How much electricity will AI data centres need? Forecasts compared

Forecasters agree that data centres will use much more electricity by 2030, but they disagree on how much. The IEA projects about 945 terawatt hours worldwide by 2030, more than double today. Grid operators in Texas and the US mid-Atlantic report even larger requests, though many requests will never be built.

This post compares the main public forecasts side by side, explains the units in plain words, and shows why the numbers differ. Every figure is labelled with who said it and when, and the post reflects what was published as of 19 October 2026. Nothing here is investment advice, and nothing is a prediction.

Key numbers at a glance

• IEA, 10 April 2025 report: global data centre demand set to rise by more than 100 percent to about 945 TWh by 2030 (projection).
• IEA, April 2025: demand from AI-optimised data centres projected to grow more than fourfold by 2030 (projection).
• IEA, April 2025: in the United States, data centres account for almost half of electricity demand growth through 2030 (projection).
• Berkeley Lab, December 2024: US data centres used 58 TWh in 2014 and 176 TWh in 2023 (estimate).
• Berkeley Lab, December 2024: 4.4 percent of US electricity in 2023, projected 6.7 to 12 percent in 2028 (scenario range).
• Berkeley Lab, December 2024: US data centre demand projected to rise by 325 to 580 TWh by 2028, as reported by the US Department of Energy (scenario range).
• PJM, 17 December 2025: forecast peak load for the 2027/2028 year about 5,250 MW higher than the year before, with nearly 5,100 MW of that from data centres (company statement).
• PJM, 17 December 2025: capacity auction price at the cap of $333.44 per MW-day, supply short of the reliability requirement by 6,623 MW (company statement).
• PJM, 14 July 2026: 2028/2029 auction cleared at the cap of $325 per MW-day, 6,831 MW short of the reliability requirement, with forecast peak about 2,000 MW above the previous auction's (company statement).
• ERCOT, April 2026 preliminary forecast, reported by the Texas Tribune on 17 June 2026: peak demand of 367,790 MW in 2032 (preliminary projection).
• EIA, 15 October 2026: ERCOT set a record peak of 91.1 GW on 22 July 2026 and PJM a record 162.6 GW on 2 July 2026 (government data, EIA-930).
• ERCOT CEO, June 2026: over 250,000 MW of the forecast comes from large-load projects (company statement).
• EIA Short-Term Energy Outlook, 6 October 2026: US wholesale power prices forecast at $52 per MWh in 2026, 11 percent above 2025 (government forecast).
• EIA, same outlook: PJM wholesale prices expected to rise 41 percent in 2026 (government forecast).

First, the units

Electricity numbers come in two kinds, and mixing them up causes most of the confusion.

A megawatt (MW) measures power, the rate of use at a moment in time. A gigawatt (GW) is 1,000 MW. A single large AI data centre campus can ask for hundreds of megawatts. Grid planners care about the peak, the single hour of the year when demand is highest, because the system has to be able to meet it.

A terawatt hour (TWh) measures energy, the total used over time. One GW running all year uses about 8.76 TWh. So when you see a forecast in TWh, it is a yearly total, and when you see one in MW or GW, it is usually about peak or connection size.

A forecast of connection requests, such as a queue, is a third thing again. A request is a developer asking a utility or grid operator for a connection. It is not a building and not a customer.

What the global forecast says

The International Energy Agency (IEA) published its Energy and AI report on 10 April 2025. The figures here are taken from the American Public Power Association's 11 April 2025 summary of the report. It states that global electricity demand from data centres is set to more than double by 2030, to about 945 TWh.

The same summary reports that demand from AI-optimised data centres, meaning facilities built around accelerator chips, is projected to grow more than fourfold by 2030. It also says data centres are on track to account for almost half of US electricity demand growth through 2030, and for more than 20 percent of demand growth in advanced economies. By 2030 the report says the United States would use more electricity for data processing than for all its energy-intensive manufacturing combined, covering aluminium, steel, cement and chemicals.

On supply, the summary says the IEA expects a diverse mix, with renewables and natural gas set to take the lead because of cost and availability in key markets. The summary does not give a percentage split or a volume for each fuel.

A caution on the IEA number: it was published in April 2025 and is a projection built on assumptions about how fast AI is adopted and how efficient the chips and buildings become. The summary says the report emphasises the significant uncertainties that remain, including how fast AI is adopted and bottlenecks in the energy sector. Treat 945 TWh as a central scenario, not a promise.

