How much electricity does an AI data center use? The biggest match hundreds of thousands of homes

Five-second answer

Claim we investigated

The biggest AI data centers use as much electricity as hundreds of thousands of homes.

True. The five largest operating AI sites each match roughly 170,000 to 770,000 U.S. homes' yearly electricity. Satellite-and-permit estimates of site power from Epoch AI, company disclosures from Meta, the EIA household benchmark and a Congressional Research Service summary all converge on hundreds of thousands of homes per site.

Quote or cite this fact-check

FactHub rates the claim “The biggest AI data centers use as much electricity as hundreds of thousands of homes” True (high confidence): The five largest operating AI sites each match roughly 170,000 to 770,000 U.S. homes' yearly electricity. Reviewed October 26, 2026. https://facthub.org/n/article/largest-ai-data-centers-electricity/1

One-minute summary

What the evidence establishes

Established. Epoch AI estimates the five largest operating AI data centers had 421 to 946 MW of IT power in mid-2026, led by xAI's Colossus 2.
Established. At full IT load, each would use as much electricity in a year as about 342,000 to 768,000 average U.S. homes.
Established. Even at half load, each still matches about 171,000 to 384,000 homes, and cooling adds more on top.
Established. Meta describes its Prometheus site as 1 GW and plans to scale Hyperion to 5 GW, about 812,000 and 4 million homes at full use.
Not shown. The exact electricity each site uses is not published; utilisation and capacity are estimates.
Context. Most data centers are far smaller; a typical large one matches about 100,000 homes.

Important context: The figures compare a data center's yearly electricity with the electricity an average U.S. home buys in a year (10,791 kWh in 2022). Data centers are rated by power capacity in megawatts, and how much of that capacity a site uses on average is not public, so this check gives a range from half to full use of the IT equipment and leaves out cooling, which adds more.

The evidence

The conclusion rests on the combined record. Expand each item to see its role and limitations.

Independence note: This check uses 5 methods across 7 institutions and 2 countries or international groupings.

Evidence map
Method	Institution / country	Year(s)	What it shows	Key limitation	Primary links
Facility-level estimates of IT power capacity from high-resolution satellite imagery, permits and company disclosures.	Epoch AI United States	2025, 2026	Epoch AI's Frontier Data Centers data, updated through August 2026, lists the largest operating AI data centers by IT power: Colossus 2 (946 MW), Anthropic-Amazon New Carlisle (910 MW), Microsoft Fairwater Atlanta (636 MW), Meta Prometheus (562 MW) and OpenAI Stargate Abilene (421 MW). FactHub's dataset records these values with their estimate dates.	These are estimates of power capacity, not metered use, built from imagery, permits and disclosures that Epoch AI says carry uncertainty in both power and dates. IT power excludes cooling, so whole-site draw is higher. Epoch AI is a single research group, though it is not a party to any data center project.	Epoch AI Epoch AI
Arithmetic on official statistics: IT power x 8,760 hours, divided by EIA's average annual electricity per U.S. residential customer (10,791 kWh).	U.S. Energy Information Administration, FactHub United States	2022, 2026	One megawatt running for a year is 8,760 MWh. Dividing each site's annual energy by 10,791 kWh gives about 342,000 to 768,000 homes at full IT load and 171,000 to 384,000 at half load. The calculation and its inputs are in the FactHub dataset 'Largest operating AI data centers'.	The benchmark is electricity sold per residential customer in 2022, which understates use in homes with rooftop solar. Site utilisation is not public, so the check brackets it between half and full IT load rather than measuring it.	U.S. Energy Information Administration FactHub
Company statements reported by TechCrunch, July 2025.	Meta, TechCrunch United States	2025	Meta's chief executive said the company would bring a 1 GW cluster called Prometheus online in 2026, and a spokesperson said Hyperion would have 2 GW by 2030 and scale to 5 GW. At full use, 1 GW is 8.76 TWh a year, about 812,000 homes.	These are company statements of planned capacity reported by a news outlet, not measurements, and the company benefits from presenting its build-out as large. Epoch AI's independent estimate for Prometheus in August 2026 (562 MW) was below the 1 GW target.	TechCrunch
Expert summary for Congress of IEA estimates.	Congressional Research Service, International Energy Agency United States, France	2025	A June 2025 CRS report says large data centers typically match about 100,000 households, the largest under construction about 3 million and the largest announced about 5 million, consistent with the range found here.	The Congressional Research Service summarizes International Energy Agency figures from a podcast rather than measuring sites itself, so it is not independent of the IEA. It uses energy rather than electricity and gives round numbers.	Congressional Research Service
Manufacturer specification for an NVIDIA GB200 NVL72 rack.	Supermicro United States	2026	Supermicro rates its GB200 NVL72 rack, with 72 GPUs, at 132 kW total power. Running all year, that is about 1.16 GWh, roughly 107 average U.S. homes, so a site with several thousand such racks reaches hundreds of megawatts.	This is a manufacturer's rated maximum for one product, not a measurement of any site, and racks do not always draw their maximum.	Supermicro

