Home / Research / AI Data Center Power Statistics

AI Data Center Power Statistics: 2026 Report

By Darin ten Bruggencate, Founder, AI Industry Reviews · September 14, 2026 · Updated October 27, 2026

Our analysts compiled power figures covering 2007 to 2026, plus forecasts to 2035, from 23 sources, including NVIDIA, Uptime Institute, Lawrence Berkeley National Laboratory, the Energy Information Administration, PJM Interconnection, and utility filings with the Federal Energy Regulatory Commission, across five levels from a single AI server to the US grid. The aim was to show how fast AI data center power demand is rising as of Q4 2026.

In the following sections, we compare the power ratings of NVIDIA AI servers, the rack densities of AI and typical data centers, the power buildings use for cooling, US data center electricity use through 2030, US commercial electricity sales through 2027, grid forecasts and connection queues, and the US share of global data center electricity.

Power Ratings of NVIDIA AI Servers

We reviewed the maximum power and heat output that NVIDIA publishes in its DGX user guides and GPU product pages for three generations of eight-GPU DGX servers, and the maximum power of the power supply built into a GB200 NVL72 rack, because a data center has to supply and cool the rated power of every server it installs.

NVIDIA's GB200 NVL72 is a whole-rack system holding 72 GPUs across 18 compute trays, fed by eight power shelves that can each deliver up to 33 kW [20]. That is 264 kW of power-shelf capacity in one rack, well above the 132 kW per rack that Schneider Electric's GB200 design plans for [22].

The International Energy Agency (IEA) estimates that the power density of AI servers increased 11 times from 2020 to 2025 and is set to rise about four times more by 2027 [7].

Maximum power ratings of NVIDIA DGX AI servers and GPUs

Source: NVIDIA A100 and H100 product pages [18][19]; NVIDIA DGX A100, DGX H100/H200, and DGX B200 user guides [15][16][17], where GPU figures for A100 and H100 are maximum thermal design power for the SXM versions and the B200 figure is the maximum power per GPU with all power supplies running.

Key Findings

• The DGX B200 is rated at 2.2 times the maximum power of the DGX A100, and each B200 GPU at 2.5 times the A100 SXM GPU. Each new system also takes more rack space, from 6 rack units (6U) for the DGX A100 to 10 rack units (10U) for the DGX B200, which limits how many fit in one rack.

AI Rack Density Compared With Typical Data Center Racks

We compared the average rack density of NVIDIA AI rack designs from Schneider Electric with the typical racks in Uptime Institute's operator surveys, because the gap shows how much more power and cooling an AI hall needs than the halls most operators run today.

Schneider Electric reports that general-purpose server racks went from about 3 kW to around 10 kW over 25 years [21]. Its published retrofit designs for existing buildings reach 40 kW per rack with air cooling and 73 kW with liquid cooling, and its new-build designs reach 132 kW for GB200 NVL72 and 142 kW for GB300 NVL72 [22]. Uptime reports that 27% of its survey respondents work in data centers more than 15 years old [8].

Average rack density of NVIDIA AI rack designs and of typical data center racks, 2022 to 2026

Source: Schneider Electric blog, July 23, 2025 [21]; Uptime Institute Global Data Center Survey 2025 [9] and 2026 [8], where Uptime's figures are the average of each operator's most common rack density (n=679 in 2026) and high-density sites are those whose most common rack is 30 kW or more.

Key Findings

• NVIDIA's DGX SuperPOD rack density rose 5.3 times from the A100 generation in 2022 to the GB200 in 2025, and a GB200 NVL72 rack at about 132 kW needs about 17 times the 7.8 kW that the typical rack at sites below 30 kW averaged in 2026 [22][8]. Over 2024 and 2025, the two years both series cover, the average typical rack in Uptime's survey rose from 8.3 kW to about 9 kW [9].

Uptime reports that the racks of 50 kW or more include some above 100 kW [8].

Highest rack density reported by data center operators, 2023 to 2026

Source: Uptime Institute Global Data Center Survey 2026, Figure 4 [8], with figures rounded by Uptime so they may not add to 100%.

