AI data center power consumption is measured at three levels: a server or rack in kilowatts, a campus in megawatts or gigawatts, and a country or the world in terawatt-hours a year. Microsoft says its Fairwater AI design runs at about 140 kW per rack; announced AI campuses range from tens of megawatts to OpenAI’s approximately 8 GW-IT PORTS-Pike site in Ohio; and a Lawrence Berkeley National Laboratory report for the US Department of Energy found that data centers used about 4.4% of US electricity in 2023. The figures below come from vendors’, operators’ and agencies’ own publications, as of October 2026.

On this page

Server and rack level

The smallest unit with a published power figure is the server. Nvidia’s DGX B200 user guide gives a maximum system power of 14.3 kW for one 10U system with eight B200 GPUs.

Rack-scale systems put far more GPUs in one cabinet. Nvidia’s GB200 NVL72 is a single liquid-cooled rack with 72 Blackwell GPUs and 36 Grace CPUs. Microsoft, which runs these systems in its Fairwater data centers, said in November 2025 that the Fairwater design operates at about 140 kW per rack and 1,360 kW per row. Specifications for individual accelerators are in the /ai-chips catalog.

Campus level: what the announced numbers measure

Data-center owners announce capacity in megawatts (MW) or gigawatts (GW), but not always the same quantity. Four meanings appear in announcements:

  • IT load (MW-IT, GW-IT, “critical IT capacity”). Power available to the servers, storage and network equipment. It excludes cooling and electrical losses.
  • Power or utility load. What the site draws from the grid, which covers IT load plus cooling and other building systems.
  • Compute capacity. A term some owners use without saying which of the two above it means.
  • MWe and contracted generation. The electrical output of a power plant, or the share of it a buyer has contracted under a power purchase agreement (PPA). This is supply, not data-center demand, and the power may serve operations across a grid region rather than a single campus.

One contract shows the gap between the first two. On September 30, 2026, AIB Data Centers said it had agreed to provide Nebius with “50 MW of critical IT capacity” in the southeastern United States, at a site supported by “a 15-year Electric Service Agreement for 65 MW of utility load.” The same facility is described by both numbers.

Project Announced figure What the figure measures Owner’s timeline
OpenAI PORTS-Pike, Ohio About 8 GW-IT IT load First 800 MW in 2028; buildout through 2032
Meta Richland Parish, Louisiana 5 GW “Compute capacity” Expansion announced July 13, 2026
OpenAI Project Camellia, Georgia 3.2 GW Power delivered by Georgia Power Phases between 2028 and 2032
NextEra and Brookfield, Paducah, Kentucky Over 1.2 GW Compute capacity, with up to 4.6 GW of dedicated generation Fully built in 2032; subject to definitive agreements
AIB Data Centers for Nebius 50 MW Critical IT capacity (65 MW utility load) 12-year initial term
Constellation and Amazon, Calvert Cliffs 690 MW Contracted nuclear generation (PPA) 20 years; about 190 MW of uprates online 2030–2032
Crusoe and Aalo Atomics, Idaho 10 MWe Reactor electrical output (Aalo-X) Proof of concept at Idaho National Laboratory in 2027

Company-wide totals use yet another basis. CoreWeave said on September 17, 2026 that its “total contracted power” was approximately 4.2 GW as of August 11, 2026, a figure for the company as a whole rather than for a single building. The /ai-data-centers tracker records every project with the capacity wording its owner used.

Why a gigawatt figure is not annual consumption

Capacity is a rate; consumption is energy over time. As a simple illustration, 1 GW drawn continuously for a full year would be 8.76 TWh (1 GW × 8,760 hours). Real campuses come online in phases over several years and do not run at full capacity every hour, so an announced gigawatt figure says little about a given year’s electricity use. PORTS-Pike, for example, starts with 800 MW in 2028 and reaches its full size only by 2032, according to OpenAI.

National statistics: the US DOE and LBNL report

The main official US estimate is the United States Data Center Energy Usage Report by Lawrence Berkeley National Laboratory, released by the Department of Energy on December 20, 2024. The DOE summary gives these figures:

  • US data centers used about 58 TWh in 2014 and 176 TWh in 2023.
  • That 2023 figure was about 4.4% of total US electricity consumption.
  • The report projects 325 to 580 TWh by 2028, or about 6.7% to 12% of US electricity.
  • Data-center load “has tripled over the past decade” and is “projected to double or triple by 2028.”

The report covers all data centers, not only AI facilities.

Global statistics: the IEA

The International Energy Agency’s Energy and AI report, published April 10, 2025, puts global data-center electricity consumption at around 415 TWh in 2024, about 1.5% of the world’s electricity use. The IEA projects that this will “more than double to around 945 TWh by 2030,” with electricity demand from AI-optimized data centers projected to more than quadruple over the same period.

The IEA also gives a scale comparison: “A typical AI-focused data centre consumes as much electricity as 100 000 households, but the largest ones under construction today will consume 20 times as much.” For the United States, it projects that data centers will account for almost half of the growth in US electricity demand between the report’s publication and 2030.

Reading power claims

When a company announces a gigawatt-scale AI data center, three questions help place the number: whether it is IT load, grid power or contracted generation; whether it is the first phase or the full build; and what year the owner expects it to be online. The official statistics answer a different question, total energy used across all data centers in a year, and the two kinds of figure should not be compared directly. Power deals that supply these campuses, including the Calvert Cliffs agreement covered in our news report, are counted in generation, not demand.

Sources