AI Data Center Energy Consumption Statistics 2026
By Axis Intelligence Research
Co-authors: Sarah Mitchell, AI & Machine Learning Analyst · Aidan Jad, EV, Green Tech & Infrastructure Analyst | Last updated: October 21, 2026 | License: CC BY 4.0
Data centers worldwide are forecast to consume 565 TWh of electricity in 2026, up 26.4% from 447 TWh in 2025, according to Gartner. Axis Intelligence Research finds that once cooling is allocated to the servers it serves, AI workloads account for 47.3% of that load, or 267.2 TWh, not the 31% headline figure.
Quick Answer: How Much Energy Do AI Data Centers Use?
Gartner forecasts global data center electricity consumption at 565 TWh in 2026, with AI-optimized servers drawing 175 TWh directly. The International Energy Agency (IEA) puts 2025 consumption at 485 TWh and projects 950 TWh by 2030. According to Axis Intelligence Research, AI-attributable electricity, including its share of cooling, reaches 267.2 TWh in 2026 and 395.4 TWh in 2027.
Key Findings
- Global data center electricity consumption is forecast to reach 565 TWh in 2026, a 26.4% increase from 447 TWh in 2025, according to Gartner’s June 2026 forecast.
- According to Axis Intelligence Research, AI workloads account for 47.3% of data center electricity in 2026 (267.2 TWh) once cooling and infrastructure load is allocated in proportion to server consumption.
- AI-focused data center electricity grew 50% in 2025, while total data center demand grew 17% and global electricity demand grew 3%, according to the IEA’s April 2026 report.
- Gartner’s 2030 floor of over 1,200 TWh sits at least 250 TWh, or 26.3%, above the IEA’s 950 TWh central projection, an Axis Intelligence Research reconciliation of the two forecasts.
- U.S. data centers used 176 TWh in 2023, or 4.4% of U.S. electricity, and could use 325 to 580 TWh by 2028, according to Lawrence Berkeley National Laboratory’s report for the Department of Energy.
How Much Electricity Do Data Centers Use Worldwide in 2026?
Global data centers will use 565 TWh in 2026 on Gartner’s numbers, and 485 TWh was the IEA’s figure for 2025. Both are estimates: no country meters data center electricity at global scale, so every global total is a model built from server shipments, power draw and facility overhead.
Gartner’s June 10, 2026 forecast is the only primary source that splits the total by workload. Its table, reproduced below, is the backbone of this page.
Data Center Electricity Consumption by Segment, Worldwide, 2025 to 2027
| Segment | 2025 (TWh) | 2025 growth | 2026 (TWh) | 2026 growth | 2027 (TWh) | 2027 growth | Source |
|---|---|---|---|---|---|---|---|
| Conventional servers | 193 | 0.9% | 195 | 1.2% | 200 | 2.4% | Gartner, Jun 2026 |
| AI-optimized servers | 95 | 83.6% | 175 | 84.2% | 258 | 47.8% | Gartner, Jun 2026 |
| Cooling and other infrastructure | 159 | 10.5% | 195 | 22.6% | 243 | 24.6% | Gartner, Jun 2026 |
| All data centers | 447 | 15.5% | 565 | 26.4% | 702 | 24.1% | Gartner, Jun 2026 |
Gartner notes that segments may not add to totals due to rounding. Figures are forecasts.
Read the growth columns before the totals. Conventional server load barely moves, 1.2% in 2026. AI-optimized servers grow 84.2%. According to Axis Intelligence Research, AI-optimized servers alone supply 67.8% of the 118 TWh added between 2025 and 2026. In 2027 they supply 60.6% of the increase, and their 258 TWh overtakes conventional servers at 200 TWh.
Gartner also forecasts worldwide data center power demand rising 27% to 132 GW in 2026, from 104 GW in 2025, and reaching 290 GW by 2030.
Aidan Jad: The cooling line is the one grid planners should circle. It grows 22.6% in 2026, against 1.2% for conventional servers, because every new AI rack drags its own heat-rejection load onto the interconnection request. A utility sizing a substation does not see “AI” and “cooling” as separate customers. It sees one meter that keeps climbing.