Key Findings

• In 2026, 24% of operators reported having at least some racks at 30 kW or more, up from 19% in 2025 [8]. Most of the increase came from racks of 50 kW or more, whose share rose from 9% to 13%.
• The share of operators whose densest rack is below 30 kW fell 6 percentage points from 2025 to 2026, after falling 4 points over the two years before. High-density racks are spreading to more sites, and three in four operators still run nothing above 30 kW.

Power Used for Cooling and Other Building Systems

We reviewed the average annual power usage effectiveness (PUE) that data center operators reported to Uptime Institute from 2007 to 2026, because it shows how much extra power a building draws for cooling and other systems on top of what its servers use, and a utility has to supply that extra power too.

Uptime reports that the average PUE of respondents' largest data centers was 1.45 in 2026, and that 23% of respondents work in a data center that averages below 1.3 [8].

Average annual PUE reported by data center operators, 2007 to 2026

Source: Uptime Institute Global Data Center Survey 2026, Figure 2 (n=644 in 2026) [8], with non-IT share calculated as (PUE minus 1) divided by PUE.

Key Findings

• Average PUE fell from 1.67 in 2019 to 1.52 in 2026, and has stayed in a range of 1.52 to 1.59 since 2020 [8].
• At the 2026 average, about a third of a data center's power goes to equipment other than servers, storage, and networking. Gains in building efficiency have slowed since the 2010s, so each kilowatt of server power now comes with about 1.5 kilowatts of total facility demand.

US Data Center Electricity Use and US Commercial Electricity Sales, 2014 to 2030

We reviewed the yearly US data center electricity use that Lawrence Berkeley National Laboratory (LBNL), in its December 2024 report to Congress, and EPRI, in its February 2026 update, both estimate and project to the end of the decade, because utilities, regulators, and developers plan new supply around how fast that use is expected to grow. LBNL builds its estimates from shipments of servers and other equipment, and EPRI builds its estimates from data on data center sites in operation, under construction, and planned [4].

EPRI puts its revised 2030 estimate about 60% above the estimates in its 2024 report and attributes the revision mainly to record data center development over the previous 18 months [4]. The 2024 report's scenarios put 2030 use at 196.3 TWh to 403.9 TWh, or 4.6% to 9.1% of US generation [5].

EPRI estimates that Virginia is the only state where data centers use more than 20% of electricity today, and that their share there could reach 39% to 57% by 2030 [4].

US data center electricity use and share of US electricity, 2014 to 2030

Source: LBNL, 2024 United States Data Center Energy Usage Report [2]; US Department of Energy release on the LBNL report [3]; EPRI, Powering Intelligence 2026 [4], where EPRI's estimates include cryptocurrency mining and LBNL's exclude it.

Key Findings

• LBNL calculates that US data center electricity use grew about 18% a year from 2018 to 2023 [2]. Its 2028 range is 1.8 to 3.3 times the 2023 level.
• EPRI's 2030 projection of 380 to 790 TWh is 2.0 to 4.5 times its 2024 estimate of 177 to 192 TWh, and its share of US electricity rises from 4% to 5% in 2024 to 9% to 17% in 2030 [4].

US data center capacity and peak load, 2024 and 2030 (EPRI)

Source: EPRI, Powering Intelligence 2026, Executive Summary (last updated February 25, 2026) [4], where nominal IT capacity is the maximum power servers and other IT equipment could draw, excluding cooling.

Key Findings

• In EPRI's Medium scenario, US data center peak load in 2030 is 3.2 to 3.4 times its 2024 level [4]. Its nominal IT capacity rises from 35 to 44 GW in 2024 to a range of 56 to 132 GW in 2030 across the Low to High scenarios [4].

EPRI's Low scenario assumes most projects under construction and one-fourth of those in advanced planning come online, and its High scenario assumes all projects under construction or in advanced planning plus 30% of those in early planning do [4].

EPRI estimates that AI workloads account for 15% to 25% of data center electricity today [4]. EPRI notes that peak load comes in below nominal IT capacity, and that cooling and other building loads, the time new buildings take to fill, daily load patterns, on-site power, and load flexibility all affect how much of that capacity shows up as peak demand [4].

We also compared EIA's October 2026 forecast of commercial electricity sales with its forecast of total US electricity consumption, because faster growth in commercial sales than in total use would show new demand gathering in the commercial sector, which includes data centers.

