Global data center electricity use is on track to roughly double to about 945 TWh by 2030, near 3% of all electricity, up from around 415 TWh in 2024 (IEA, Energy and AI 2025). AI is the primary driver: on the same IEA accounting, electricity use at AI-focused data centers surged 50% in 2025 alone while overall data center demand grew 17% that year. Yet the per-query picture cuts the other way. Google reports the energy behind a median Gemini prompt fell 33x in twelve months (Google, 2025), and Epoch AI pegs a typical ChatGPT query at about 0.3 watt-hours, ten times below older estimates (Epoch AI, 2025). This analysis consolidates data from the IEA, EPRI, Lawrence Berkeley National Laboratory, Google, Epoch AI, Stanford HAI, McKinsey, and 10 other primary sources to separate the headline projections from the measured per-unit reality.
TL;DR
- Global data center electricity use is projected to roughly double from about 415 TWh in 2024 to around 945 TWh by 2030, with AI-focused sites up 50% in 2025 alone (IEA, Energy and AI 2025).
- Gartner puts global data center consumption at 565 TWh in 2026, up 26% from 447 TWh in 2025 (Gartner, 2026).
- US data centers used about 4.4% of national electricity in 2023 and could reach 6.7% to 12% by 2028 (LBNL, 2024).
- A typical GPT-4o query uses roughly 0.3 watt-hours, about ten times below older estimates (Epoch AI, 2025).
- Training Meta’s Llama 3.1 405B emitted about 8,930 tons of CO2 in 2024 (Stanford HAI, AI Index 2025).
- Microsoft’s emissions rose 25% in 2025 to 20 million metric tons of CO2 equivalent (Microsoft FY2025 disclosure, via Fortune 2026).
- PJM capacity prices climbed from 28.92 to 329.17 dollars per megawatt-day, with data centers driving 63% of the increase (PJM, 2025; IEEFA, 2025).
- The five largest hyperscalers are projected to spend around 725 billion dollars on AI infrastructure in 2026 (analyst estimates, 2026).
- McKinsey projects up to 6.7 trillion dollars in data center capital investment by 2030 (McKinsey, 2025).
1. Global data center and AI electricity demand
The macro story is a doubling in roughly six years, and AI is the accelerant. Data centers ran on about 415 TWh in 2024, but the growth rate matters more than the level: AI-focused data center electricity jumped 50% in a single year (2025) while the broader category grew 17% (IEA, Energy and AI 2025). By 2030 the AI share of data center power is set to rise from today’s 5-15% band to 35-50%, making it the single largest driver of the increase. Estimates diverge by scope, so read the source basis, not just the headline.
| Metric | Value | Source |
|---|---|---|
| Global data center electricity, 2024 | ~415 TWh (~1.5% of global) | IEA, Energy and AI 2025 |
| Global data center electricity, 2030 (Base Case) | ~945 TWh (~3% of global) | IEA, Energy and AI 2025 |
| AI-focused data center electricity growth, 2025 | +50% | IEA, Energy and AI 2025 |
| Global data center electricity, 2026 (Gartner) | 565 TWh (+26% YoY) | Gartner, 2026 |
| AI-optimized server electricity, 2026 | 175 TWh (+84% YoY) | Gartner, 2026 |
| Global power demand growth by 2030 vs 2023 | +165% | Goldman Sachs, 2025 |
| AI-optimized server share of data center power, 2026 | 31% | Gartner, 2026 |
Scope note: the IEA sees the AI share of data center power rising from 5-15% today toward 35-50% by 2030. Gartner’s 2025 baseline (447 TWh) sits below other 2025 estimates near 485 TWh because methodologies use different facility definitions and boundaries; treat cross-source TWh comparisons with care. The chips inside these racks are the demand engine; see our AI chips statistics for 2026 for the silicon side. Full report: IEA, Energy demand from AI.