How Much of Data Center Electricity Is Actually AI?
AI workloads used 47.3% of data center electricity in 2026 on an attributed basis, according to Axis Intelligence Research, not the 31% widely quoted. The 31% figure is real, but it counts only the electricity flowing into AI-optimized servers. It leaves the chillers, pumps and power conversion serving those servers in a separate bucket.
That separation matters for anyone asking “how much electricity does AI use?” A cooling plant runs because the servers produce heat. Assigning all 195 TWh of 2026 cooling and infrastructure load to non-AI work, which the headline share implicitly does, understates AI’s footprint.
The AI Electricity Attribution Share (AEAS): An Axis Intelligence Research Metric
The AI Electricity Attribution Share (AEAS) measures the share of total data center electricity attributable to AI workloads after allocating cooling and infrastructure load in proportion to server consumption.
Formula: AEAS = AI-optimized server TWh ÷ (AI-optimized server TWh + conventional server TWh) AI-attributable TWh = AEAS × total data center TWh
Inputs: Gartner’s June 2026 segment table (above).
Worked example, 2026: 175 ÷ (175 + 195) = 47.30%. Then 47.30% × 565 TWh = 267.2 TWh.
| Year | AI-optimized servers (TWh) | Gartner headline AI share | AEAS | AI-attributable electricity (TWh) | Source |
|---|---|---|---|---|---|
| 2025 | 95 | n/a | 32.99% | 147.4 | Axis calculation from Gartner |
| 2026 | 175 | 31% | 47.30% | 267.2 | Axis calculation from Gartner |
| 2027 | 258 | n/a | 56.33% | 395.4 | Axis calculation from Gartner |
AEAS is an Axis Intelligence Research calculation. Inputs are Gartner estimates; AEAS readings are estimates.
What the AEAS shows: AI crosses from one-third of data center electricity in 2025 to more than half in 2027. The 2027 crossover that Gartner flags at server level arrives a year earlier, in 2026, once cooling follows the heat.
Scope of the allocation: AEAS assumes AI racks and conventional racks carry the same overhead per kWh of server load. Liquid-cooled AI halls can run lower overhead than legacy air-cooled rooms, which would pull AI’s true share somewhat below the AEAS reading. Dense AI halls in hot climates push the other way. Axis Intelligence Research publishes the proportional version because it is the only allocation reproducible from public inputs.
A Cross-Check: Gartner’s Split Implies a PUE of 1.527
Dividing Gartner’s total by its server load gives a power usage effectiveness (PUE)-like ratio. PUE is total facility energy divided by IT equipment energy; 1.0 would mean zero overhead. The ratio comes to 1.552 in 2025, 1.527 in 2026 and 1.533 in 2027.
The Uptime Institute’s 2025 Global Data Center Survey independently reports a weighted average PUE of 1.54. Two methods, one a forecast model and one an operator survey, land within a few hundredths of each other. According to Axis Intelligence Research, that agreement is the strongest public evidence that Gartner’s segment split is internally sound, and it is why AEAS uses it.
How Fast Is AI Data Center Electricity Demand Growing?
AI-focused data center electricity rose 50% in 2025, according to the IEA’s Key Questions on Energy and AI, published April 16, 2026. Total data center demand rose 17%, in line with the IEA’s earlier projection, and global electricity demand rose 3%.
Those three rates give two ratios Axis Intelligence Research tracks:
| Ratio | 2025 value | Formula | Source |
|---|---|---|---|
| AI Energy Intensity Ratio (AEIR) | 2.94 | AI-focused growth (50%) ÷ total data center growth (17%) | Axis calculation from IEA |
| AI-focused growth vs global electricity growth | 16.7 times | 50 ÷ 3 | Axis calculation from IEA |
| Data center growth vs global electricity growth | 5.7 times | 17 ÷ 3 | Axis calculation from IEA |
The AI Energy Intensity Ratio (AEIR) is a growth-speed measure: a reading of 1.0 would mean AI-focused facilities grow at the same pace as the whole sector. The 2025 reading of 2.94 is the baseline reading. AEAS answers a different question, how big AI’s share is; AEIR answers how fast it is pulling ahead.
