GIP Index / GIP Explainer

Built, planned, or used? Sorting out AI's power numbers

The big data center number in the headlines is usually a nameplate or a promise. The electricity actually used is about half as large, and that's the number the GIP counts.

When you see “61 gigawatts” next to the words “data centers,” assume it’s a ceiling or a wish list. It is almost never the electricity those buildings burned.

That matters, because the gap between those numbers is where most of the confusion about AI and power lives. Capacity tells you how big the plumbing is. Consumption tells you how much water went through it.

I’ve seen roughly 61 GW used for two different things this year. One is how much data center capacity the United States already has. The other is how many gigawatts AI companies say they want to switch on.

Both are real numbers. Neither is a measure of work done.

So let’s pull them apart into three numbers: what’s installed today, what’s in the pipeline, and what was actually consumed.

Number one: installed capacity

Start with the counts of what already exists. Three credible sources land in the same neighborhood, which is part of why “61 GW” sticks.

The International Energy Agency puts total installed US data center capacity at 52 GW in 2025, up from 35 GW in 2023. Its annex also gives the IT load inside that, the power the servers themselves can draw: 39 GW. The rest is cooling and other overhead.

451 Research, part of S&P Global, measures something slightly different: utility power supplied to hyperscale, leased and crypto-mining data centers. Its September 2025 forecast said that would reach 61.8 GW by the end of 2025, up 22%.

In May 2026 it revised the outcome up to about 64.4 GW, up 25%. That series includes crypto mining, which the IEA leaves out, and excludes enterprise-owned data centers outside the big tech firms.

The tracker DC Hub shows “operating now” at 61 GW as of Nov 1, 2026. That figure is the sum of megawatts across facility records, and records without a megawatt figure are excluded. It’s a database total, not a meter reading.

So the installed number is somewhere between about 52 and 64 GW depending on what you count. That’s a fair answer to “how big is America’s data center fleet?” It’s the wrong answer to “how much power does AI use?”

Number two: the pipeline

The second 61 GW is a promise. Here the numbers get much bigger and much softer.

The US AI Data Center Atlas, which tracks campuses of 100 MW or more by planned IT capacity, counts 29 GW at campuses that are partly or fully in service. Behind that sits 192 GW that isn’t. Of that 192 GW, 79 GW is at campuses where building has started, 57 GW is permitted and 55 GW is at the announced or land-purchase stage.

The Atlas also asks whether a power supply is lined up. Of the 192 GW not yet in service, 84 GW has firm power on record. Another 61 GW has nothing more than a named utility.

DC Hub’s pipeline is smaller, at 63 GW across 201 projects in construction, permitting or planning. Different tracker, different rules, same lesson: the pipeline is at least as large as everything already built.

Company announcements add to the noise. OpenAI said in April it had committed to 10 GW of US AI infrastructure by 2029 and had “already surpassed that milestone.”

DatacenterDynamics noted the 10 GW is likely contracted rather than live, and that OpenAI ended 2025 with 1.9 GW of compute capacity, according to its CFO. Secured and running are not the same verb.

Grid queues are bigger still. The IEA reports that ERCOT, the Texas grid operator, has received 160 GW of data center interconnection requests, “twice the existing global capacity of hyperscale and colocation data centres.” The agency adds that “not all interconnection requests result in approvals” and that some are speculative.

There’s a clean natural experiment on that point. In Ohio, after regulators required large data centers to pay for at least 85% of the energy they sign up for whether they use it or not, AEP Ohio’s data center requests fell from more than 30 GW to 13 GW. More than half the queue disappeared once a request carried a bill.

None of this means the build-out is fake. The IEA’s base case has US installed capacity roughly doubling to 100 GW by 2030, and 451 Research now projects about 183.2 GW of US grid demand by 2030, up from 117.1 GW in its June 2025 forecast.

Something big is happening. But a pipeline gigawatt is an option on the future, and options expire.

Number three: electricity actually consumed

Now the number that does the work. The IEA estimates US data centers consumed 224 terawatt-hours in 2025.

Divide by the 8,760 hours in a year and you get an average draw of about 25.6 GW. That’s under half the 52 GW the IEA counts as installed, which matches its own estimate of a 50% average US capacity factor. Data centers are built for peak demand plus a safety margin, and the IEA notes they “often oversize their grid connections initially.”

The US government’s own model gives a similar picture. Lawrence Berkeley National Laboratory’s 2024 report estimated 176 TWh in 2023, 4.4% of US electricity, excluding crypto. For 2028 it projected 325 to 580 TWh, and it did the conversion out loud: “Assuming an average capacity utilization rate of 50%,” that range means 74 to 132 GW of total power demand.

LBNL also explains why it builds bottom-up from shipped servers. Doing so “avoids overestimation that can be caused by tracking data center load for projects that have not yet selected a power provider.” That is the pipeline problem, stated by the people who model it.

Why the GIP counts consumption

The Gross Intelligence Product is built from three pieces: electricity used by AI, computation per unit of electricity, and capability per unit of computation. The first piece has to be electricity used, not capacity built. The methodology page spells out the formula.