What the US national laboratory says

The Lawrence Berkeley National Laboratory published a United States Data Center Energy Usage Report on 19 December 2024. It was requested under the Energy Act of 2020 and is an update to the laboratory's 2016 study. The US Department of Energy summarised it on 20 December 2024.

According to that summary, US data centres used 58 TWh in 2014 and 176 TWh in 2023, which was 4.4 percent of all US electricity. The report offers a range of scenarios for 2028, and the Department says data centres could use 6.7 to 12 percent of US electricity by then, an increase of 325 to 580 TWh. Load growth, the report found, tripled over the past decade.

The range matters more than any single point. The gap between 6.7 and 12 percent is not an error. It reflects different assumptions about how many AI servers get shipped and how much the rest of the economy grows, since the percentage depends on the total as well as the data centre part.

The laboratory's publication page, dated 19 December 2024 and listing Arman Shehabi and nine co-authors, says the report estimates historical consumption back to 2014 and gives a scenario range to 2028. Its abstract does not list the numbers, so the figures above are attributed to the Department of Energy's summary of the report.

What grid operators report

Grid operators run the wires and the markets, and they have to plan for peaks, so their numbers are in megawatts.

PJM Interconnection serves more than 67 million people across 13 states and the District of Columbia. On 17 December 2025 it announced the result of its auction for the 2027/2028 delivery year. A capacity auction is a market where generators promise to be available in a future year in return for a payment, so that the system has enough supply at peak. PJM said the forecast peak for that year is about 5,250 MW higher than the forecast used for the previous auction, and that nearly 5,100 MW of the increase is attributable to data centres. Its release said the auction left no doubt that data centre demand continues to far outstrip new supply.

The next auction, for the 2028/2029 delivery year, followed on 14 July 2026. PJM said it again cleared at the price cap, this time $325 per MW-day, 2.5 percent below the previous cap, and procured 138,318 MW. Committed supply fell 6,831 MW short of the reliability requirement, and PJM said the last two auctions were the first in its history in which the whole region fell short. The forecast peak for that year was about 2,000 MW higher than the one used a year earlier, and the release noted "the continued trend of the addition of large data center loads to the load forecast". PJM said it would seek approval for a special Backstop Procurement and is developing "Connect and Manage" rules that would let large loads such as data centres connect but operate flexibly when needed. The companion explainer on who pays for AI data centre power covers what these auctions mean for bills.

The Electric Reliability Council of Texas (ERCOT) manages most of the Texas grid, which is largely separate from the rest of the country. According to the Texas Tribune on 17 June 2026, ERCOT's preliminary April 2026 long-term forecast reached 367,790 MW of peak demand in 2032. The Tribune gave the record at that time as 85,508 MW, set in August 2023; the EIA's hourly data dates the earlier 85.5 GW record to 20 August 2024, so the two sources differ on the date. Either way, the record has since moved: the EIA reports a new ERCOT peak of 91.1 GW on 22 July 2026. The 2032 forecast is about four times that new record. ERCOT's chief executive told the board in June that over 250,000 MW of forecast demand comes from large-load projects.

Two warnings come with that number. First, the article calls it a preliminary forecast. Second, ERCOT told reporters that about 70 percent of large-load projects asking to connect are data centres, and that it has no exact count of requests that will never be built. Filing a request needs only an information form.

Why the forecasts differ

Four things explain most of the gaps.

What is being counted. The IEA counts global electricity used. The laboratory counts US use. PJM and ERCOT count peak megawatts in a region. A queue counts requests. None of these are the same thing, so a table of them side by side is a comparison of ideas, not of like with like.

Real demand versus speculative requests. A developer can ask several utilities for the same project, to see which can deliver fastest. Texas Senate Bill 6, as enrolled in 2025, directs the state regulator to require applicants to disclose whether they are pursuing substantially similar requests elsewhere in the state, which suggests the practice is common enough to need a rule. The Texas Tribune quotes an industry figure saying the old process was not built for hundreds of gigawatts of requests, and ERCOT is moving to study data centres in batches, starting with a first combined study called Batch Zero, approved by its board on 2 June 2026 and open to projects that already have financing and land.

Timing. A forecast for 2028 is mostly projects already under construction. A forecast for 2032 includes projects that exist only as announcements.