The IEA also checks its own arithmetic. Its 2024 estimate of 415 TWh, grown by 17%, gives 485.6 TWh, matching the 485 TWh it reports for 2025.
Sarah Mitchell: Efficiency is not the brake people hope it is. The IEA says energy per AI task has been falling by at least an order of magnitude a year. Demand still grew 50%. When a resource gets ten times cheaper per unit of work and consumption rises anyway, the binding constraint is the grid connection, not the chip.
Gartner vs IEA: Why the 2030 Forecasts Disagree by 250 TWh
Gartner and the IEA nearly agree on the present and split sharply on 2030. For 2025, the IEA’s 485 TWh sits 38 TWh, or 8.5%, above Gartner’s 447 TWh. For 2030, the order flips: Gartner expects over 1,200 TWh, at least 250 TWh above the IEA’s 950 TWh.
Global Data Center Electricity Forecasts Compared
| Metric | Gartner (Jun 2026) | IEA (Apr 2026) | Axis reconciliation |
|---|---|---|---|
| 2025 consumption | 447 TWh | 485 TWh | IEA higher by 38 TWh |
| 2030 consumption | over 1,200 TWh | 950 TWh | Gartner higher by at least 250 TWh |
| Implied annual growth, 2025 to 2030 | at least 21.8% | 14.4% | Gartner path is steeper |
| AI-focused growth to 2030 | n/a | triples (24.6% a year) | Axis calculation from IEA |
Sources: Gartner press release, June 10, 2026; IEA, Key Questions on Energy and AI, April 16, 2026. Implied growth rates are Axis Intelligence Research calculations: (2030 value ÷ 2025 value) to the power 1/5, minus 1.
Gartner’s 2030 floor equals the IEA’s 2035 Base Case of 1,200 TWh, published in the IEA’s 2025 Energy and AI report. In other words, Gartner expects the industry to arrive five years earlier than the IEA does.
The difference is not a rounding dispute. The IEA explicitly writes that bottlenecks, from transformers to high-bandwidth memory, are reducing the likelihood of more aggressive near-term scenarios; it expects the memory shortage to last through at least the end of 2027. Gartner’s figure reflects demand pressure from AI buildout and warns that grid supply will be insufficient. One forecast prices in the queue; the other prices in the order book.
The IEA has barely moved its own number. Its 2025 report projected 945 TWh for 2030; the 2026 update says 950 TWh, a 5 TWh revision. Across all its 2025 sensitivity cases, 2035 demand ranged from 700 TWh to 1,700 TWh.
Aidan Jad: If you plan generation or sign a long-dated power purchase agreement, do not average these two numbers. They rest on different assumptions about whether turbines, transformers and memory arrive on time. The right move is to pick the bottleneck you believe and use the forecast that matches it.
How Much Electricity Do U.S. Data Centers Use?
U.S. data centers consumed 176 TWh in 2023, equal to 4.4% of total U.S. electricity, according to the Department of Energy’s release of the Lawrence Berkeley National Laboratory (LBNL) report. That was up from 58 TWh in 2014, a 3.0-times increase in nine years.
LBNL projects 325 to 580 TWh by 2028, between 6.7% and 12% of U.S. electricity. According to Axis Intelligence Research, that range implies annual growth of 13.1% to 26.9% from 2023.
U.S. Data Center Electricity: Official Data and Projections
| Year | Metric | Value | Source |
|---|---|---|---|
| 2014 | Data center electricity | 58 TWh | DOE / LBNL, Dec 2024 |
| 2023 | Data center electricity | 176 TWh (4.4% of U.S. total) | DOE / LBNL, Dec 2024 |
| 2025 | Server share of commercial-sector electricity | 7% | EIA, AEO2026 |
| 2028 | Data center electricity, scenario range | 325 to 580 TWh (6.7% to 12%) | DOE / LBNL, Dec 2024 |
| 2050 | Server electricity, case range | 446 to 818 TWh (22% to 33% of commercial buildings) | EIA, AEO2026 |
The U.S. Energy Information Administration (EIA) estimates that servers alone accounted for 7% of commercial-sector electricity in 2025 in its Annual Energy Outlook 2026 (AEO2026), released April 8, 2026. EIA also assumes data center floorspace needs up to 2.9 times as much cooling energy as other commercial floorspace.
