] PJM Interconnection, 2026 Load Forecast Report, January 2026. ↗](https://www.pjm.com/-/media/DotCom/library/reports-notices/load-forecast/2026-load-report.pdf) [[6] ERCOT, Large Load Update to the Texas Senate Committee on Business and Commerce… ↗](https://www.ercot.com/files/docs/2026/04/01/ERCOT_LargeLoad_Update_April2026_B-C_-Hearing.pdf)+10 more The stories that matter, in one email. Free — unsubscribe anytime.