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: July 23, 2026 | License: CC BY 4.0
Global data center electricity consumption reached 565 TWh in 2026 — a 26.4% increase from 447 TWh in 2025 — driven overwhelmingly by AI-optimized servers, which alone account for 31% of all data center power this year. The IEA projects this figure to nearly double by 2030, crossing 950 TWh and approximately 3% of all global electricity demand.
Quick Answer
Global data centers will consume 565 TWh of electricity in 2026, up 26.4% from 447 TWh in 2025, according to Gartner’s June 2026 Forecast. AI-optimized servers now represent 31% of data center power draw and will surpass conventional servers in 2027. The IEA’s April 2026 report projects consumption doubling to 950 TWh by 2030, with AI-specific facilities tripling in that window. U.S. data center servers accounted for an estimated 7% of all commercial-sector electricity in 2025, per the EIA’s Annual Energy Outlook 2026.
Key Findings
- Global data center electricity demand hit 565 TWh in 2026, a 26.4% year-over-year increase, with AI-optimized servers the primary driver — accounting for 31% of all data center power consumption and growing at 84.2% annually (Gartner, June 2026).
- AI-focused data centers surged 50% in electricity consumption in 2025, growing at more than 16 times the rate of overall global electricity demand, which rose just 3%, according to the IEA’s Key Questions on Energy and AI (April 2026).
- Between 2020 and 2025, AI server power density increased 11 times, and the IEA projects a further fourfold increase by 2027 — meaning a single refrigerator-sized rack could draw power equivalent to 65 households.
- U.S. data centers consumed 176 TWh in 2023, equal to 4.4% of total U.S. electricity; the Lawrence Berkeley National Laboratory projects that share rising to between 6.7% and 12% by 2028.
- According to Axis Intelligence Research’s AI Energy Intensity Ratio (AEIR™), AI-focused data center electricity growth ran at 2.94× the rate of overall data center electricity growth in 2025 — the highest ratio recorded since the metric’s baseline year.
How Much Electricity Do Data Centers Use Globally?
The cleanest answer is also the most uncomfortable one for grid planners: demand is accelerating faster than the infrastructure built to meet it. Gartner published the most granular current breakdown in June 2026, splitting data center power by workload type across a three-year window:
| Segment | 2025 TWh | 2025 Growth | 2026 TWh | 2026 Growth | 2027 TWh | 2027 Growth |
|---|---|---|---|---|---|---|
| Conventional Server | 193 | +0.9% | 195 | +1.2% | 200 | +2.4% |
| AI-Optimized Server | 95 | +83.6% | 175 | +84.2% | 258 | +47.8% |
| Cooling & Infrastructure | 159 | +10.5% | 195 | +22.6% | 243 | +24.6% |
| Total | 447 | +15.5% | 565 | +26.4% | 702 | +24.1% |
Source: Gartner, June 2026. Note: figures may not sum precisely due to rounding.
The table makes the structural shift unmistakable. Conventional server consumption is essentially flat — growing less than 2.5% even by 2027. AI-optimized servers, running at 84% growth this year, are the entire story. By 2027, their power draw (258 TWh) will exceed that of conventional servers (200 TWh) — a crossover Gartner identifies as the defining event in data center power history.
The IEA’s perspective is consistent. Its April 2026 report Key Questions on Energy and AI — the most authoritative public dataset on this topic — finds that overall data center electricity grew 17% in 2025, while AI-specific facilities grew 50%. That gap is not a rounding artifact; it reflects a structural divergence between workloads that have plateaued (traditional enterprise computing) and workloads that are compounding (inference and training for large AI models).
For the IEA’s base case, total data center consumption roughly doubles from 485 TWh in 2025 to 950 TWh in 2030, reaching approximately 3% of global electricity demand by that date. AI-focused data centers triple in the same window.
Aidan Jad’s read on the 2026–2030 trajectory: this is not a demand curve that bends smoothly. Power density per rack — the single variable that drives everything downstream, from cooling design to grid interconnection size — is on a discontinuous path. The 2027 crossover, when AI racks consume more than conventional servers industry-wide, is the inflection point that changes grid planning horizons from five years to fifteen.
How Much Data Center Electricity Does the United States Use?
The U.S. is the world’s dominant data center market, and the gap between U.S. and global growth trajectories has widened in 2025–2026.