Globally, the United States accounted for 45% of data center electricity in 2024, followed by China at 25% and Europe at 15%, per the IEA. The IEA’s 2025 report expects data centers to drive nearly half of U.S. electricity demand growth to 2030. For the household-scale comparison and grid pricing effects, see Axis Intelligence Research’s data center grid impact statistics.
Which Country Depends Most on Data Center Electricity?
Ireland is the clearest measured case. Data centres consumed 23% of Ireland’s metered electricity in 2025, up from 22% in 2024 and 5% in 2015, according to the Central Statistics Office Ireland. Urban homes used 18%, so Irish data centres now draw more metered power than every urban dwelling in the country combined.
Ireland matters beyond its size because CSO data is metered, not modeled. It is a rare national series that tests what global models assume about concentration. Project-level permitting responses are logged in the Axis Intelligence data center moratorium tracker, and quarterly demand announcements in the data center electricity demand tracker.
How Much Energy Does One AI Query Use?
A median Gemini text prompt used 0.24 Wh in May 2025, according to Google, and an average ChatGPT query uses about 0.34 Wh, according to OpenAI CEO Sam Altman. Both figures are company-reported and have not been independently verified.
| Measure | Value | Basis | Source |
|---|---|---|---|
| Gemini Apps median text prompt | 0.24 Wh | Includes idle capacity, host CPU and data center overhead | Google, Aug 2025 |
| Same prompt, active accelerators only | 0.10 Wh | Narrow boundary | Google, Aug 2025 |
| Reduction over 12 months | 33x | May 2024 vs May 2025 | Google, Aug 2025 |
| ChatGPT average query | 0.34 Wh | Average, boundary not detailed | OpenAI, Jun 2025 |
| All conventional searches run as simple AI text queries | less than 4 TWh a year | Less than 1% of data center use | IEA, Apr 2026 |
The boundary changes the answer by a factor of 2.4: Google’s full-boundary method gives 0.24 Wh, its narrow method 0.10 Wh. The ChatGPT figure is 1.42 times the Gemini figure, but one is an average and the other a median, so the comparison is indicative only.
Sarah Mitchell: Read the footnotes before the headline. Google’s 33x drop compares its own median prompt in May 2024 with May 2025, on its own hardware, and Google states the figures have not been verified by a third party. The number is plausible and useful. It is also a text-prompt median, and the workloads now growing fastest are not text prompts.
The IEA’s warning sits next to these small numbers. Video generation, reasoning and agentic tasks can use hundreds or thousands of times more energy per query than simple text generation. Per-query efficiency is improving at a rate the IEA calls unprecedented in energy history, yet total demand still rose 50% in 2025.
How Power-Dense Are AI Data Centers?
AI server power density rose 11 times between 2020 and 2025 and is set to rise another fourfold by 2027, according to the IEA. According to Axis Intelligence Research, that puts 2027 density at 44 times the 2020 level.
By 2027, one advanced rack the size of a large refrigerator could have peak demand equal to 65 households. The IEA also reports AI racks shedding or adding more than 50% of rated load within a second, and needing to evacuate heat equivalent to 30 natural gas boilers per rack. A typical AI-focused data center already uses as much electricity as 100,000 households; the largest under construction will use 20 times as much.
The installed base has not caught up. Uptime Institute’s 2025 survey found the average of operators’ typical rack densities at almost 9 kW, up from 8.3 kW in 2024. More than 80% of operators have no rack above 30 kW. Facilities with racks of 50 kW or more rose to 9% in 2025 from 6% in 2023.
Aidan Jad: A 9 kW average next to a 65-household rack is not a contradiction. It describes two industries sharing one name. Most enterprise rooms will never host AI training. The new load concentrates in a small number of very large campuses, which is exactly why it lands on a handful of substations rather than spreading across the grid.
