Data Center Power Demand Forecasts Differ by 222 TWh in 2030

Two public documents give different values for the same quantity. Berkeley Lab’s 2025 update puts U.S. data center electricity use in 2030 at 649 TWh in its Reference Case, while the IEA’s Energy and AI report implies about 427 TWh for the same country and year. The 222 TWh between them is larger than all U.S. data center use in 2024. The reason they split decides how much a cooling plan can lean on any data center power demand forecast.

The two U.S. outlooks share a 2024 base and split on shipments

Berkeley Lab estimates U.S. data center electricity use at 192 TWh in 2024. The IEA puts the global figure at 415 TWh for that year and assigns the United States 45% of it. The IEA page prints the U.S. increase to 2030, not the total.

IEA U.S. base: 45% × 415 TWh = 186.75 TWh, which is within 3% of Berkeley Lab’s 192 TWh. IEA U.S. 2030: 186.75 TWh + 240 TWh = 427 TWh. Cross-check on the same page’s “up 130%”: 240 TWh ÷ 1.30 = 185 TWh as the base, so 425 TWh in 2030. Both are derived here, and Berkeley Lab’s text describes the IEA projection as roughly 430 TWh.

The bases agree and the paths do not. The sixth root of (649 ÷ 192), minus 1, is 22.5% a year for Berkeley Lab. The sixth root of (427 ÷ 187), minus 1, is 14.8% a year for the IEA. Neither rate is published; both are computed here.

Berkeley Lab attributes the divergence to assumptions about server and accelerator shipment trajectories, and calls the IEA method broadly consistent with its own. The IEA names near-term server shipment projections as the key input to its modeling.

Source: Berkeley Lab, United States Data Center Energy Usage Report: 2025 Update (June 2026), Sections 2 and 3; IEA, Energy and AI, “Executive summary” and “Energy demand from AI” (2025).

Table 1. U.S. data center electricity in 2030 by outlook (TWh, 2030)

* Derived: 45% of 415 TWh plus the 240 TWh increase; not printed in the report. PUE is not varied across the Berkeley Lab scenarios.

Source: Lawrence Berkeley National Laboratory, “United States Data Center Energy Usage Report: 2025 Update” (June 2026); IEA, “Energy and AI” (2025)

The IEA adds more than 45 TWh in Europe by 2030. The European Commission, citing the IEA, gives 70 TWh for the EU in 2024 and 115 TWh by 2030.

The IEA’s 15% European share of 415 TWh is about 62 TWh, below the Commission’s EU figure for the same year. The pages read do not show how the IEA draws the European boundary, and the Commission notes that precise consumption data are missing. Until the boundary is known, the European series cannot size anything.

Berkeley Lab dropped its PUE uncertainty range between editions while the 2028 total held

The 2024 report treated cooling as a scenario input. Average U.S. PUE falls from 1.4 in 2023 to a 2028 band of 1.15 to 1.35, which the report ties to the shift into hyperscale and colocation facilities and to liquid-cooled AI servers.

For 2024 to 2028 it also varied the PUE and WUE of each space type by plus or minus 10%.

The 2025 update drops that lever. It estimates an average national PUE of 1.145 in 2024 for facilities serving AI equipment, easing to 1.136 in 2030, and an aggregate that reaches 1.36 in 2030 in the Reference Case. PUE uncertainty is no longer carried because the authors hold measured or simulated data for it. It also states that evolving cooling designs do not materially affect its totals.

The 2025 aggregate for 2030 sits above the top of the 2024 band for 2028. The editions also start from different baselines, 1.4 for 2023 against 1.45 for 2024. The aggregate path has therefore moved from a band to a single line, while the AI-facility value stays near 1.14.

The headline barely moved. The update’s 2028 Reference Case of 464 TWh falls in the middle of the earlier 325 to 580 TWh range.

Infrastructure energy = total energy × (1 − 1 ÷ PUE). It covers cooling and power distribution together and is a typical value derived from the aggregate PUE. For 2024: 192 TWh × (1 − 1 ÷ 1.45) = 59.6 TWh, which matches the 31% infrastructure share the update reports. For 2030: 649 TWh × (1 − 1 ÷ 1.36) = 171.8 TWh. Share of growth: (171.8 − 59.6) ÷ (649 − 192) = 112 ÷ 457 = 24.6%. IEA global: 20% × (945 − 415) TWh = 106 TWh.

The IEA puts cooling and other infrastructure at around 20% of net global growth. Two institutions with different methods land at a fifth to a quarter of growth, which makes cooling a follower of IT expansion in both outlooks.