The most authoritative U.S. dataset remains the December 2024 report from the Lawrence Berkeley National Laboratory (LBNL), published for the U.S. Department of Energy under the Energy Act of 2020. LBNL’s findings:
- U.S. data centers consumed 176 TWh in 2023, equal to 4.4% of total U.S. electricity
- The compound annual growth rate rose from approximately 7% (2014–2018) to 18% (2018–2023)
- LBNL projects 2028 U.S. data center consumption in a range of 325–580 TWh, representing 6.7%–12% of total U.S. electricity
The U.S. Energy Information Administration’s Annual Energy Outlook 2026 (released May 19, 2026) sharpens the picture for the nearer term. EIA now estimates that data center servers accounted for 7% of all commercial-sector electricity in 2025 — and projects server electricity use rising to 22%–33% of commercial building electricity use by 2050. The EIA’s AEO2026 projects total U.S. electricity demand reaching 4,283 billion kWh in 2026, with the commercial sector — led by data centers — as the fastest-growing segment, expanding 5% annually.
One structural dynamic that aggregate national figures obscure: U.S. data center demand is geographically concentrated. Northern Virginia (Dominion Energy’s territory within PJM) hosts the largest data center cluster on earth. Capacity auction prices for that zone rose more than tenfold between the 2024 and 2026–2027 delivery years, a pricing signal that reflects the true cost of serving dense, continuous AI workloads on a grid that was not designed for them.
U.S. Data Center Electricity: Historical and Forecast
| Year | U.S. TWh | % of U.S. Electricity | Source |
|---|---|---|---|
| 2014 | ~60 | ~1.5% | LBNL 2024 |
| 2018 | ~90 | ~2.3% | LBNL 2024 |
| 2023 | 176 | 4.4% | LBNL 2024 |
| 2025 | ~200–220 (est.) | ~7% commercial servers | EIA AEO2026 |
| 2028 (forecast) | 325–580 | 6.7%–12% | LBNL 2024 |
| 2050 (server only, forecast) | 446–818 | 22%–33% commercial | EIA AEO2026 |
Sources: LBNL 2024 United States Data Center Energy Usage Report (primary); EIA Annual Energy Outlook 2026 (primary).
AI Data Center Energy Consumption vs. Traditional Data Centers
The power density comparison between AI and conventional workloads is one of the most commonly misunderstood dimensions of this topic. Global averages mask the physical reality inside an AI-optimized facility.
Standard colocation racks draw 5–10 kW. AI-optimized GPU racks currently draw 40–80+ kW, with NVIDIA’s next-generation configurations reaching 120–600 kW. The IEA’s April 2026 report documents that between 2020 and 2025, AI server power density increased 11 times — and projects a further fourfold increase by 2027, at which point a single rack the size of a household refrigerator would carry peak demand equivalent to 65 households.
This density jump cascades through every downstream cost: cooling infrastructure (liquid cooling versus air cooling, which consumes 20–25% of total power in conventional facilities), electrical distribution architecture (the sector is migrating toward high-voltage direct current systems), structural requirements (an AI rack can weigh more than a pickup truck), and grid interconnection sizing.
Power Usage Effectiveness (PUE) — total facility power divided by IT equipment power, where 1.0 is perfect — varies sharply across the installed base:
- Hyperscale leaders (Google, Meta): PUE of 1.08–1.09 (2025 reported figures)
- Modern colocation facilities: PUE of 1.3–1.6
- Enterprise data centers: PUE of 1.5–1.8
- Global average: approximately 1.5–1.6
The practical implication: for every 100 MW of AI compute load in an average-efficiency facility, another 50–60 MW is consumed by cooling, power distribution, and other overhead. In a best-in-class hyperscale facility, that overhead drops to 8–9 MW. The gap is the efficiency opportunity — and it is why hyperscalers’ capital expenditure allocation to liquid cooling infrastructure has grown from a footnote to a headline item.
Inference — running trained models for users — now accounts for 80–90% of AI computing by volume, and will represent approximately 75% of total AI energy demand by 2030, per IEA 2025 analysis. Training events are episodic but intensely power-dense. The IEA notes that the energy required per individual AI task has fallen by at least an order of magnitude annually in recent years; simple text queries now use less electricity than a television runs over the same period. But energy-intensive use cases — video generation, autonomous agents, reasoning models — consume hundreds to thousands of times more per query than simple text generation, and their adoption is accelerating.