How Efficient Are Data Centers? PUE Statistics 2025
The global average PUE was 1.54 in 2025, according to Uptime Institute, the sixth consecutive year it has barely moved. It fell 38.4% from 2.50 in 2007, mostly before 2014, when it reached 1.65. Facilities commissioned since around 2020 average 1.48, and data centers of 20 MW and above average 1.44.
Google reports a fleet-wide PUE of 1.09 for 2025 and says that means 83% less overhead energy than the industry average, per its data center efficiency page. This is company-reported.
| Benchmark | PUE | Source |
|---|---|---|
| Global survey average, 2025 | 1.54 | Uptime Institute |
| Facilities commissioned since around 2020 | 1.48 | Uptime Institute |
| Data centers 20 MW and above | 1.44 | Uptime Institute |
| Google fleet, 2025 | 1.09 | Google (company-reported) |
| Implied by Gartner 2026 segment split | 1.527 | Axis calculation from Gartner |
The efficiency gap in TWh: if all 370 TWh of 2026 server load ran at Google’s 1.09, total data center electricity would be 403.3 TWh instead of 565 TWh. According to Axis Intelligence Research, that illustrative scenario removes 161.7 TWh, or 28.6% of the 2026 total. It cannot happen fleet-wide, because climate and legacy buildings set floors, but it sizes the prize that liquid cooling and new builds compete for.
What Powers AI Data Centers, and What Are the Emissions?
The tech sector signed around 40% of all corporate renewable power purchase agreements (PPAs) in 2025, according to the IEA. That is a share of the corporate PPA market, not a share of data center demand covered by renewables.
Firm power is moving faster. The pipeline of conditional offtake agreements between data center operators and small modular reactor (SMR) projects grew from 25 GW at the end of 2024 to 45 GW by April 2026, an 80% increase. Deal-level detail is tracked on Axis Intelligence Research’s nuclear energy for data centers page.
Grid delays are pushing U.S. developers toward onsite gas. The IEA estimates 15 to 27 GW of onsite natural gas may power data centers by 2030, mostly in the United States, and that reliable onsite gas supply needs 30% to 70% more generation than the load it serves. Gas turbine orders surged 70% in 2025. The IEA also expects 20 to 25 GW of battery storage in data centers by 2030 to manage load swings.
On emissions, the IEA’s 2026 update projects data center emissions roughly doubling to around 350 million tonnes in 2035, about 2% of electricity-sector emissions. Its 2025 report had projected 300 Mt in the 2035 Base Case, up from 180 Mt. Axis Intelligence Research notes the 2035 figure was revised up by 50 Mt, or 16.7%, in one year, while the 2030 electricity projection moved only 5 TWh. The revision sits in the fuel mix, not the demand curve.
Methodology
Collection. Every figure on this page was retrieved from its producer on October 21, 2026: Gartner (forecast), the IEA (two reports and one press release), the U.S. Department of Energy and LBNL, the EIA, the Central Statistics Office Ireland, Uptime Institute, Google and OpenAI. Gartner, IEA and Uptime figures are estimates from measurement or modeling organizations. Google and OpenAI figures are company-reported. Each figure is a row in the CSV with its source URL and retrieval date.
Axis Intelligence Research calculations.
- AEAS = AI-optimized server TWh ÷ (AI-optimized + conventional server TWh); AI-attributable TWh = AEAS × total TWh. Inputs: Gartner, June 2026.
- Implied overhead ratio = Gartner total TWh ÷ (AI-optimized + conventional server TWh).
- AEIR = IEA AI-focused growth ÷ IEA total data center growth.
- Implied annual growth = (end value ÷ start value)^(1/years) − 1.
- PUE scenario = 2026 server TWh × 1.09; avoided TWh = 565 − scenario total.
Why we do not merge Gartner and the IEA. The two use different models and different definitions of a data center’s boundary. Averaging them would produce a number neither organization supports, so this page publishes both and quantifies the gap.
Scope. No primary source publishes AI-specific electricity in TWh; AEAS is a transparent allocation, not a measurement. Gartner’s segments do not separate storage or networking. Per-query figures cover text prompts only and exclude training. Global and regional totals are modeled; only Ireland’s figures here are metered.