1. Independent satellite and permit estimates put the largest sites at 421 to 946 MW Facility-level estimates of IT power capacity from high-resolution satellite imagery, permits and company disclosures.

Epoch AI's Frontier Data Centers data, updated through August 2026, lists the largest operating AI data centers by IT power: Colossus 2 (946 MW), Anthropic-Amazon New Carlisle (910 MW), Microsoft Fairwater Atlanta (636 MW), Meta Prometheus (562 MW) and OpenAI Stargate Abilene (421 MW). FactHub's dataset records these values with their estimate dates.

Limit: These are estimates of power capacity, not metered use, built from imagery, permits and disclosures that Epoch AI says carry uncertainty in both power and dates. IT power excludes cooling, so whole-site draw is higher. Epoch AI is a single research group, though it is not a party to any data center project.

Verify: The record for power capacity in a single data center has doubled every 10 months AI data centers (Frontier Data Centers Hub)

Epoch AI estimates the largest AI data center by IT power, xAI's Colossus 2, at about 950 MW. Epoch AI
Epoch AI's capacity figures are independent estimates built from satellite imagery and permits. Epoch AI
IT power excludes cooling and other overhead, so whole-site use is higher than the IT figure. Epoch AI
Epoch AI says its power and date estimates are uncertain. Epoch AI · Qualifies

2. Converted with the federal household benchmark, each site matches hundreds of thousands of homes Arithmetic on official statistics: IT power x 8,760 hours, divided by EIA's average annual electricity per U.S. residential customer (10,791 kWh).

One megawatt running for a year is 8,760 MWh. Dividing each site's annual energy by 10,791 kWh gives about 342,000 to 768,000 homes at full IT load and 171,000 to 384,000 at half load. The calculation and its inputs are in the FactHub dataset 'Largest operating AI data centers'.

Limit: The benchmark is electricity sold per residential customer in 2022, which understates use in homes with rooftop solar. Site utilisation is not public, so the check brackets it between half and full IT load rather than measuring it.

Verify: How much electricity does an American home use? Largest operating AI data centers: household-equivalent electricity

An average U.S. residential customer bought 10,791 kWh of electricity in 2022. U.S. Energy Information Administration
Purchases understate total use for homes with their own solar panels. U.S. Energy Information Administration · Qualifies
At full IT load the five largest sites each match about 342,000 to 768,000 homes; at half load, about 171,000 to 384,000. FactHub

Each of the five largest operating AI data centers uses as much electricity as at least 170,000 U.S. homes even at half load, and up to about 768,000 at full load. Cooling is excluded, so true totals are higher. Data: Epoch AI capacity estimates (mid-2026), EIA household benchmark. View source chart

3. Meta says its Prometheus site is 1 GW and Hyperion will reach 5 GW Company statements reported by TechCrunch, July 2025.

Meta's chief executive said the company would bring a 1 GW cluster called Prometheus online in 2026, and a spokesperson said Hyperion would have 2 GW by 2030 and scale to 5 GW. At full use, 1 GW is 8.76 TWh a year, about 812,000 homes.

Limit: These are company statements of planned capacity reported by a news outlet, not measurements, and the company benefits from presenting its build-out as large. Epoch AI's independent estimate for Prometheus in August 2026 (562 MW) was below the 1 GW target.

Verify: Mark Zuckerberg says Meta is building a 5GW AI data center

Meta planned to bring a 1 GW cluster, Prometheus, online in 2026. TechCrunch
Meta's Hyperion campus is planned at 2 GW by 2030, scaling to 5 GW. TechCrunch

4. Congressional researchers describe large data centers in household terms Expert summary for Congress of IEA estimates.