The reason is simple. An idle rack produces no answers.

A campus with signed land and a named utility produces no answers either. If we used capacity, the index would rise every time a press release landed, and fall when a project was cancelled, with no change in the amount of thinking done.

Consumption has its own flaws. It counts the cooling, the lighting and every inefficient chip along with the good ones. But it’s a physical quantity someone measured or modeled from shipments, and it moves when work moves.

AI-focused facilities are not AI workloads

One more definition trips people up. The IEA reports that “AI-focused” data centers worldwide used about 155 TWh in 2025, which Our World in Data puts at 0.49% of global electricity. That category grew 50% in the year.

The IEA defines AI-focused data centers as facilities “with large numbers of servers equipped with specialised accelerators.” That’s a hardware label on a building. By construction it counts everything those buildings do, including the cooling and any non-AI jobs, and it misses AI work that runs on ordinary servers elsewhere.

So 155 TWh is the best available proxy for AI’s electricity, not a meter on AI itself. Spread over the year it works out to about 17.7 GW of average draw worldwide.

Compare that with Epoch AI’s estimate of global AI data center power capacity: about 30 GW by late 2025. Epoch builds its number from chip power ratings multiplied by 2.5 for overhead.

The methods differ, so treat this as rough, but 17.7 against 31 GW implies the AI fleet ran at a little over half its rated power on average. Same story as the whole fleet.

How the US compares

The obvious next question is how America stacks up against everyone else. The honest answer is that only one source compares the major regions on one basis, and that’s the IEA’s annex. It reports installed capacity and consumption for the same regions, defined the same way, for the same year.

On that basis, the US had 52 GW installed in 2025 and used 224 TWh. China had 28 GW and 117 TWh. Europe, which in the IEA tables is wider than the EU, had 17 GW and 72 TWh.

The whole Asia Pacific region, China included, had 41 GW and 173 TWh. The Middle East had 0.4 GW and 1.7 TWh.

The US is about half of the world’s 114 GW of installed capacity, which the IEA says outright. Worldwide consumption was 485 TWh, about 55 GW of average draw.

The ratio of draw to capacity is similar everywhere, which is reassuring. The IEA’s capacity factors for 2025 are 50% for the US, 48% for Europe and China, and 45% for the Middle East.

The overhead differs more. The IEA’s power usage effectiveness, total power divided by IT power, is 1.30 in the US, 1.42 in Europe, 1.47 in China and 1.88 in the Middle East, where cooling in the heat costs more.

National meters tell a slightly different story

Some governments publish metered or official consumption figures. These are better data for their own countries, but they don’t share a definition, so I’ve kept them out of the chart.

China’s National Energy Administration reported 170 TWh for 2025, well above the IEA’s 117 TWh, and expects computing electricity to reach 800 TWh, about 6% of national use, by 2030. The gap almost certainly comes from what counts as a “computing center,” which is exactly why I don’t put the two numbers in the same bar.

The small countries show the concentration problem. Ireland’s data centers used 7,663 GWh in 2025, 23% of metered electricity, up from 5% in 2015, per the CSO.

In Britain, DESNZ found data centers took 2% of grid use nationally but 65% of Slough’s. The Netherlands was at 4.6%, up 37% from 2021.

Pipelines abroad

Announced pipelines outside the US are real but harder to pin down. The IEA puts the Middle East’s total project pipeline at almost 10 GW, against 0.4 GW installed, and says, citing Epoch AI, that the region’s gigawatt-scale projects alone are an order of magnitude larger than everything already built there. For the EU, it estimates the project pipeline at 2.3 times current installed capacity, alongside a European Commission goal of tripling capacity in five to seven years.

The most concrete Gulf project I could verify is Stargate UAE, a 1 GW cluster in Abu Dhabi with 200 MW expected to go live in 2026. That’s a company announcement, not a meter.

I’ve left Japan, India, Germany and Saudi Arabia out of the chart and tables. I couldn’t find capacity or consumption figures for them on the same basis as the rest, and I’d rather leave a gap than fill it with a guess. The IEA does note pipeline growth in India after new tax incentives, without giving a number.

Who pays for the promise

Here’s where the distinction stops being academic. Grid operators plan and buy capacity for forecast demand, and forecasts are built partly from the pipeline.

In the PJM market, which covers much of the mid-Atlantic and Midwest, the independent market monitor calculated that existing and forecast data center load raised capacity market revenues by about $29.4 billion across four auctions covering 2025/26 to 2028/29. Its verdict on the inputs: “Forecast data center load growth has been the primary cause to date and the accuracy of those forecasts is highly questionable.”

Those costs, the monitor says, are “significant and irreversible” and will be paid through May 31, 2029. The IEA adds that PJM’s load forecasts “may include duplicative or speculative requests.”

Household bills are rising, though not only because of data centers. The EIA expects the average US residential price to reach 18.2 cents per kilowatt-hour in 2026, up from 17.3 cents in 2025. It projects PJM wholesale prices up 41% in 2026, the most of any region, but attributes wholesale moves mainly to weather and natural gas prices.

