US data centers used 176 TWh of electricity in 2023, about 4.4% of the national total, after sitting near a flat ~60 TWh through 2014 to 2016 as efficiency gains offset rising compute. The generative-AI surge that followed broke that trend, and forecasts for 2030 now diverge widely: from 9% to 17% of US electricity in EPRI's February 2026 update. The open question is how much new load the grid can absorb, and how flexible that load can be.
Last updated May 28, 2026.
Total facility electricity (IT, cooling, and infrastructure): 58 TWh in 2014, 76 TWh in 2018, and 176 TWh in 2023, with a 325 to 580 TWh scenario band for 2028. The report states these points in text but does not publish a year-by-year table, so the chart shows only the named anchor points, not a continuous series. The 2014 to 2016 plateau held near 60 TWh; the slope is not constant within periods, which is why intervening years are left out.
Source: LBNL, 2024 United States Data Center Energy Usage Report (LBNL-2001637) · As of 2014 to 2023; 2028 scenario
Credible forecasters diverge widely. LBNL's range is for 2028; the others are 2030, so the horizons are not identical and are annotated per row. EPRI's February 2026 update (9% to 17%) is shown alongside its May 2024 figure (up to ~9%) to show how fast the forecasts moved. McKinsey's ~11.7% is a single-scenario point rendered as a marker. BloombergNEF is omitted here because its cleanest figure (~106 GW by 2035) is peak power on a different horizon. The dashed line is the 4.4% 2023 actual.
Source: LBNL (2028); EPRI 2024 and 2026 (2030); McKinsey (2030) · As of 2024 to 2026 forecasts
New load the existing system could absorb if the load accepts occasional curtailment, across the 22 largest balancing authorities (~95% of US load): 76 GW at 0.25% annual curtailment (~85 hours/yr), 98 GW at 0.5% (~177 hours/yr), and 126 GW at 1.0% (~366 hours/yr). The 5.0% / 215 GW rung appears only in the slide text, has no stated hours value, and is less corroborated than the lower rungs, so it is shaded differently.
Source: Norris et al., "Rethinking Load Growth," Duke Nicholas Institute (February 2025) · As of February 2025
Per-rack power has risen from ~10 kW for conventional CPU racks (the ~5 to 15 kW range) to ~40 kW for air-cooled H100-class AI racks and ~132 kW for NVIDIA's liquid-cooled GB200 NVL72. The vendor-roadmap rung is charted at its 250 kW low end; vendors project Rubin-generation racks at 250 to 900 kW and have referenced racks up to ~1 MW. These are vendor and industry reference points, not audited statistics, and the roadmap rung is a projection.
Source: Introl (OCP 2025 write-up); NVIDIA GB200 NVL72 specifications · As of Retrieved May 28, 2026
US data center electricity was a non-story for most of a decade. LBNL's 2024 report puts total facility electricity (IT, cooling, and supporting infrastructure) at about 60 TWh across 2014 to 2016, roughly flat, even as compute kept rising. Efficiency offset the growth: server virtualization, the shift of workloads into hyperscale facilities, and steadily improving power usage effectiveness held total energy nearly constant. LBNL's prior 2016 report documented this plateau, and the 2024 report uses it as the baseline.
The trend broke after 2017. LBNL dates the start of the climb to 2017, when the installed server base began growing again, and attributes the post-2017 doubling primarily to AI servers. Total facility electricity reached about 76 TWh in 2018 (1.9% of US electricity) and 176 TWh in 2023 (4.4%). The compound annual growth rate ran about 7% from 2014 to 2018, then jumped to roughly 18% from 2018 to 2023. The generative-AI surge that followed ChatGPT's November 2022 release is the most-cited accelerant, and CBRE and EPRI both date the steepest capacity and forecast increases to roughly 2023 onward.
Globally, the pattern is similar but earlier-stage. The IEA's April 2025 "Energy and AI" report estimates global data center electricity at about 415 TWh in 2024, roughly 1.5% of global electricity, growing about 15% per year, more than four times faster than total electricity demand. The US and China account for about 80% of that growth.