US Commercial Electricity Sales vs Total Consumption, 2025 to 2027

Source: EIA, October 2026 Short-Term Energy Outlook [14].

Key Findings

• The EIA projects commercial electricity sales to grow 6.6% from 2025 to 2027, against 3.8% for total US electricity consumption [14].

Grid Forecasts and Connection Queues

We reviewed the utility forecasts of US peak demand filed with FERC since 2022, to see how they have changed from one year's filings to the next, along with PJM Interconnection's load forecast, Dominion Energy's data center contracts, and LBNL's data on how long new power plants take to connect, because a new data center cannot run until the grid can supply and connect it.

In the 2025 forecasts, US summer peak demand rises from 829 GW in 2025 to 995 GW in 2030, a 3.7% annual rate [10]. Grid Strategies links roughly 90 GW of the 166 GW increase to data centers [10].

Grid Strategies estimates that the data center portion of utility forecasts is overstated by roughly 25 GW, based on market analyst reports, and notes that TD Cowen projected 65 GW of new data center demand by 2030 from expected chip shipments [10]. Grid Strategies says utility forecasts need to better reflect the probability that projects are completed, their total loads, supply constraints, and the timing of load growth [10].

The average proposed US data center doubled in size from 150 MW in 2023 to 300 MW in 2024, while the average operating data center draws roughly 30 to 60 MW [10]. Sixteen campuses of a gigawatt or more, with nearly 30 GW of combined demand, are scheduled to come online in 2026 and 2027 [10].

Five-year US summer peak growth in utility forecasts filed with FERC, 2022 to 2025

Source: Grid Strategies, Power Demand Forecasts Revised Up for Third Year Running, Led by Data Centers (November 2025), aggregating forecasts submitted to FERC [10].

Key Findings

• Forecast five-year peak growth rose from 24 GW in the 2022 filings to 166 GW in the 2025 filings, nearly seven times as much [10].

Grid operator, utility, and interconnection queue figures

Source: PJM 2026 Long-Term Load Forecast Report [12]; Dominion Energy second-quarter 2026 earnings call presentation, July 31, 2026 [13]; LBNL, Queued Up: 2025 Edition [11].

Key Findings

• PJM forecasts a summer peak 42.0% higher in 2036 than in 2026 [12].
• Of the capacity that asked to connect to US grids from 2000 to 2019, only 13% was operating by the end of 2024 and 77% had been withdrawn [11]. The median project built from 2018 to 2024 took more than 4 years from connection request to operation [11].

PJM adjusted its forecasts for Dominion's Virginia zone and many other zones to account for data center load growth [12]. Dominion chief financial officer Steven Ridge said on the July 31, 2026 earnings call that nine of the Dominion Zone's top 10 all-time peak days occurred in the first seven months of 2026 [23].

New power plants spend years in connection studies [11], while EPRI notes that data centers can be developed and connected in a few years [4].

US Data Centers Compared With Global Demand

We compared the US data center electricity totals with the global figures that the IEA publishes, because the US share shows how much of the worldwide rise in data center power use lands on US utilities.

The IEA finds that data centers account for nearly half of US electricity demand growth through 2030 [6]. Worldwide, the IEA expects data centers to account for around one-tenth of electricity demand growth through 2030 [6].

Global data center electricity use, 2024 to 2035

Source: IEA, Energy and AI (April 2025) [6]; IEA, Key Questions on Energy and AI (2026) [7].

Key Findings

• The IEA's 45% US share of 415 TWh implies about 187 TWh of US data center use in 2024, and its 2030 projection of 950 TWh implies average global growth of about 14% a year from the 485 TWh in 2025 [6][7].
• The IEA projects global data center electricity use of about 1,200 TWh in 2035 in its Base Case, within a range of 700 to 1,700 TWh across cases [6].

What These Numbers Mean for AI Data Centers

Utilities and state regulators now base decisions on the power a single AI server draws. EPRI's state estimates point to developers giving priority to power access and land, with new capacity planned in states such as Indiana, Louisiana, and Mississippi that have very little operating capacity today [4].

To judge when a planned AI campus can run, compare its announced power need with utility peak demand forecasts, grid operator and utility load figures, and interconnection queue times.

← All research · Power, cooling & facility ops