Energy Per Query: Representative AI vs. Conventional Tasks
| Task | Approximate Energy | Notes |
|---|---|---|
| Google Search (traditional) | ~0.0003 kWh | Pre-AI baseline |
| Simple AI text query | <0.001 kWh | IEA estimate, 2026 |
| If all searches used simple AI text | <4 TWh/year total | <1% of current data center use |
| AI video generation (per output minute) | 10–100× text query | Range; dependent on model |
| AI reasoning/agentic task | 100–1,000× text query | IEA estimate range |
Sources: IEA, Key Questions on Energy and AI (April 2026); IEA, Energy and AI (2025).
AI Data Center Energy Consumption by Region
Data center demand is not distributed uniformly across the planet. Three regions — North America, Europe, and Asia-Pacific — account for the overwhelming majority of installed capacity, but their trajectories differ materially.
United States remains the world’s single largest data center electricity consumer, driven by the Northern Virginia cluster, followed by Phoenix, Dallas–Fort Worth, Silicon Valley, and Chicago. The EIA’s AEO2026 projects the commercial sector — dominated by data centers — growing 5% in 2026, the fastest rate recorded.
Europe presents a constrained growth picture. Ireland alone draws approximately 21% of its national electricity from data centers. The Netherlands, Germany, and Sweden host significant concentrations, and each is encountering grid capacity limits, permitting delays, and community resistance that are slowing buildout. The IEA notes that permitting and regulatory systems in multiple jurisdictions are “being stretched by the wave of project applications.”
Asia-Pacific is the fastest-growing region by number of projects announced, though its installed base lags North America. Singapore has imposed data center construction moratoriums and then partial lifts in response to grid and water concerns. Japan, South Korea, and India are all seeing acceleration driven by domestic AI investment and government policy support.
China operates largely outside the IEA’s primary dataset (the IEA does not include China in its “advanced economies” tracking), but satellite-based capacity tracking used in the IEA’s April 2026 report confirms that “AI factories” — cutting-edge training facilities — have more than tripled in capacity globally over the past 18 months, with significant Chinese contribution.
Axis Intelligence Research’s view: the regional divergence that matters most for 2026–2028 is not between continents — it is within the United States, between states with available grid capacity (Texas, Ohio, Georgia) and those with constrained interconnection queues (Northern Virginia, the Pacific Northwest). The average wait time for a new U.S. grid connection is now four years. That four-year lead time is the binding constraint on AI scaling plans, not computing hardware.
Regional Data Center Power Demand Indicators, 2025–2026
| Region / Market | Key Indicator | Source |
|---|---|---|
| Northern Virginia (VA) | Summer peak load +23% since 2019; winter +45% | PJM / cited in IEA 2026 |
| Ireland | ~21% of national electricity from data centers | IEA 2026 |
| Singapore | Active moratorium-then-partial-lift on new construction | Singapore EMA |
| United States (total) | 176 TWh in 2023; est. 200–220 TWh in 2025 | LBNL 2024; EIA 2026 |
| Global (all data centers) | 447 TWh (2025); 565 TWh (2026 forecast) | Gartner, June 2026 |
| Global (AI-focused only) | +50% in 2025; tripling projected by 2030 | IEA, April 2026 |
What Powers AI Data Centers? The Energy Mix Shift
In 2025, hyperscalers signed renewable power purchase agreements (PPAs) covering approximately 40% of their data center electricity demand, according to the IEA. That figure is significant — but it reveals the gap as much as the progress. The remaining 60% is drawing from the grid’s existing generation mix, which in the U.S. still includes substantial natural gas and coal.
The more consequential 2025–2026 development is the scale of nuclear commitments:
- Meta executed agreements in January 2026 to procure up to 6.6 GW of nuclear power, including a 20-year PPA with Constellation Energy for 1.1 GW from the Clinton Clean Energy Center, and forward agreements with advanced reactor developers TerraPower and Oklo.
- Amazon struck a $20 billion arrangement with Talen Energy for nuclear power from its Susquehanna facility (1.9 GW), alongside a separate $50 billion partnership with X-energy for 960 MW of small modular reactors (SMRs).
- Google signed a 25-year PPA with NextEra Energy for 615 MW from the restarted Duane Arnold nuclear plant in Iowa, and separately committed to 1,800 MW with Elementl Power for three SMR sites.