Commercial-market data comes mainly from CBRE's "North America Data Center Trends." One caution applies throughout: CBRE measures critical/IT power capacity in megawatts, a different unit from LBNL's total facility electricity in TWh. The MW figures below are IT/critical power, not total facility draw, and the two are not interchangeable.
Northern Virginia, centered on Ashburn and Loudoun County, is the largest data center market in the world, with about 4,040 MW of inventory at the end of 2025 per CBRE's H2 2025 edition, up roughly 37% year over year and including more than 1 GW delivered in 2025. National vacancy fell to a record-low 1.4% in H2 2025, and colocation vacancy in Northern Virginia fell to 0.5%, with demand running ahead of deliverable supply. The other primary US markets are Dallas-Fort Worth, Phoenix, Chicago, Silicon Valley, Atlanta, Hillsboro (Oregon), and the New York Tri-State area, with Phoenix and Dallas-Fort Worth the fastest-growing.
Under-construction capacity across the primary markets fell to 5,994.4 MW at the end of 2025, down from 6,350.1 MW a year earlier. CBRE describes that as the first year-over-year decline in primary-market construction since 2020, attributed largely to power-availability constraints. The spending behind the buildout has run the other direction: combined capex at Microsoft, Alphabet, Amazon, and Meta roughly doubled from about $160B in 2023 to about $300B+ in 2025, the large majority for AI data centers and compute. Those capex totals mix data centers, chips, networking, and other spending, so they are a directional signal rather than a clean data center figure.
Forecasts for US data center electricity diverge by a wide margin. On annual energy, LBNL's December 2024 report projects 325 to 580 TWh by 2028, or 6.7% to 12.0% of 2028 US electricity. McKinsey models US data center energy rising to about 606 TWh by 2030, roughly 11.7% of US power demand in that scenario. McKinsey's 2023 base of 147 TWh differs from LBNL's 176 TWh because of different boundary definitions, which is one reason forecasts are hard to line up directly.
EPRI's figures show how fast the projections are moving. Its May 2024 "Powering Intelligence" study put data centers at up to about 9% of US electricity generation by 2030, across four scenarios spanning 4.6% to 9.1% of US load. Its February 2026 update revised that to 9% to 17% by 2030, about 60% higher, citing accelerated buildout over the prior 18 months. The 2026 figure is the current EPRI number and supersedes the 2024 one.
On peak power rather than annual energy, BloombergNEF's December 2025 outlook puts US data center power demand at about 106 GW by 2035, a 36% upward revision from its April 2025 figure, reaching roughly 8.6% of US electricity by 2035 versus about 3.5% today. Grid Strategies' December 2025 "National Load Growth Report" finds that US utility 5-year peak load forecasts now total about +166 GW by ~2030, roughly a 20% increase over the 2025 peak and about six times the growth utilities were forecasting in 2022. Grid Strategies attributes roughly 90 GW of that to data centers. Much of the new transmission needed to serve this load is covered in Blizzard Power's transmission focus topic.
The high-end utility numbers are widely believed to overstate real demand, though the magnitude is contested. Grid Strategies names two mechanisms. First, double-counting: a single project shops multiple utility jurisdictions for the best deal, so one data center can appear in several utilities' queues and forecasts at once. Second, speculative requests: queues include early-stage, non-binding requests with low odds of being built. Comparing utility forecasts to bottom-up technology-bottleneck analysis, Grid Strategies estimates utilities may be overstating data center demand by as much as ~40%. The defensible framing is a range: the bottom-up floor (LBNL, IEA) and the aggregated-interconnection-request ceiling can differ by a factor of two or more, and where actual demand lands depends on how much of the queue gets built and powered.