A June 2025 CRS report says large data centers typically match about 100,000 households, the largest under construction about 3 million and the largest announced about 5 million, consistent with the range found here.

Limit: The Congressional Research Service summarizes International Energy Agency figures from a podcast rather than measuring sites itself, so it is not independent of the IEA. It uses energy rather than electricity and gives round numbers.

Verify: Energy Tax Benefits for Data Centers: In Brief (R48583)

A typical large data center uses about as much energy as 100,000 households. Congressional Research Service
The largest data center under construction is projected to match about 3 million households. Congressional Research Service

5. One current AI server rack is rated at about 132 kW Manufacturer specification for an NVIDIA GB200 NVL72 rack.

Supermicro rates its GB200 NVL72 rack, with 72 GPUs, at 132 kW total power. Running all year, that is about 1.16 GWh, roughly 107 average U.S. homes, so a site with several thousand such racks reaches hundreds of megawatts.

Limit: This is a manufacturer's rated maximum for one product, not a measurement of any site, and racks do not always draw their maximum.

Verify: SuperServer SRS-GB200-NVL72 rack specifications

One GB200 NVL72 rack is rated at 132 kW total power. Supermicro

Common questions

“Data centers don't run at full power all the time, so the comparison overstates their use.”

Short answer: Still hundreds of thousands

True that sites keep spare capacity: Lawrence Berkeley National Laboratory estimates U.S. data centers use about half their grid connection on average. But this check already shows the half-load case, and each of the five largest sites still matches about 171,000 to 384,000 homes. Cooling also adds to the IT load, which these figures leave out.

LBNL estimates U.S. data centers use about half of their grid connection on average. GreenCalculus · Qualifies
IT power excludes cooling and other overhead, so whole-site use is higher than the IT figure. Epoch AI

“The gigawatt figures are plans, not what is actually running.”

Short answer: Partly

Company targets such as Meta's 1 GW and 5 GW are plans. That is why this check rests on Epoch AI's estimates of capacity already built, which reached 946 MW at Colossus 2 by June 2026. The claim holds on built capacity alone.

Epoch AI estimates the largest AI data center by IT power, xAI's Colossus 2, at about 950 MW. Epoch AI
Meta's Hyperion campus is planned at 2 GW by 2030, scaling to 5 GW. TechCrunch

Conclusion

The claim holds under every reasonable assumption we tested. Epoch AI, an independent research group that estimates data center power from satellite imagery and permits, lists five operating AI campuses with 421 to 946 MW of IT power as of mid-2026. Running a year at full IT load, those sites would use 3.7 to 8.3 TWh, the electricity bought by about 342,000 to 768,000 average U.S. homes. Even if their chips averaged only half of capacity, each would still match about 171,000 to 384,000 homes. Both figures leave out cooling and power conversion, so the true totals are likely higher. The independent lines agree. Meta has said its Prometheus cluster is a 1 GW site and that its Hyperion campus will reach 5 GW; a gigawatt running all year is 8.76 TWh, about 812,000 homes. The Congressional Research Service, summarizing the International Energy Agency, describes a typical large data center as matching about 100,000 households and the largest under construction as matching about 3 million. Rack-level specifications point the same way: one current AI rack is rated at about 132 kW, roughly a hundred homes' worth, and these campuses hold thousands of them. What remains uncertain is precision, not direction. Capacity figures are estimates, the sites' actual utilisation is not published, and the household comparison uses electricity purchases, which understate use in homes with rooftop solar. None of these changes the order of magnitude. The claim describes the biggest AI sites; most data centers are far smaller, and national totals are a separate question: U.S. data centers as a whole could use 9.5% to 15.3% of the country's electricity by 2030, according to Lawrence Berkeley National Laboratory.

How we checked

Defined the claim as: each of the largest AI data centers uses, in a year, as much electricity as hundreds of thousands (at least 100,000) of average U.S. homes.
Took facility capacity from an independent tracker (Epoch AI) rather than from company announcements, and checked it against Meta's own statements.
Converted capacity to annual energy and household equivalents with EIA's 2022 figure, at full and at half IT load, and published the calculation as a FactHub dataset.
Cross-checked the order of magnitude against a Congressional Research Service summary and a manufacturer's rack specification.
Recorded the limits of each source: estimated capacity, unknown utilisation, excluded cooling and the household benchmark's treatment of rooftop solar.