The IEA’s satellite-based tracking of SMR conditional offtake agreements shows the pipeline growing from 25 GW at end-2024 to 45 GW by April 2026 — an 80% expansion in 18 months. Big tech has collectively signed contracts for more than 10 GW of possible new U.S. nuclear capacity in the past year.
Why nuclear, specifically? The answer is not ideological — it is operational. AI data centers require near-100% uptime with flat, predictable load. Intermittent renewables require storage or backup. Nuclear provides 24/7 baseload at a predictable cost, which is the profile that best matches an inference workload running continuously at high GPU utilization.
The IEA notes a parallel development: constrained grid connections are pushing some U.S. data center operators toward onsite natural gas generation. Satellite tracking of these projects indicates that around one-fifth have begun land clearing or construction. The IEA estimates 15–27 GW of onsite natural gas could be powering U.S. data centers by 2030 — a development with significant carbon implications.
Hyperscaler Nuclear Power Agreements, 2025–2026
| Company | Capacity | Technology | Counterparty | Status |
|---|---|---|---|---|
| Meta | Up to 6.6 GW | Existing nuclear + advanced reactors | Constellation, Vistra, TerraPower, Oklo | Signed 2025–2026 |
| Amazon | 1.9 GW + 960 MW SMR | Existing nuclear + SMR | Talen Energy, X-energy | Signed 2025 |
| 615 MW + 1,800 MW | Existing restart + SMR | NextEra (Duane Arnold), Elementl | Signed 2025–2026 | |
| Microsoft | 835 MW | Existing nuclear (Three Mile Island restart) | Constellation | Signed 2025 |
| Oracle | 800 MW+ | Nuclear-adjacent campus | Multiple | In development |
Sources: Perkins Coie nuclear deal analysis (January 2026, confirmed via IEA April 2026); EnkiAI deal trackers.
AI Data Center Electricity Demand Forecast to 2030
The IEA’s central projection — roughly doubling total data center electricity from 485 TWh in 2025 to 950 TWh in 2030 — is grounded in what the IEA describes as its most data-intensive forecasting effort to date, using satellite-based tracking of AI factory construction plus bottom-up modeling from server shipments and power draw.
Gartner’s forecast is directionally consistent but extends the near-term trajectory further: 565 TWh in 2026 rising to 702 TWh by 2027 (a 24.1% further increase), with installed power capacity reaching 132 GW in 2026 and a projected 290 GW by 2030.
The EIA’s AEO2026 frames the long-range U.S. outlook as a range: server electricity use reaches 446–818 billion kWh by 2050, with the high end reflecting faster AI server adoption and no efficiency improvements after 2040.
One dimension worth flagging: the IEA notes that bottlenecks — high-bandwidth memory shortages expected to persist through at least end-2027, grid interconnection delays, and permitting timelines — are reducing the probability of more aggressive near-term scenarios. The central projection is not the upper bound; it is the middle of a range that includes significant supply-side constraint.
Global Data Center Electricity Demand: Historical and Forecast
| Year | Consumption (TWh) | AI-Focused (TWh) | % of Global Electricity | Source |
|---|---|---|---|---|
| 2022 | ~460 | ~55–65 | ~1.8% | IEA 2025 |
| 2024 | 415 | ~75–90 | ~1.5% | IEA 2025 |
| 2025 | 447–485 | 95–120 | ~1.7%–1.8% | IEA 2026; Gartner 2026 |
| 2026 (forecast) | 565 | 175 | ~2.1% | Gartner, June 2026 |
| 2027 (forecast) | 702 | 258 | ~2.6% | Gartner, June 2026 |
| 2030 (IEA base case) | 950 | ~285–320 | ~3% | IEA, April 2026 |
| 2030 (Gartner power demand) | n/a | n/a | 290 GW installed capacity | Gartner, June 2026 |
Note: Gartner’s 2025 figure of 447 TWh and IEA’s 2025 figure of 485 TWh reflect different methodological scopes and cut-off dates; both are primary-sourced. IEA 2025 AI-focused figures are Axis Intelligence Research estimates based on the 50% growth rate applied to IEA’s base year figures.
The AI Energy Intensity Ratio (AEIR™): An Axis Intelligence Research Original Metric
Published by Axis Intelligence Research as a baseline reading, July 23, 2026.