LBNL groups facilities by space type. Hyperscale facilities are the largest, operated by cloud and AI providers such as AWS, Microsoft, Google, Meta, and Oracle, with single campuses now planned at hundreds of MW to over 1 GW. They account for the bulk of recent growth. Colocation facilities (large-scale and small/medium) are multi-tenant sites run by operators such as Equinix, Digital Realty, and QTS that lease space and power to enterprises and to hyperscalers; large-scale colocation is the second major growth category. Enterprise or internal data centers are privately owned single-company facilities, a declining share as workloads move to the cloud. Edge facilities are small, geographically distributed sites for low-latency workloads, and a small share of total energy.
AI workloads split into two modes with different grid implications. Training builds or refines a model: it is large, scheduled, batch-oriented, and somewhat delay-tolerant, can run in concentrated multi-week campaigns, and can be sited remotely. That delay-tolerance is the basis for the temporal flexibility discussed below. Inference runs a trained model to answer queries: it is real-time, latency-sensitive, more like conventional cloud load, and less flexible.
AI workloads combined are projected at 50% to 70% of data center demand by 2030, per the Norris presentation citing the LBNL data and corroborated by EPRI and IEA framing. EPRI has cited an AI query as drawing roughly 10 times the electricity of a traditional web search, an order-of-magnitude estimate rather than a precise measurement.
The driver is the chips. AI training and inference run on GPUs and accelerators (NVIDIA H100, B200, and GB200; AMD MI300; Google TPU) that draw far more power per chip and per rack than CPUs. Rack density has risen from about 5 to 15 kW for conventional CPU racks to roughly 40 kW for air-cooled H100-class AI racks and about 132 kW for NVIDIA's GB200 NVL72, which requires liquid cooling. Vendor roadmaps project Rubin-generation racks at 250 to 900 kW and have referenced racks up to about 1 MW. These rack figures are vendor and industry reference points, not audited statistics.
More chip power means more heat to remove, and cooling is the largest non-IT energy use. At GB200-class densities, air cooling is insufficient and liquid cooling (direct-to-chip or immersion) becomes necessary. The standard efficiency metric is Power Usage Effectiveness (PUE), defined as total facility energy divided by IT equipment energy. A PUE of 1.0 is the theoretical ideal, where all energy goes to computing; a PUE of 1.5 means 50% overhead, mostly cooling and power conversion, on top of the IT load. The industry average was about 1.56 in the Uptime Institute's 2024 Global Data Center Survey, roughly flat for five consecutive years, while hyperscale leaders report PUE near 1.1.
Data centers also run near-continuously, so they have a high load factor at the facility level. That makes them attractive to utilities as steady, baseload-like revenue and demanding for the grid because they add to peak as well as off-peak. The grid they connect to does not run flat: Norris's "Rethinking Load Growth" documents US balancing authorities operating at about a 53% average load factor (range 43% to 61%), which means substantial unused capacity most of the year. That gap is the empirical basis for the flexibility argument below.
The most-cited rigorous state analysis is the Virginia Joint Legislative Audit and Review Commission's "Data Centers in Virginia" study, Report 598, published December 9, 2024, with modeling by the consultancy E3. It finds that unconstrained data center growth could raise Virginia generation and transmission costs by as much as $18 billion by 2040, with costs shared by all ratepayers under current rate structures, and that a typical Dominion residential customer could see generation and transmission costs rise an estimated $14 to $37 per month in constant dollars by 2040. JLARC also found that current rates appropriately allocate costs to the customers responsible, but that growth still raises system costs for all customers because new generation and transmission must be built. Much of that transmission cost is the buildout covered in Blizzard Power's transmission focus topic.
The same study sets the economic contribution against those costs. Virginia's data center sales-tax exemption provided $928 million in tax savings in FY23, and about 90% of the industry uses it. The industry contributes roughly 74,000 jobs, $5.5B in labor income, and $9.1B in GDP annually in the state. On water, JLARC found data centers typically use about the same water as a large office building or less, though some use substantially more.