Editorial note: Scope: this check covers the largest individual AI data centers operating in 2026, not data centers in general or AI's share of national electricity. Household equivalents are FactHub calculations from published capacity estimates; they are not metered consumption. No external specialist review has been recorded.

Expert-review status

No external specialist review has been recorded.

Funding and conflicts

Funding: No external funding for this piece.

Ownership: Published by the FactHub Research Team, an editorial group within FactHub.

Conflicts: No known conflicts of interest with the companies or researchers named.

Sources

Sources are labeled by kind, with what this fact check takes from each, so readers can check the evidence independently.

Academic Epoch AI. The record for power capacity in a single data center has doubled every 10 months
Data insight by Axel Campos and Ben Cottier (September 4, 2026) with per-facility IT power estimates from Epoch AI's Frontier Data Centers data.

Reliability: Epoch AI is an independent nonprofit research group that estimates data center power from satellite imagery, permits and disclosures. It does not build or operate data centers.
- Epoch AI estimates the largest AI data center by IT power, xAI's Colossus 2, at about 950 MW.
  
  What the source says
  
  The current record-holder is xAI’s Colossus 2, at an estimated 950 MW.
  Opening paragraph
- IT power excludes cooling and other overhead, so whole-site use is higher than the IT figure.
  
  What the source says
  
  IT power is the power capacity available to chips, servers, and other IT equipment, excluding cooling and other overhead.
  Learn more about this graph
- Qualifies: Epoch AI says its power and date estimates are uncertain.
  
  What the source says
  
  because of the small sample size and the uncertainty in our power and date estimates, it’s unclear whether the power trend will actually slow down.
  Paragraph below the chart
Academic Epoch AI. AI data centers (Frontier Data Centers Hub)
Epoch AI's open database of large AI data centers, updated October 26, 2026.

Reliability: Same publisher as the data insight; this page documents how the estimates are made and offers the data under CC BY.
- Epoch AI's capacity figures are independent estimates built from satellite imagery and permits.
  
  What the source says
  
  Independent estimates of power, AI compute, and cost built from high-resolution satellite imagery, permits, and our understanding of how data centers are constructed.
  Page introduction
Government U.S. Energy Information Administration. How much electricity does an American home use?
EIA frequently asked question giving average annual electricity sales per U.S. residential customer.

Reliability: EIA is the U.S. Department of Energy's statistical agency; the figure comes from utility sales reports.
- An average U.S. residential customer bought 10,791 kWh of electricity in 2022.
  
  What the source says
  
  In 2022, the average annual amount of electricity sold to (purchased by) a U.S. residential electric-utility customer was 10,791 kilowatthours (kWh), an average of about 899 kWh per month.
  First paragraph
- Qualifies: Purchases understate total use for homes with their own solar panels.
  
  What the source says
  
  Electricity purchases may not represent the total electricity consumption for some residential electricity customers.
  Second paragraph
News report TechCrunch. Mark Zuckerberg says Meta is building a 5GW AI data center
News report by Maxwell Zeff (July 14, 2025) on Meta's statements about its Prometheus and Hyperion data centers.

Reliability: Technology news outlet reporting a Threads post by Meta's chief executive and an emailed statement from a Meta spokesperson. Meta, the subject, has an interest in presenting its build-out as large.
- Meta planned to bring a 1 GW cluster, Prometheus, online in 2026.
  
  What the source says
  
  Zuckerberg also noted that Meta plans to bring a 1 GW super cluster, called Prometheus, online in 2026, making it one of the first tech companies to control an AI data center of this size.
  Fourth paragraph · Mark Zuckerberg, Meta chief executive, as reported
- Meta's Hyperion campus is planned at 2 GW by 2030, scaling to 5 GW.
  
  What the source says
  
  Gabriel says Meta plans to bring two gigawatts of data center capacity online by 2030 with Hyperion, but that it would scale to five gigawatts in several years.
  Third paragraph · Ashley Gabriel, Meta spokesperson
Government Congressional Research Service. Energy Tax Benefits for Data Centers: In Brief (R48583) Identifier: R48583
CRS report by Nicholas E. Buffie (June 25, 2025) for Members of Congress.