No single existing metric captures the structural divergence between AI-specific data center energy growth and overall data center energy growth — the ratio that reveals whether the sector’s energy appetite is driven by AI or by general expansion. Axis Intelligence Research has constructed the AI Energy Intensity Ratio (AEIR™) for this purpose.
Formula:
AEIR™ = AI-focused data center electricity growth rate ÷ Total data center electricity growth rate
Inputs and derivation:
Using IEA primary data for 2025:
- AI-focused data center electricity growth rate: +50% (IEA, Key Questions on Energy and AI, April 2026)
- Total data center electricity growth rate: +17% (IEA, Key Questions on Energy and AI, April 2026)
AEIR™ (2025 reading) = 50% ÷ 17% = 2.94
Interpretation: An AEIR™ of 2.94 means AI-focused facilities consumed electricity at nearly three times the growth rate of the broader data center sector in 2025. A reading of 1.0 would indicate AI facilities grew at the same pace as the overall sector. A reading above 1.0 — particularly one as high as 2.94 — signals that AI is not simply pulling the sector forward at its historical pace; it is restructuring the demand profile of the entire industry.
This is a baseline reading. It is the first computation of this metric. Subsequent readings will be published as IEA and Gartner release updated annual data, enabling trend tracking over time. Axis Intelligence Research will update the AEIR™ annually.
Limitations: The metric depends on IEA’s definition of “AI-focused data centers,” which encompasses facilities with a majority of AI accelerator workloads and does not fully capture AI inference workloads distributed across general-purpose facilities. The denominator (total growth) includes AI-focused growth, meaning the ratio understates the divergence between pure AI and pure traditional workloads.
License: CC BY 4.0. Attribution: “Axis Intelligence Research, AI Energy Intensity Ratio (AEIR™), 2026, axis-intelligence.com/ai-data-center-energy-consumption-statistics/”
AI Data Center Power Density Statistics
The phrase “power density” describes how much electricity a given physical space requires — measured in kilowatts per rack (kW/rack) for individual server units, or megawatts per square meter at the facility level. It is the number that determines whether existing cooling, electrical, and structural infrastructure can support new AI hardware. In almost every case — from a legacy enterprise data center to a recent hyperscale build — the answer is no without significant retrofit.
The IEA’s April 2026 report documents the historical trajectory of AI server power density:
| Period | Change in AI Server Power Density |
|---|---|
| 2020–2025 | ×11 (eleven-fold increase) |
| 2025–2027 (IEA projected) | ×4 (fourfold further increase) |
| Net 2020–2027 (projected) | ×44 (forty-four-fold increase) |
Source: IEA, Key Questions on Energy and AI (April 2026).
To express this in practical terms: in 2020, a typical server rack drew roughly 5–10 kW. By 2027, the IEA projects a rack equivalent to 65 households of peak power demand — placing a single refrigerator-sized unit at roughly 200–300 kW at peak. Current NVIDIA configurations (DGX H100, HGX H200) already place GPU-dense racks in the 60–120 kW range. Next-generation designs are projected above 600 kW.
This density increase has three direct infrastructure consequences:
Cooling: Air cooling, which works by moving air past hot components, becomes physically inadequate above approximately 30–40 kW/rack. High-density AI facilities are migrating to direct liquid cooling (pumping coolant directly to chip surfaces), immersion cooling (submerging hardware in dielectric fluid), and rear-door heat exchangers. Liquid cooling can reduce data center power costs by 20–40% versus equivalent air-cooled setups, while also reducing water usage significantly (Microsoft research published in Nature in April 2025 found a 52% water use reduction with cooling plates and two-phase immersion cooling).
Grid interconnection: A single AI hyperscale campus drawing 500 MW+ requires a dedicated high-voltage transmission interconnection. These interconnections take four years on average to build in the United States, creating the fundamental bottleneck between announced AI data center capacity and operational capacity.
Battery storage: AI training and inference workloads induce large, rapid power swings — server loads varying by more than 50% of rated capacity within a single second. This variability requires onsite battery storage to smooth grid impact. The IEA projects 20–25 GW of battery storage installed in data centers globally by 2030.
AI Data Center CO₂ Emissions Statistics
Energy consumption translates to emissions through the carbon intensity of the electricity grid serving each facility. Global data center CO₂ emissions are on a trajectory that the IEA projects at approximately 350 million tonnes by 2035 — roughly double current levels — but still representing about 2% of global electricity sector emissions at that date.