2. The US grid under load
The United States is where the demand curve meets a physical grid, and forecasts keep getting revised upward. Berkeley Lab clocked data centers at 4.4% of national electricity in 2023; the newer EPRI scenarios stretch to 9-17% of US electricity by 2030. EPRI’s peak-load estimate now runs as high as 94 GW by 2030, up from 21-22 GW in 2024, and its new numbers are 60% above its own 2024 estimate (EPRI, Powering Intelligence 2026). Grid Strategies frames the national picture: 166 GW of five-year load growth, roughly 90 GW of it data centers.
| Metric | Value | Source |
|---|---|---|
| US data center share of electricity, 2023 | 4.4% (176 TWh) | LBNL, 2024 |
| US data center share by 2028 | 6.7% to 12% (325-580 TWh) | LBNL, 2024 |
| US data center share by 2030 | 9% to 17% | EPRI, Powering Intelligence 2026 |
| US data center peak load, 2024 | 21-22 GW | EPRI, Powering Intelligence 2026 |
| US data center peak load by 2030 | 45 / 71 / 94 GW (Low/Med/High) | EPRI, Powering Intelligence 2026 |
| National 5-year peak load growth | 166 GW (~90 GW data centers) | Grid Strategies, 2025 |
Outlier note: Grid Strategies cautions that utilities may be overstating data center demand by as much as 40%, because speculative and duplicate interconnection requests inflate the queue (Grid Strategies, National Load Growth Report 2025). Reports: LBNL 2024 Data Center Energy Usage Report and EPRI, Powering Intelligence 2026.
3. What one AI query actually costs
Per-query numbers have collapsed as vendors published real telemetry, which reframes the “AI is boiling the oceans” narrative around a single prompt. A median Gemini text prompt draws 0.24 watt-hours, 0.03 grams of CO2e, and 0.26 milliliters of water, roughly five drops (Google, 2025). Epoch AI independently lands at about 0.3 Wh for a typical GPT-4o query, and OpenAI’s own figure is 0.34 Wh. The catch: long-context prompts are far heavier, and the aggregate still adds up across billions of daily queries.
| Metric | Value | Source |
|---|---|---|
| Typical GPT-4o query energy | ~0.3 Wh (vs 3 Wh old estimate) | Epoch AI, 2025 |
| Median Gemini text prompt energy | 0.24 Wh | Google, 2025 |
| Average ChatGPT query energy | 0.34 Wh | Sam Altman, The Gentle Singularity 2025 |
| Long query (100k input tokens) | ~40 Wh | Epoch AI, 2025 |
| Median Gemini prompt carbon | 0.03 gCO2e | Google, 2025 |
| Gemini per-prompt energy drop, 12 months to May 2025 | 33x lower | Google, 2025 |
| 1 billion ChatGPT messages/day | ~12.5 MW continuous | Epoch AI, 2025 |
Context: the per-query fall reflects more efficient models and hardware, not lower demand. For the dollar side of the same trend, see our AI inference cost statistics for 2026. Primary methodology: Google Cloud, Measuring the environmental impact of AI inference and Epoch AI, How much energy does ChatGPT use?.
4. Training frontier models
Training is a one-time spike per model, but the spikes are getting steeper. Stanford’s AI Index puts the training emissions of Llama 3.1 405B at about 8,930 tons of CO2, up from 588 tons for GPT-3 in 2020 and effectively nothing for AlexNet in 2012 (Stanford HAI, AI Index 2025). The report’s structural finding is the compounding: training compute for notable models doubles roughly every five months, and the power to train frontier systems doubles about annually.
| Metric | Value | Source |
|---|---|---|
| Llama 3.1 405B training emissions (2024) | ~8,930 tons CO2 | Stanford HAI, AI Index 2025 |
| GPT-4 training emissions (2023) | ~5,184 tons CO2 | Stanford HAI, AI Index 2025 |
| GPT-3 training emissions (2020) | ~588 tons CO2 | Stanford HAI, AI Index 2025 |
| AlexNet training emissions (2012) | ~0.01 tons CO2 | Stanford HAI, AI Index 2025 |
| Training compute doubling time | ~every 5 months | Stanford HAI, AI Index 2025 |
| GPT-4 training energy (leaked-spec estimate) | ~50-62 GWh | Independent estimates via TDS, 2023 |
Flag: the GPT-4 training-energy figure is an external estimate derived from leaked hardware specs (25,000 A100 GPUs over 90-100 days), not a vendor disclosure; treat it as indicative. Training is also a small share of lifetime footprint relative to inference at scale. For the broader model-building picture, see our generative AI statistics for 2026. Source: Stanford HAI, 2025 AI Index Report, Chapter 1.