PUE sensitivity for 2030: IT energy = 649 TWh ÷ 1.36 = 477 TWh. Holding the 2024 aggregate of 1.45 gives 477 TWh × 1.45 = 692 TWh, which is 43 TWh higher. Running every facility at the AI-facility value of 1.145 gives 546 TWh, which is 103 TWh lower. Both are illustrative bounds.

Even the favorable bound recovers less than half of the 222 TWh between the two outlooks. Shipments decide the total, and an energy forecast cannot settle a cooling choice on its own.

Source: Berkeley Lab, 2024 Report, Sections 4 and 5; 2025 Update, Sections 2, 3.5 and 3.6; IEA, Energy and AI, “Energy demand from AI” (2025).

How much IT load does one grid connection carry after cooling?

Energy forecasts become power capacity only through a utilization assumption. Berkeley Lab’s 2024 report converted its 2028 range into 74 to 132 GW at 50% average capacity utilization. The 2025 update applies the same 50% to its 2030 Reference Case and warns that converting annual consumption to nameplate or requested capacity is hard.

Interconnection capacity = annual energy ÷ 8,760 h ÷ utilization. For 2030: 649 TWh ÷ 8,760 h = 74.1 GW average draw, and 74.1 GW ÷ 0.50 = 148 GW. The same steps on the 2024 base give 192 TWh → 21.9 GW → 43.8 GW. The difference of 104 GW over six years is 17.4 GW a year, the figure the update reports.

A fixed utility connection serves IT load = connection ÷ PUE, if the stated PUE holds at peak. For a 100 MW connection: 100 ÷ 1.45 = 69.0 MW of IT at the 2024 national aggregate, 100 ÷ 1.36 = 73.5 MW at the 2030 aggregate, and 100 ÷ 1.145 = 87.3 MW at the AI-facility value. These are idealized upper bounds, since annual-average PUE does not guarantee the same ratio at peak.

The same move from 1.45 to 1.145 cuts facility energy for a given IT load by 21%. On a connection that cannot grow it lifts IT load by 26.6%. The annual-cost side of that gap is worked in what a PUE 1.05 facility saves over a PUE 1.4 facility in a year.

The IEA notes that a data center can be operational in two to three years, while the energy system needs longer lead times. The grid capacity data center projects can secure at the start is therefore the fixed quantity, and the cooling design decides how much IT load fits inside it.

Source: Berkeley Lab, 2024 Report, Executive Summary; 2025 Update, Section 2 “Power Capacity”; IEA, Energy and AI, “Energy demand from AI” (2025).

Korea’s 2028 figures count transformers, servers and summer peaks

Savills Korea’s 2025 market report puts total data center receiving capacity at about 1.9 GW in 2024 and about 4.8 GW in 2028, a 26.4% compound annual rate in the report’s own wording. Hankyung Market Insight’s account of the report defines receiving capacity as the sum of transformer capacity a customer installs to take power from outside. The scope is commercial data centers, and the release warns that supply limits, permitting and local opposition may delay planned projects.

Korea IDC publishes a different series for the same country and years. Its data center power demand rises from 4,461 MW in 2025 to 6,175 MW in 2028, or 11% a year, as reported by Sisa Journal-e. Korea IDC says the figure adds enterprise data centers, including small in-house server rooms, to service-provider facilities, and that it is estimated from the number of servers supplied to the market.

The government series sits lower. The Ministry of Trade, Industry and Energy’s 11th Basic Plan for Electricity Supply and Demand, notice 2025-169 of February 2025, projects data center summer peak power of 1.2 GW in 2025 and 3.3 GW in 2030. It reaches 6.2 GW in 2038. The plan multiplies new contracted capacity by a utilization rate and assumes PUE improves to the global average.

Same-year comparison for 2025: 4,461 MW ÷ 1,200 MW = 3.7. One series estimates demand from servers supplied, and the other is the summer peak the grid plan carries.

The plan’s own table shows a second gap. Applications to KEPCO for data center supply reach a cumulative 7,343 MW by 2027, and 4,718 MW of that was judged supplyable in a review dated August 2023. The plan’s total summer peak for 2027 is 2.3 GW.

Contracted supply capacity ÷ planned peak = 4,718 MW ÷ 2,300 MW = 2.05. The contracts alone are about twice the peak the plan expects. A contract number is a ceiling on what a site may draw, not a forecast of the heat it will reject.

A December 2025 Cushman & Wakefield Korea report states that consumption grows from 5.0 TWh in 2023 to 15.5 TWh by 2038. In the plan’s table 15.5 TWh is the 2038 increase over trend, and the total for that year is 30.0 TWh. A reader who takes the consultancy’s sentence in place of the table halves the 2038 energy figure.