The emissions math is complicated by several factors:
- Renewable PPAs reduce reported emissions but do not always reduce real-time grid emissions (a renewable credit for solar generated at noon does not offset a data center drawing from the grid at 3 a.m.)
- Nuclear PPAs do reduce real-time grid carbon intensity, which is why hyperscaler nuclear commitments are being taken more seriously by sustainability-focused analysts than renewable credit purchases
- Onsite natural gas — the solution emerging in grid-constrained U.S. markets — increases emissions relative to grid-supplied electricity in states with high renewable penetration
- Hardware efficiency gains — AI energy consumption per task is falling by at least an order of magnitude annually per the IEA, which partially offsets absolute consumption growth, but does not reverse it given the explosion in AI usage volume
The IEA survey of energy companies found lack of digital skills as the single largest barrier to greater AI adoption in the energy sector itself — meaning AI’s potential to reduce emissions through grid optimization, transformer monitoring, and demand forecasting is not yet being fully realized. Less than half of global energy demand is covered by policy frameworks promoting AI adoption in energy.
Methodology: How Axis Intelligence Research Compiled This Dataset
Data collection: All statistics in this article were retrieved from primary sources during the production session of July 23, 2026. Every figure has a corresponding row in the downloadable CSV, including source URL and retrieval date.
Primary sources used:
- Gartner, Forecast: Data Center Power Capacity and Consumption, Worldwide, 2024–2030, 1Q26 (press release, June 10, 2026)
- International Energy Agency (IEA), Key Questions on Energy and AI (April 16, 2026)
- IEA, Energy and AI (2025)
- Lawrence Berkeley National Laboratory, 2024 United States Data Center Energy Usage Report, LBNL-2001637 (December 2024)
- U.S. Energy Information Administration (EIA), Annual Energy Outlook 2026 (AEO2026), May 19, 2026 Today in Energy analysis
- Perkins Coie, nuclear PPA deal summary (January 2026) — for nuclear agreement details
Axis-originated figures:
- The AI Energy Intensity Ratio (AEIR™) is an original Axis Intelligence Research metric. Formula disclosed inline. Inputs are fetched primary data (IEA April 2026). No third-party source publishes this ratio.
- The estimated 2025 U.S. data center electricity range (200–220 TWh) is an Axis Intelligence Research interpolation: IEA projects 2024 U.S. consumption at approximately 183 TWh, with EIA’s commercial-sector growth rate of 3% in 2025 applied. Labeled as an estimate; the denominator uses EIA’s 7%-of-commercial-sector figure as a cross-check.
Limitations:
- Gartner’s 2025 total (447 TWh) and IEA’s 2025 total (485 TWh) differ by ~38 TWh; both are primary-sourced. The difference reflects methodological scope (Gartner segments by server type and includes infrastructure; IEA uses bottom-up server shipment modeling). Axis Intelligence Research presents both without blending them.
- IEA AI-focused consumption data are model estimates, not measured consumption; no global metering system for AI-specific electricity exists.
- LBNL’s 2028 forecast (325–580 TWh) is a wide range by design, reflecting genuine uncertainty in AI hardware adoption trajectories.
- The AEIR™ is a baseline reading; no prior computation exists for trend comparison. We state this explicitly and do not describe this reading relative to a non-existent history.
About This Dataset
Dataset title: AI Data Center Energy Consumption Statistics 2026
Creator/Publisher: Axis Intelligence Research
Coverage period: Historical data 2020–2025; forecasts to 2030 (Gartner) and 2050 (EIA)
Geographic coverage: Global; United States breakout; regional indicators
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/”
Downloadable CSV: Available at the link above; also archived on Hugging Face, Kaggle, and GitHub (Axis Intelligence Research organization)
Citation Block
APA: Axis Intelligence Research. (2026). AI Data Center Energy Consumption Statistics 2026. axis-intelligence.com. https://axis-intelligence.com/ai-data-center-energy-consumption-statistics/
MLA: Axis Intelligence Research. “AI Data Center Energy Consumption Statistics 2026.” Axis Intelligence, 23 July 2026, axis-intelligence.com/ai-data-center-energy-consumption-statistics/.
Chicago: Axis Intelligence Research. “AI Data Center Energy Consumption Statistics 2026.” Axis Intelligence, July 23, 2026. https://axis-intelligence.com/ai-data-center-energy-consumption-statistics/.
