# What does a "gigawatt" of AI actually mean? Date: 2026-11-01 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/gigawatt-ai-explained/ **Answer:** A gigawatt is a measure of power, a billion watts, roughly what a million homes draw, and AI companies now use it as their main unit for the size of a data center. It tells you how much electricity a site can use at one moment, but not how many chips it holds, how much of it is running or whether it has been built. AI companies used to brag about how many chips they had. Now they announce gigawatts. The change makes sense: chips keep changing, but electricity is the limit everyone shares, and it is much easier to compare. The catch is that the same word can describe very different things. Meta's site in Louisiana is described as 5 gigawatts, while one careful estimate shows about 1.6 gigawatts of computing power there by 2028. Both can be right, and this article shows you why. ## Power versus energy **Power** is the rate of using electricity at a given moment, measured in watts. **Energy** is the total used over time, measured in watt-hours. A 60-watt bulb has that power whenever it is on; left on for 10 hours, it uses 600 watt-hours. | Unit | Size | Everyday reference | |---|---|---| | Kilowatt (kW) | 1,000 W | A few household appliances running together | | Megawatt (MW) | 1,000 kW | A 50 MW site uses about what 40,500 US households use in a year | | Gigawatt (GW) | 1,000 MW | Roughly 1 million homes, ignoring swings in demand | | Terawatt-hour (TWh) | Energy, not power | 1 GW running all year uses 8.76 TWh | The household figures come from the Rockefeller Institute and Epoch AI, a research group that tracks AI data centers. The last line is simple arithmetic: 1 GW × 8,760 hours in a year. That conversion is useful. The International Energy Agency (IEA) says the world's data centers used 485 TWh in 2025. Dividing by 8.76 gives an average of about 55 GW of all data centers running around the clock, not only AI ones. The IEA expects about 950 TWh in 2030, which is about 108 GW on average and roughly 3% of the world's electricity. ## Three numbers hiding in one **Planned or running.** This is the largest gap. Epoch's November 2026 ranking separates capacity already available from the capacity planned: | Site | IT power now (estimate) | Planned IT power (estimate) | |---|---|---| | Anthropic-Amazon New Carlisle, Indiana | 1,037 MW | 1,925 MW by 2028 | | Colossus 2, Memphis (SpaceX) | 946 MW | 1,531 MW by 2027 | | OpenAI Stargate Abilene, Texas | 824 MW | not listed in the top five | | Microsoft Fairwater Atlanta | 636 MW | not listed in the top five | | Microsoft Fairwater Wisconsin | not listed in the top five | 2,263 MW by 2028 | So as of this month, Epoch's estimate is that no single site has much more than 1 GW of computing power running. Claims of 5 or 10 GW describe plans. **IT or facility.** IT power counts only the computing equipment. Facility power adds cooling and other systems. Epoch's example is a Texas site reported at 1.4 GW of IT power, or about 2 GW in total. Companies do not always say which one they mean. **Site or cluster.** A company may give the eventual size of a whole property. [Meta says](https://datacenters.atmeta.com/2026/07/deepening-our-investment-in-richland-parish-louisiana/) its Richland Parish site will expand to 5 GW with an investment of more than $50 billion, and will house its Hyperion training cluster. Epoch's estimate for Hyperion is about 1.6 GW of IT power by 2028. The first figure is a long-term ambition for the property. The second is a nearer estimate of computing power. ## What a gigawatt buys Epoch estimates that Colossus 2 will reach about 1,531 MW of IT power, about 1.8 million **H100-equivalents** and about $58 billion in capital cost. An H100-equivalent is a way to count computing power as if every chip were one Nvidia H100, a widely used GPU. | Per 1 GW of IT power | Colossus 2 estimate | |---|---| | Computing power | about 1.2 million H100-equivalents | | Capital cost | about $38 billion | *Illustrative division of Epoch's figures. Newer chips give more computing per megawatt, so this ratio will keep moving.* The same gigawatt built with older chips delivers less, which is why a power figure alone cannot rank two companies' computing strength. ## The Canadian scale Comparing these sites to Canadian figures shows how large they are. [Hydro-Québec](https://news.hydroquebec.com/news/press-releases/all-quebec/hydro-quebec-proposing-regie-energie-new-rate-large-data-centres-adjustment-rate-cryptographic-use-applied-blockchains.html) says data centers in the province reach about 200 MW at peak, and could reach about 1,000 MW by 2035. So a single 1 GW American campus is about five times Québec's current data center demand, and about what Hydro-Québec projects for the whole province in nine years. Alberta's grid operator [allocated a 1,200 MW interim limit](https://www.cbc.ca/news/canada/calgary/ai-data-centre-alberta-electricity-9.6977136) to just two projects. British Columbia [capped AI data centers at 300 MW](https://www.blg.com/en/insights/2026/07/data-centre-regulation-in-british-columbia-competing-for-a-limited-supply), plus any unused conventional allotment, for its first two-year allocation period. > **Try it:** Take any data center headline with a gigawatt figure. Write down whether it is planned or running, IT or facility power, and which year it refers to. If the story does not say, treat the number as an upper limit. ## How can you tell whether a gigawatt announcement is real? Look for commitments that are hard to reverse. A signed power agreement, an approved permit, steel on the ground and a named customer each make a project more likely to happen. A press release or a planning letter alone does not. The IEA itself notes that not all projects in the pipeline will be built. Independent checks help as well. Epoch publishes satellite images with its estimates, and warns that about one time in five its IT power figure misses the real value by more than a factor of 1.4, so treat them as ranges. Plans also change after construction starts: OpenAI dropped a planned expansion in Abilene, as covered in [our status report on Stargate](/en/learn/stargate-status-who-builds/). A good habit is to read "up to" as a ceiling, and to look for the same project again in a year. ## Sources - Largest AI data centers ranked by power capacity, 2026: https://epoch.ai/graphs/largest-ai-data-centers-by-power-capacity - Deepening our investment in Richland Parish, Louisiana: https://datacenters.atmeta.com/2026/07/deepening-our-investment-in-richland-parish-louisiana/ - Key Questions on Energy and AI: Executive summary: https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary --- From the Guide to Artificial Intelligence, by Robert Scales, published by Holistic Hybrid Inc. Share and adapt it under CC BY-NC-SA 4.0 (https://creativecommons.org/licenses/by-nc-sa/4.0/): credit the guide, link to https://guidetoartificialintelligence.com/en/learn/gigawatt-ai-explained/, no commercial use without permission, and keep the same licence.