Local opposition and utility responses have produced concrete actions. Tucson's City Council voted unanimously in August 2025 to oppose "Project Blue," a proposed ~$3.6B Amazon data center campus, after sustained concern over water and ratepayer cost. In Ohio, after confirming about 5 GW of new data centers and receiving roughly 30 GW in requests, AEP issued a temporary moratorium on data center service requests in 2023, then reached a settlement adding longer contracts, load-ramp schedules, minimum demand charges, and collateral for loads above 25 MW. In Georgia, after about 7.3 GW of large-load customers committed to Georgia Power, the Georgia PSC changed contract provisions to add PSC review, longer contracts, and minimum billing for cost recovery.
Aggregate opposition figures come from advocacy and tracking outlets and should be read as such. Data Center Watch reports about $64 billion of data center projects blocked or delayed amid local opposition, and trackers count moratoria in at least 14 states. These are tracker estimates, not official counts. On regional ratepayer exposure, Synapse Energy Economics projected that PJM consumers could pay about $100 billion extra through 2033 as data center load outpaces supply, a single modeling estimate cited via FERC colocation coverage. On water, US data centers consumed an estimated ~17 billion gallons directly for cooling in 2023, a figure attributed to EESI and related reporting drawing on LBNL and company disclosures.
The central finding in this area comes from Tyler Norris and co-authors at Duke's Nicholas Institute, in "Rethinking Load Growth" (February 2025). If a new large load accepts occasional curtailment, the existing US grid can absorb far more of it than the construction pace suggests, because most of the year the system runs well below its peak.
Across the 22 largest US balancing authorities, which account for about 95% of US load, Norris estimates the grid could host 76 GW of new load at 0.25% annual curtailment (about 85 hours per year), 98 GW at 0.5% (about 177 hours per year), and 126 GW at 1.0% (about 366 hours per year). A 5.0% / 215 GW figure appears in the slide text but is less corroborated than the lower rungs and has no stated hours value. The enabler is the roughly 53% average load factor noted above: there is large unused capacity most of the year. The headroom ladder chart above shows these rungs.
The curtailment is shallow as well as rare. Norris finds that 88% of the hours requiring curtailment retain at least half of the new load, 60% retain at least 75%, and 29% retain at least 90%. So even in the hours when the new load is curtailed, most of it usually keeps running. Norris frames the result as a first-order estimate and is running a more detailed production-cost simulation as follow-up.
Norris groups flexibility into four types. On-site power and storage means co-located batteries, renewables, or generators. Temporal flexibility means scheduling delay-tolerant compute, such as model training and batch jobs, before or after high-stress periods. Spatial flexibility means shifting workloads across data centers in different geographies and grids. Reduced operations means planned workload reduction during defined windows. Training's delay-tolerance, noted earlier, is what makes the temporal and spatial options workable.
EPRI's DCFlex initiative, launched in October 2024, is a roughly 3-year program to test data center demand flexibility in real-world conditions, targeting 5 to 10 flexibility hubs. It has grown from 14 members at launch to about 45 collaborators, including Google, Meta, NVIDIA, PG&E, PJM, ERCOT, Microsoft, and Oracle. A Phoenix demonstration tested whether a data center could sustain a 25% power reduction during grid stress via a 15-minute ramp-down over a 3-hour event. Google has integrated about 1 GW of data center demand response with US utilities and participates in carbon-aware temporal and geospatial workload shifting. ERCOT has established a Large Flexible Load Task Force and an interim process to study loads as flexible "Controllable Load Resources," and PG&E runs a "Flex Connect" pilot offering faster interconnection to large loads in exchange for flexibility during constraint.
A growing number of AI data centers are installing on-site generation, mostly natural gas turbines and some fuel cells, to avoid multi-year interconnection waits. One trade-press analysis identified about 47 buildouts using on-site generation, roughly 23 GW of it (about 75%) natural-gas-powered, with equipment from GE Vernova, Siemens, Wartsila, and Bloom Energy. Both responses, curtailment and on-site generation, answer the same constraint: interconnection delay, which Blizzard Power covers in its interconnection queue focus topic.