Frequently Asked Questions
How much electricity do AI data centers use in 2026?
Globally, all data centers — including AI-focused facilities — are forecast to consume 565 TWh of electricity in 2026, according to Gartner’s June 2026 forecast. Of that total, AI-optimized servers account for 175 TWh (31%), growing 84.2% year-over-year. U.S. data centers consumed 176 TWh in 2023 (the most recent confirmed figure from LBNL); the IEA estimates global data centers reached 485 TWh in 2025, with AI-focused facilities growing 50% in that year alone.
What is the IEA forecast for data center electricity consumption by 2030?
The IEA’s central case projects global data center electricity consumption approximately doubling from 485 TWh in 2025 to roughly 950 TWh by 2030, which would represent approximately 3% of total global electricity demand. AI-focused data center consumption is projected to triple in the same window. The IEA notes that near-term bottlenecks — chip shortages, grid interconnection delays, and permitting backlogs — reduce the probability of more aggressive near-term scenarios.
How fast is AI server power density increasing?
According to the IEA’s April 2026 report Key Questions on Energy and AI, AI server power density increased 11 times between 2020 and 2025 and is projected to increase a further fourfold by 2027. By 2027, a single server rack the size of a household refrigerator could carry peak power demand equivalent to 65 households. Standard colocation racks draw 5–10 kW; current AI GPU racks draw 40–80+ kW, with next-generation designs projected above 600 kW.
What percentage of U.S. electricity do data centers use?
U.S. data centers consumed 176 TWh in 2023, equal to 4.4% of total U.S. electricity, per the Lawrence Berkeley National Laboratory’s 2024 report. The EIA’s Annual Energy Outlook 2026 estimates that data center servers accounted for 7% of commercial-sector electricity in 2025. LBNL projects U.S. data center consumption reaching 325–580 TWh by 2028, representing 6.7%–12% of total U.S. electricity — a range that reflects genuine uncertainty about AI hardware adoption rates.
Why are tech companies signing nuclear power deals for data centers?
The operational profile of AI data centers — near-100% uptime, flat baseload demand, no tolerance for intermittency — is a better match for nuclear generation than for intermittent renewables. Nuclear delivers 24/7 power at a predictable cost without requiring battery storage backup. Meta, Microsoft, Amazon, and Google collectively signed contracts for more than 10 GW of possible new U.S. nuclear capacity in 2025–2026. The IEA’s satellite-based tracking shows the global pipeline of SMR offtake agreements growing from 25 GW (end-2024) to 45 GW (April 2026).
What is PUE, and how does it affect AI data center energy consumption?
Power Usage Effectiveness (PUE) equals total facility power divided by IT equipment power. A PUE of 1.0 is theoretical perfection. Best-in-class hyperscale facilities (Google, Meta) report PUE of 1.08–1.09; the global average is approximately 1.5–1.6. For an AI data center drawing 100 MW of compute load, a PUE of 1.5 means 150 MW of total facility power — 50 MW of overhead. Liquid and immersion cooling can reduce that overhead by 20–40% versus conventional air cooling, making cooling technology a material factor in total energy consumption.
How does AI data center energy growth compare to total electricity demand growth?
The divergence is large. In 2025, global electricity demand grew approximately 3% overall. AI-focused data centers grew 50% in the same period — more than 16 times the global average rate. According to Axis Intelligence Research’s AI Energy Intensity Ratio (AEIR™), AI-focused data center electricity grew at 2.94 times the rate of total data center electricity in 2025. Overall data centers grew 17%, which was itself nearly six times the global electricity demand growth rate.
What happens to data center CO₂ emissions by 2030?
The IEA projects global data center CO₂ emissions roughly doubling by 2035, reaching approximately 350 million tonnes — still about 2% of global electricity sector emissions at that date. The actual trajectory depends heavily on three variables: the carbon intensity of the grid supplying each facility, the pace of nuclear and renewable PPA signings, and whether the efficiency gains in AI hardware outpace the raw volume growth in AI usage. Hyperscaler nuclear PPAs — covering more than 10 GW of capacity — represent the most consequential variable for near-term emissions trajectories.
Related Axis Intelligence Research: AI Data Center Statistics · Data Center Grid Impact Statistics · Nuclear Energy Statistics 2026 · AI Data Center Water Usage Statistics · Data Center Construction Cost Statistics · AI Data Center Forecast 2030