Efficiency. Chips and software keep changing how much energy a unit of AI work needs. Forecasts that assume steady efficiency can miss in either direction.

What it is doing to prices and records

Demand is already showing up in measured data. The US Energy Information Administration's Today in Energy brief of 15 October 2026 reports that three grid operators set hourly peak load records in summer 2026, based on its Form EIA-930 data: PJM reached 162.6 GW on 2 July, 1.3 percent above its June 2025 record; ERCOT reached 91.1 GW on 22 July, 6.5 percent above its previous record; and the Southwest Power Pool reached 58.0 GW on 27 July, 3.5 percent above its 2023 record. The EIA says most of these peaks came during regional heat waves in the hottest summer on record nationally. It does not attribute them to data centres, so the records show a tighter system, not a measured data centre share.

The EIA's Short-Term Energy Outlook, released on 6 October 2026, forecasts US wholesale power prices averaging $52 per MWh in 2026, 11 percent more than in 2025, which it attributes mainly to extreme weather such as winter storm Fern and hot summer temperatures. It expects PJM prices to rise 41 percent, while Northwest Mid-Columbia prices fall 23 percent. In the same outlook natural gas supplies 40 percent of US generation in 2026, coal 16 percent, nuclear 18 percent, wind 11 percent and solar 8 percent, with solar projected to reach 9 percent in 2027. Because the EIA attributes the 2026 price rise mainly to weather, do not read it as caused only by data centres.

What is firm and what is uncertain

Firm, in the sense of measured or officially stated:

• US data centre use rose from 58 TWh in 2014 to 176 TWh in 2023, per the national laboratory.
• PJM's own forecast peak for 2027/2028 includes nearly 5,100 MW of added data centre load.
• ERCOT's record peak is 91.1 GW, set on 22 July 2026 according to the EIA, and its large-load forecast is far beyond anything it has served.
• PJM's last two capacity auctions both cleared at the price cap and short of its reliability requirement.

Uncertain:

• The 2030 and 2032 figures. They are projections that depend on AI adoption, chip efficiency, financing and construction speed.
• How much of any queue is real. ERCOT says it has no exact count of requests that will not be built.
• The split of supply. The IEA says gas and renewables lead, but the share each will take depends on prices and policy.
• Queue totals. Press reports quote ERCOT's queue at very different sizes, and ERCOT has not published a single confirmed total in the sources cited here, so none is used.

What to watch next

• PJM's 2029/2030 base residual auction, which its July 2026 release says is scheduled for December, and its planned Backstop Procurement filing at FERC.
• Whether ERCOT's batch studies shrink the queue to a real number.
• The EIA's next Short-Term Energy Outlook, due 10 November 2026.
• Any update to the Berkeley Lab report, which covers scenarios only to 2028.
• FERC's follow-up to its June 2026 orders on large-load interconnection. Law firm summaries describe show-cause orders to the six grid operators issued on 18 June 2026; no final rule had been published as of 19 October 2026, as covered in the FERC large load explainer.

If you want the money side of the same story, the compute spending posts on this blog cover who pays for the chips: Anthropic's compute commitments and OpenAI's Stargate and 26 gigawatts. For one country's view, see the explainer on AI data centres and power in Japan.

For teams that want to know what software they depend on, wherever it runs, Safeguard offers SBOM and dependency scanning and a free scan at https://gold.safeguard.sh/scan.

Quick check

Sources

• AI to double global data centre electricity demand by 2030, summary of the IEA Energy and AI report (American Public Power Association, 11 April 2025)
• DOE releases new report evaluating increase in electricity demand from data centers (US Department of Energy, 20 December 2024)
• 2024 United States Data Center Energy Usage Report (Lawrence Berkeley National Laboratory, 19 December 2024)
• PJM auction procures 134,479 MW of generation resources (PJM news release, 17 December 2025)
• PJM capacity auction procures 138,318 MW of generation resources (PJM news release, 14 July 2026)
• PJM, ERCOT, and SWPP all set peak load records this summer (US Energy Information Administration, Today in Energy, 15 October 2026)
• Texas ERCOT data center energy grid (Texas Tribune, 17 June 2026)
• Short-Term Energy Outlook, October 2026 (US Energy Information Administration, released 6 October 2026)
• Texas S.B. 6 bill analysis, enrolled (Texas Legislature)
• Large load show cause orders, law firm summaries index (JD Supra, June 2026)

Last reviewed: 19 October 2026.