The IEA’s own conclusion is that “with the right conditions, data centres do not necessarily raise electricity prices.” I agree with the conditional. The condition is that the people asking for the gigawatts pay for them, as Ohio now requires.

What to do with a gigawatt headline

My read: the confusion is mostly a units problem, and treating it as a scandal misses the fix. Every number above is defensible on its own terms. The trouble starts when a pipeline figure gets quoted as if it were a meter reading, and nobody asks which one it is.

So when you see a gigawatt number, ask three things. Is it installed, planned or consumed? Does it include crypto and cooling?

And is there a TWh figure nearby? If the answer to the last question is no, divide by two before you draw any conclusions about AI, and go find the consumption number.

For regulators and utilities, the Ohio result is the most useful fact in this piece. Make queue positions cost something. Speculative gigawatts vanish when they carry a bill, and the people who would have paid for them are mostly households.

Sources

1. “Key Questions on Energy and AI: Executive summary,” International Energy Agency, Apr 2026. https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary
2. “Key Questions on Energy and AI” (full report and Annex A tables, CC BY-NC-SA 4.0), International Energy Agency, Apr 2026. https://www.iea.org/files/3179f7f8-01f6-4dd6-bffa-c9f7b73f1dc9/KeyQuestionsonEnergyandAI.pdf
3. “Data center grid-power demand to rise 22% in 2025, nearly triple by 2030,” S&P Global, Oct 14, 2025. https://www.spglobal.com/energy/en/news-research/latest-news/electric-power/101425-data-center-grid-power-demand-to-rise-22-in-2025-nearly-triple-by-2030
4. “US data center grid-power demand to climb 22% in 2025,” RCRTech, Oct 2025. https://rcrtech.com/ai-infrastructure-news/us-data-center/
5. “Surging US data center power demand tests sustainability targets,” S&P Global Market Intelligence, May 6, 2026. https://www.spglobal.com/market-intelligence/en/news-insights/articles/2026/5/surging-us-data-center-power-demand-tests-sustainability-targets-101100707
6. “Data Center Power Index: totals,” DC Hub, accessed Nov 1, 2026. https://dchub.cloud/dcpi/totals
7. “Power,” US AI Data Center Atlas, accessed Nov 1, 2026. https://everydatacenter.com/power.html
8. “Building the compute infrastructure for the Intelligence Age,” OpenAI, Apr 29, 2026. https://openai.com/index/building-the-compute-infrastructure-for-the-intelligence-age/
9. “OpenAI claims to have secured 10GW of AI infrastructure capacity ahead of 2029 target,” DatacenterDynamics, May 1, 2026. https://www.datacenterdynamics.com/en/news/openai-claims-to-have-secured-10gw-of-ai-infrastructure-capacity-ahead-of-2029-target/
10. “Introducing Stargate UAE,” OpenAI, May 22, 2025. https://openai.com/index/introducing-stargate-uae/
11. “Global AI power capacity,” Epoch AI, Jan 16, 2026. https://epoch.ai/data-insights/ai-datacenter-power
12. “How much energy do data centers and artificial intelligence use?” Our World in Data, 2026. https://ourworldindata.org/how-much-energy-do-data-centers-and-artificial-intelligence-use
13. “2024 United States Data Center Energy Usage Report,” Lawrence Berkeley National Laboratory, Dec 2024. https://eta-publications.lbl.gov/sites/default/files/2024-12/lbnl-2024-united-states-data-center-energy-usage-report_1.pdf
14. “Data Centres Metered Electricity Consumption 2025,” Central Statistics Office (Ireland), Jul 7, 2026. https://www.cso.ie/en/releasesandpublications/ep/p-dcmec/datacentresmeteredelectricityconsumption2025/keyfindings/
15. “Datacenters verbruiken 4,6 procent van de elektriciteit,” Statistics Netherlands (CBS), Dec 15, 2025. https://www.cbs.nl/nl-nl/nieuws/2025/51/datacenters-verbruiken-4-6-procent-van-de-elektriciteit
16. “Data centre electricity consumption in Great Britain, 2020 to 2024,” Department for Energy Security and Net Zero, Jun 30, 2026. https://assets.publishing.service.gov.uk/media/6a3ea4334c7605ab56723bfd/Data_centre_electricity_consumption_in_Great_Britain__2020_to_2024.pdf
17. “去年全国算力中心总用电量1700亿千瓦时,” National Energy Administration (China), May 29, 2026. https://www.nea.gov.cn/20260529/18c09743776c4dddb55288ceea38b8b1/c.html
18. “2026 Quarterly State of the Market Report for PJM: January through June,” Monitoring Analytics, Aug 13, 2026. https://www.monitoringanalytics.com/reports/PJM_State_of_the_Market/2026/2026q2-som-pjm.pdf
19. “Short-Term Energy Outlook: Electricity, coal, and renewables,” U.S. Energy Information Administration, Oct 6, 2026. https://www.eia.gov/outlooks/steo/report/elec_coal_renew.php