Source: Ministry of Trade, Industry and Energy, 11th Basic Plan for Electricity Supply and Demand (2025), Reference 6.

Energy outlooks disagree about shipments and capacity figures disagree about definitions

The U.S. spread is a forecast question. It narrows or widens as GPU shipments, chip lifetimes and inference utilization become known, and Berkeley Lab’s single sensitivities run from 11% below to 21% above its Reference Case. The Korean spread is a reading question. Savills, Korea IDC and the ministry measure different things, and the gap between them closes by reading the definitions, not by waiting.

That difference shapes how a cooling plan is written. A plan tied to a utility contract inherits the contract’s ceiling, and the PUE it can hold at that ceiling sets the IT load. A plan tied to shipment-driven growth needs a bracket, and the 521 to 843 TWh compounded range gives one. Neither choice follows from the forecasts alone, and qualified engineering review is required before any of it reaches a design.

The limits are specific. The IEA’s U.S. 2030 value is derived here, not printed. Korea IDC is read through trade-press accounts because its release could not be opened. The ministry’s application figures are a cumulative snapshot from August 2023. Annual-average PUE is not a peak value, and none of the forecasts states rack density, which is what separates air from liquid at the room level.

Table 2. What each forecast counts and how far it reaches (Verdict by case)

* Derived in this article from IEA figures. Rows are not on a common unit and are not sorted by value.

Source: Berkeley Lab (2026); IEA (2025); European Commission (2025); Savills Korea (2025); Sisa Journal-e on Korea IDC (2025); Ministry of Trade, Industry and Energy, 11th Basic Plan (2025)

A phased approach, such as retrofitting liquid cooling without rebuilding the data center, can follow whichever U.S. case materializes, while a single build-out commits to one. A Korean plan should name the series it sized to, because 2025 carries 1.2 GW in one source and 4,461 MW in another.

References

• United States Data Center Energy Usage Report: 2025 Update / Lawrence Berkeley National Laboratory / June 2026 / https://eta-publications.lbl.gov/publications/united-states-data-center-energy-2025
• 2024 United States Data Center Energy Usage Report (LBNL-2001637) / Lawrence Berkeley National Laboratory / December 2024 / https://eta-publications.lbl.gov/sites/default/files/2024-12/lbnl-2024-united-states-data-center-energy-usage-report_1.pdf
• Energy and AI, “Energy demand from AI” / International Energy Agency / 2025 / https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai
• Energy and AI, “Executive summary” / International Energy Agency / 2025 / https://www.iea.org/reports/energy-and-ai/executive-summary
• In focus: Data centres, an energy-hungry challenge / European Commission, Directorate-General for Energy / November 17, 2025 / https://energy.ec.europa.eu/news/focus-data-centres-energy-hungry-challenge-2025-11-17_en
• 제11차 전력수급기본계획 (2024~2038), 산업통상자원부 공고 제2025-169호 (11th Basic Plan for Electricity Supply and Demand, Ministry of Trade, Industry and Energy notice 2025-169) / Ministry of Trade, Industry and Energy, Korea (file hosted by Korea Power Exchange) / February 21, 2025 / https://new.kpx.or.kr/boardDownload.es?bid=ATT&list_no=74548&seq=1
• 2025년 한국 데이터센터 시장 리포트 발행 (Savills Korea 2025 Korea Data Center Market Report, news release) / Savills Korea / July 2025 / https://www.savills.co.kr/insight-and-opinion/savills-news/223374/-2025%EB%85%84-%ED%95%9C%EA%B5%AD-%EB%8D%B0%EC%9D%B4%ED%84%B0%EC%84%BC%ED%84%B0-%EC%8B%9C%EC%9E%A5-%EB%A6%AC%ED%8F%AC%ED%8A%B8–%EB%B0%9C%ED%96%89
• 한국IDC “데이터센터 전력 수요 2028년까지 연평균 11% 성장” (Korea IDC: data center power demand to grow 11% a year to 2028) / Sisa Journal-e, reporting Korea IDC / June 30, 2025 / https://www.sisajournal-e.com/news/articleView.html?idxno=412958
• 국내 데이터센터 수요 급증…3년간 年 26% ‘고속 성장’ (Domestic data center demand surges, 26% a year for three years; report on Savills Korea’s market report) / Hankyung Market Insight / July 3, 2025 / https://marketinsight.hankyung.com/article/202507035396r
• Data Centres & the Power Challenge / Cushman & Wakefield Korea / December 2025 / https://assets.cushmanwakefield.com/-/media/cw/apac/south-korea/insights/research/cwresearchdatacentresthepowerchallenge2025eng.pdf