Regulators are still writing the rules for how these loads connect at power plants. In November 2024, FERC rejected an amended interconnection service agreement that would have expanded power sales to a colocated Amazon data center at Talen's 2,475 MW Susquehanna nuclear plant, and directed PJM to develop colocation rules. On December 18, 2025, FERC issued a unanimous order directing PJM to establish colocation rules for data centers and other large loads at power plants. In June 2025, Talen and AWS had restructured the arrangement into a grid-connected, front-of-the-meter 17-year PPA worth about $18B for up to 1,920 MW from Susquehanna.
More articles on data centers and grid economics coming soon.
Two units, kept separate: Two metrics appear on this page and are not interchangeable. LBNL reports total facility electricity (IT, cooling, and infrastructure) in TWh; the 176 TWh / 4.4% and 325 to 580 TWh figures are this metric. CBRE reports critical/IT power capacity in MW; the ~4,040 MW Northern Virginia inventory and the under-construction figures are this metric. The Norris and forecast-GW figures are peak power in GW. No chart mixes TWh and GW on a single axis.
The LBNL time series: LBNL-2001637 presents its annual series only as figures, not as a year-by-year numeric table. The electricity chart on this page therefore plots only the discrete points the report states in text (58 TWh in 2014, 76 TWh in 2018, 176 TWh in 2023, and the 325 to 580 TWh range for 2028) and does not interpolate intervening years, because the growth rate is not constant within periods (about 7% per year 2014 to 2018, about 18% 2018 to 2023). The 2014 to 2016 plateau is shown as a single 2014 anchor with the plateau noted in the caption.
Forecast horizon mismatch:The forecast-spread chart uses share of US electricity as the common metric, but LBNL's horizon is 2028 while EPRI and McKinsey are 2030. That difference is annotated per row. EPRI 2024 (up to ~9%) is shown alongside EPRI 2026 (9% to 17%) to make the revision visible; the 2026 figure is current. McKinsey's ~11.7% is a single-scenario point rendered as a thin marker. BloombergNEF was kept off this chart because its cleanest figure (~106 GW by 2035) is peak power on a 2035 horizon, which would mislead on a 2030 share axis.
The "phantom load" caveat:The aggregated utility forecast (+166 GW by ~2030, Grid Strategies, December 2025) is the ceiling, not a consensus. Grid Strategies itself estimates utilities may be overstating data center demand by as much as ~40%, driven by the same project appearing in multiple jurisdictions' queues (double-counting) and by speculative, non-binding requests. The page shows the range rather than picking a point estimate.
Rack-density figures: The conventional-CPU (~10 kW), air-cooled H100 (~40 kW), GB200 NVL72 (~132 kW), and vendor-roadmap (250 to 900 kW, up to ~1 MW) points are vendor and industry reference points from Introl and NVIDIA, not government data or audited statistics. The roadmap rung is a projection, charted at its 250 kW low end so no single invented number stands in for the band.
Norris method:"Rethinking Load Growth" adds a constant new load in all hours, curtails it only when it would exceed the historical seasonal peak, and defines the curtailment rate as annual curtailed MWh divided by the new load's maximum potential annual consumption, across the 22 largest balancing authorities (~95% of US load). The headroom ladder and hours-per-year values are read from the ISO-NE CLG presentation, slides 16 to 17. The 5.0% / 215 GW figure appears only in slide text, has no stated hours value, and is less corroborated than the 76/98/126 GW rungs, so it is shaded differently on the chart.
Advocacy-sourced figures: The aggregate opposition figures (~$64 billion of projects blocked or delayed, moratoria in at least 14 states) come from advocacy and tracking outlets, mainly Data Center Watch, and are attributed as tracker estimates rather than official counts. The Synapse ~$100 billion PJM figure is a single modeling estimate cited via news coverage. The Ohio, Georgia, and Tucson examples are from the Norris presentation and news coverage.
Last updated: May 28, 2026.
This page is for informational purposes only and does not constitute investment, legal, or engineering advice. Hyperscaler capex totals mix data centers, chips, networking, and other spending and are a directional signal, not a clean data center figure. Construction-status and capacity figures reflect the most recent public reporting available as of the last-updated date and may change.