5. The water footprint
Water is the second resource story, split between direct cooling and the water embedded in electricity generation. Texas data centers are projected to use as much as 399 billion gallons of water in 2030, up from 49 billion in 2025 (HARC and University of Houston, 2025). Vendors are pushing efficiency: Microsoft reported a water-use effectiveness of 0.30 L/kWh in FY2025, a 39% improvement over 2021, and its zero-water-evaporation cooling avoids more than 125 million liters per data center each year.
| Metric | Value | Source |
|---|---|---|
| Median Gemini prompt water use | 0.26 mL | Google, 2025 |
| Microsoft water-use effectiveness, FY2025 | 0.30 L/kWh (-39% vs 2021) | Microsoft, 2025 |
| Microsoft zero-water cooling saving | >125 million liters/data center/year | Microsoft, 2025 |
| US data center direct water use, 2023 | ~17 billion gallons | LBNL, 2024 |
| US direct water use projected by 2028 | 38-73 billion gallons | LBNL, 2024 |
| Texas data center water use, 2025 to 2030 | 49 billion to 399 billion gallons | HARC and University of Houston, 2025 |
| Large single data center | up to 5 million gallons/day | EESI, 2025 |
Context: US data centers also drove an estimated 211 billion gallons of indirect water use through electricity generation in 2023, which typically exceeds on-site cooling water (LBNL, 2024). Background: EESI, Data Centers and Water Consumption.
6. Carbon emissions climbing
Efficiency per query is improving, but corporate footprints are rising because absolute compute is exploding. Microsoft’s emissions rose 25% in 2025 to 20 million metric tons of CO2 equivalent, up from 16 million, driven by data center construction and a pullback on some renewable-energy-credit purchases (Microsoft FY2025 disclosure, via Fortune 2026). Amazon and Google reported similar directional increases. At the sector level, the IEA still frames data centers as a small but rising slice of global carbon.
| Metric | Value | Source |
|---|---|---|
| Microsoft CO2e emissions, 2025 | 20 million tons (+25% YoY) | Microsoft FY2025 disclosure, via Fortune 2026 |
| Microsoft CO2e emissions, 2024 | 16 million tons | Microsoft FY2025 disclosure, via Fortune 2026 |
| Amazon carbon footprint change (latest year) | +16% | Carbon Brief, 2026 |
| Google greenhouse gas emissions change (latest year) | +18% | Carbon Brief, 2026 |
| Data center share of global CO2, 2024 to 2030 | ~0.5% rising to ~1% (1.4% high case) | IEA, Energy and AI 2025 |
| AI systems carbon footprint, 2025 | 32.6 to 79.7 million tons CO2 | Carbon Brief, 2026 (published research) |
Outlier note: the AI-systems carbon range of 32.6 to 79.7 million tons for 2025 is wide because workload attribution is genuinely uncertain, and several firms booked emissions increases partly from pausing some renewable-energy-credit purchases rather than from raw consumption alone (Carbon Brief, 2026). Context and charts: Carbon Brief, AI: Five charts on data-centre energy use and emissions and Fortune, Microsoft carbon emissions 2025.
7. Powering it: prices, capex, and nuclear
Someone pays for the new load, and increasingly it shows up on the grid and on utility bills. In the PJM region, capacity prices rocketed from 28.92 dollars per megawatt-day in 2024/25 to 329.17 dollars in 2026/27, with data centers responsible for 63% of the 2025/2026 jump (PJM, 2025; IEEFA, 2025). To meet demand, hyperscalers are committing historic capital and turning to nuclear power, while Gartner warns grid supply falls short once consumption passes 1,200 TWh around 2030.
| Metric | Value | Source |
|---|---|---|
| PJM capacity price, 2024/25 to 2026/27 | 28.92 to 329.17 dollars/MW-day | PJM, 2025 |
| Data center share of December PJM capacity costs | 40% (6.5 of 16.4 billion dollars) | PJM market monitor, via Utility Dive 2026 |
| Added PJM ratepayer cost from data centers | ~9.3 billion dollars/year | IEEFA, 2025 |
| Big-5 hyperscaler AI infrastructure spend, 2026 | ~725 billion dollars (+77% YoY) | Analyst estimates, 2026 |
| Global data center capex by 2030 | up to 6.7 trillion dollars (5.2T for AI) | McKinsey, 2025 |
| Data center projects blocked, early 2026 | 75+ projects worth 130 billion dollars | Gartner, 2026 |
| Average PUE at largest data centers, 2025 | 1.54 (flat for 6 years) | Uptime Institute, 2025 |
Context: the same buildout underpins the cloud-capacity boom; see our cloud computing statistics for 2026. Sources: IEEFA, PJM capacity prices, McKinsey, The cost of compute, and Utility Dive, Data centers were 40% of PJM capacity costs.