Reliability: Nonpartisan research arm of Congress. The household comparisons are attributed to an IEA podcast, so they are not independent of the IEA.
- A typical large data center uses about as much energy as 100,000 households.
  
  What the source says
  
  Large data centers typically have the same energy usage as roughly 100,000 households.
  p. 1, Introduction
- The largest data center under construction is projected to match about 3 million households.
  
  What the source says
  
  The largest data center currently under construction is projected to use as much energy as 3 million households
  Summary page
Industry Supermicro. SuperServer SRS-GB200-NVL72 rack specifications
Manufacturer product page for a liquid-cooled rack with 72 NVIDIA B200 GPUs.

Reliability: Hardware vendor's published specification; it states rated power, not measured draw in service.
- One GB200 NVL72 rack is rated at 132 kW total power.
  
  What the source says
  
  8x 1U 33kW (6x 5.5kW PSUs), total power 132kW
  Rack Configuration, Power
Other GreenCalculus. Data Center Energy Consumption: How Much Electricity Data Centers Use
Reference page (October 2026) summarizing LBNL, IEA and operator data on data center electricity.

Reliability: Commercial carbon-data publisher that summarizes Lawrence Berkeley National Laboratory's 2025 Update; used here only for LBNL's utilisation estimate, which we could not read in the full report.
- Qualifies: LBNL estimates U.S. data centers use about half of their grid connection on average.
  
  What the source says
  
  LBNL estimates that US data centers use about 50% of their interconnection capacity on average, because they keep spare capacity for redundancy and maintenance.
  How much power does a data center use?
Government Lawrence Berkeley National Laboratory. United States Data Center Energy Usage Report: 2025 Update DOI: 10.71468/P1RP4F
Report by Sarah Smith, Alexander Hubbard, Alex Newkirk and colleagues (June 2026), the standard U.S. estimate of data center electricity.

Reliability: U.S. national laboratory report requested by Congress, built from equipment shipment data and cooling simulations.
- Data centers could use 11.8% of U.S. electricity by 2030, within a 9.5% to 15.3% range.
  
  What the source says
  
  estimates that data centers could account for 11.8% of total U.S. electricity by 2030. The estimate also includes a range of scenarios that indicate the energy use could be between 9.5 and 15.3% of total U.S. electricity use by 2030.
  Abstract
- LBNL's reference case is 649 TWh of U.S. data center electricity in 2030.
  
  What the source says
  
  The Reference Case estimate for electricity use (649 TWh in 2030) is calculated based on the current understanding of expected shipments and equipment design across the data center industry.
  Abstract
FactHub record FactHub. Largest operating AI data centers: household-equivalent electricity
FactHub saved analysis converting Epoch AI capacity estimates to household equivalents with the EIA benchmark.
- At full IT load the five largest sites each match about 342,000 to 768,000 homes; at half load, about 171,000 to 384,000.
FactHub dataset Largest operating AI data centers: power capacity and household-equivalent electricity, mid-2026
Tables: largest-ai-dc

Saved analysis 'household-equivalents' and table 'largest-ai-dc'
FactHub chart largest-ai-data-centers-homes

Update history

October 26, 2026 First version.

For search, research, and AI systems

Machine-readable fact-check

This page includes standardized ClaimReview metadata describing the exact claim, FactHub's finding, publisher, and review dates. You can inspect the record below and confirm that it matches the article you just read.

View ClaimReview JSON

{"@context":"https://schema.org","@type":"ClaimReview","url":"https://facthub.org/n/article/largest-ai-data-centers-electricity/1","datePublished":"2026-10-26","dateModified":"2026-10-26","author":{"@type":"Organization","name":"FactHub","url":"https://facthub.org/"},"claimReviewed":"The biggest AI data centers use as much electricity as hundreds of thousands of homes.","reviewRating":{"@type":"Rating","alternateName":"True"},"image":"https://facthub.org/node-assets-v2/article/largest-ai-data-centers-electricity/1/main-illustration/social-1200x630.jpg"}

This metadata helps machines discover the fact-check. It does not prove the conclusion or guarantee special treatment by a search engine.

Challenge this fact-check

Every source above is linked so you can check it yourself. Identify the specific conclusion you dispute and provide verifiable evidence that could change it. Credible challenges deserve review and a visible response.

Submit evidence Send general feedback instead

Technical details: node handle article:largest-ai-data-centers-electricity