Corrections policy. Errors can be reported to [email protected].
About This Dataset
Title: AI Data Center Energy Consumption Statistics 2026 Creator and publisher: Axis Intelligence Research
Contents: 160 rows covering global data center electricity by segment (2025 to 2027), IEA and Gartner forecasts to 2030 and 2035, U.S. official data (2014 to 2050), Ireland metered shares, PUE benchmarks, per-query energy, power density, supply and emissions, every Axis calculation with its formula, and retracted figures from the previous version with reasons. Coverage: Global, United States, Ireland; 2007 to 2050.
License: CC BY 4.0.
Citation: Axis Intelligence Research, AI Data Center Energy Consumption Statistics 2026, axis-intelligence.com/ai-data-center-energy-consumption-statistics/
Cite This Research
APA: Axis Intelligence Research. (2026). AI data center energy consumption statistics 2026. Axis Intelligence. https://axis-intelligence.com/ai-data-center-energy-consumption-statistics/
MLA: Axis Intelligence Research. “AI Data Center Energy Consumption Statistics 2026.” Axis Intelligence, 21 Oct. 2026, axis-intelligence.com/ai-data-center-energy-consumption-statistics/.
Chicago: Axis Intelligence Research. “AI Data Center Energy Consumption Statistics 2026.” Axis Intelligence, October 21, 2026. https://axis-intelligence.com/ai-data-center-energy-consumption-statistics/.
Frequently Asked Questions About AI Data Center Energy Consumption
Is AI 31% or nearly half of data center electricity?
Both, depending on what you count. Gartner’s 31% for 2026 counts only electricity into AI-optimized servers. Allocating cooling and infrastructure in proportion to server load, Axis Intelligence Research’s AEAS puts AI at 47.3% of data center electricity in 2026, or 267.2 TWh.
Will data centers use 1,200 TWh of electricity by 2030?
The two main forecasts disagree. Gartner expects over 1,200 TWh by 2030; the IEA projects 950 TWh. The gap of at least 250 TWh reflects whether transformer, turbine and memory bottlenecks hold back construction, which the IEA expects through at least 2027.
How much has AI data center electricity grown compared with the grid?
In 2025, AI-focused data center electricity grew 50%, total data center demand grew 17% and global electricity demand grew 3%, per the IEA. That makes AI-focused growth 16.7 times faster than the grid as a whole, an Axis Intelligence Research calculation.
Does a ChatGPT or Gemini query use a lot of electricity?
Not individually. Google reports 0.24 Wh for a median Gemini text prompt and OpenAI’s CEO cites about 0.34 Wh for an average ChatGPT query, both company-reported. The IEA says reasoning, agentic and video tasks can use hundreds or thousands of times more per query.
What share of U.S. electricity goes to data centers?
U.S. data centers used 4.4% of U.S. electricity in 2023, or 176 TWh, according to DOE and LBNL. The same report projects 6.7% to 12% by 2028. EIA separately estimates servers alone used 7% of commercial-sector electricity in 2025.
Why hasn’t data center efficiency improved since 2019?
Uptime Institute’s average PUE has stayed near 1.54 for six years because legacy facilities keep running and hot climates limit efficient cooling. New builds are better: facilities from around 2020 onward average 1.48, and Google reports 1.09 for its fleet.
Will AI data center emissions double?
Yes, on IEA projections. Data center emissions reach around 350 million tonnes by 2035, about 2% of electricity-sector emissions. The IEA raised that 2035 figure from 300 Mt in its 2025 report, even though its 2030 electricity projection changed by only 5 TWh.
How dense will AI server racks get?
The IEA expects one advanced rack to reach peak demand equal to 65 households by 2027, after an 11-times density increase from 2020 to 2025 and a further fourfold rise. Most of today’s installed base is far lower: Uptime puts the average typical rack at almost 9 kW.
Related Axis Intelligence Research: AI data center statistics · AI data center forecast 2030 · AI data center water usage statistics · data center construction cost statistics