Summary: AI Energy Consumption by the Numbers
| Metric | Value | Source |
|---|---|---|
| Global data center electricity by 2030 | ~945 TWh (~3% of global) | IEA, Energy and AI 2025 |
| Global data center electricity, 2026 | 565 TWh (+26% YoY) | Gartner, 2026 |
| Grid supply shortfall threshold | ~1,200 TWh by 2030 | Gartner, 2026 |
| AI-optimized server electricity, 2026 | 175 TWh (+84% YoY) | Gartner, 2026 |
| Global power demand growth by 2030 vs 2023 | +165% | Goldman Sachs, 2025 |
| US data center share of electricity, 2023 | 4.4% | LBNL, 2024 |
| US data center share by 2030 | 9% to 17% | EPRI, Powering Intelligence 2026 |
| US data center peak load by 2030 (High) | 94 GW | EPRI, Powering Intelligence 2026 |
| Typical GPT-4o query energy | ~0.3 Wh | Epoch AI, 2025 |
| Median Gemini text prompt energy | 0.24 Wh | Google, 2025 |
| Llama 3.1 405B training emissions | ~8,930 tons CO2 | Stanford HAI, AI Index 2025 |
| US data center direct water use, 2023 | ~17 billion gallons | LBNL, 2024 |
| Texas data center water use by 2030 | up to 399 billion gallons | HARC and University of Houston, 2025 |
| Microsoft emissions, 2025 | 20 million tons (+25% YoY) | Microsoft FY2025, via Fortune 2026 |
| Added PJM ratepayer cost from data centers | ~9.3 billion dollars/year | IEEFA, 2025 |
| PJM capacity price, 2026/27 | 329.17 dollars/MW-day | PJM, 2025 |
| Global data center capex by 2030 | up to 6.7 trillion dollars | McKinsey, 2025 |
| Big-5 hyperscaler AI spend, 2026 | ~725 billion dollars | Analyst estimates, 2026 |
| Average data center PUE, 2025 | 1.54 | Uptime Institute, 2025 |
Methodology and Sources
Data was gathered by aggregating figures from primary reports, corporate disclosures, grid operator filings, and named research trackers published mainly in 2025 and the first half of 2026, with each statistic traced to the originating organization rather than to secondary coverage where possible.
Sources cited:
- International Energy Agency (IEA), Energy and AI 2025 - Energy demand from AI
- Lawrence Berkeley National Laboratory (LBNL), 2024 United States Data Center Energy Usage Report - publication
- Electric Power Research Institute (EPRI), Powering Intelligence 2026 - executive summary
- Grid Strategies, National Load Growth Report 2025 - PDF
- Gartner, data center electricity forecast press release, June 2026 - release
- Goldman Sachs Research, AI data center power demand, 2025 - article
- Epoch AI, How much energy does ChatGPT use?, 2025 - analysis
- Google Cloud, Measuring the environmental impact of AI inference, 2025 - blog
- OpenAI / Sam Altman, The Gentle Singularity, 2025 - via DCD
- Stanford HAI, 2025 AI Index Report - Chapter 1
- Microsoft FY2025 sustainability disclosure - via Fortune
- Carbon Brief, AI: Five charts on data-centre energy and emissions, 2026 - article
- McKinsey, The cost of compute, 2025 - report
- IEEFA, PJM capacity prices analysis, 2025 - resource
- PJM Interconnection independent market monitor - via Utility Dive
- HARC and University of Houston, Texas data center water study, 2025
- Uptime Institute, 2025 Global Data Center Survey (PUE)
- Environmental and Energy Study Institute (EESI), Data Centers and Water Consumption, 2025 - article
Data watch: The IEA typically refreshes its energy outlooks annually, so an updated Energy and AI edition is expected within the next year; Gartner reissues its data center electricity forecast on a recurring basis; EPRI updates its Powering Intelligence scenarios roughly annually; Grid Strategies publishes a National Load Growth Report each year; Stanford HAI releases a new AI Index annually; Google, Microsoft, Amazon, and Alphabet publish fresh environmental and emissions disclosures in each new fiscal cycle; and PJM runs new capacity auctions on a set delivery-year schedule.
Last updated: July 10, 2026.
We review and update this page quarterly as new data is published.