AI Energy Consumption Statistics: How Much Electricity AI Uses

AI electricity demand is rising quickly, but public data rarely isolates every AI workload. Most reliable figures cover AI-focused facilities or data centers as a sector.
Per-query estimates change faster. Models, chips, output length, and reasoning effort can alter each request's demand.
AI energy consumption helped drive a 17% rise in global data-center electricity use during 2025. The International Energy Agency measured 485 terawatt-hours in 2025. It projects about 950 TWh in 2030. AI-focused facilities could triple their use during that period.
Key AI Electricity Statistics
Diagram
Key AI Electricity Statistics
- Global data-center electricity use
- Projected global data-center electricity use
- Global data-center demand growth
- AI-focused data-center demand growth
- Projected AI-focused data-center growth
- US data-center reference case
- US data-center uncertainty range
- US share of electricity use
| Measure | Figure | Scope |
|---|---|---|
| Global data-center electricity use | 485 TWh | 2025 estimate |
| Projected global data-center electricity use | 950 TWh | 2030 central projection |
| Global data-center demand growth | 17% | 2025 annual increase |
| AI-focused data-center demand growth | About 50% | 2025 annual increase |
| Projected AI-focused data-center growth | Threefold | 2025 to 2030 |
| US data-center reference case | 649 TWh | 2030 LBNL model |
| US data-center uncertainty range | 521 to 843 TWh | 2030 scenarios |
| US share of electricity use | 11.8% reference case | 2030 projection |
These totals include infrastructure serving many workloads. They do not provide one energy number for every AI prompt.
How Much Electricity Does AI Use Today?
No public dataset measures every training and inference task worldwide. The best sector estimate comes from data-center demand.
The International Energy Agency's 2026 update estimated 485 TWh of global data-center electricity during 2025.
That sector total increased by 17% in one year. Electricity use at AI-focused facilities rose about 50%.
AI was the fastest-growing component, a trend our AI statistics on investment and production use also track. Data centers also serve cloud storage, streaming, enterprise software, and other digital services.
How Fast Could AI Electricity Demand Grow?
The IEA expects global data-center electricity use to reach roughly 950 TWh in 2030. That would represent about 3% of global electricity demand.
AI-focused data centers could triple their electricity use from 2025 to 2030. The projection assumes continued deployment despite grid and supply-chain constraints.
The agency also sees upside risk after 2030. More energy-intensive agents, reasoning systems, and video generation could increase demand.
AI energy consumption forecasts are scenarios, rather than observed use. Grid connections, transformers, chips, financing, and regulation can change the result.
Why Is Per-query Energy Difficult to Measure?
The energy behind one AI task depends on several variables:
- model architecture and size;
- input and output length;
- hardware generation;
- server utilization and batching;
- reasoning or agentic steps;
- data-center cooling and power overhead; and
- location and electricity supply.
Providers also update models and serving systems frequently. An estimate from one version may become stale within months.
The AI token calculator can estimate model charges. Token cost does not reveal electricity consumption.
Does Training or Inference Use More Electricity?
Training concentrates heavy computing into model development. Inference spreads demand across every user request after release.
For a lightly used model, training may dominate. For a product with hundreds of millions of users, inference can become much larger.
Reasoning models can perform many internal steps before producing an answer. Agents may call several models and tools for one task.
This makes "one query" an unstable unit. Water estimates face the same problem, as our AI water usage research shows. Two prompts can trigger different amounts of computation on the same service.
Is AI Becoming More Energy Efficient?
Yes, energy per task can fall as chips and software improve. The IEA describes the pace of efficiency improvement as unusually fast.
Total demand can still rise. Lower cost encourages more users, longer outputs, and new applications.
This effect resembles rebound in other technologies. Efficiency reduces the unit cost, while volume grows faster than savings.
The result explains why falling energy per request can coexist with rising sector demand.
How Much Electricity Could US Data Centers Use?
The 2026 Lawrence Berkeley National Laboratory update modeled 649 TWh in 2030 as its reference case.
That amount equals 11.8% of projected US electricity use. Its compounded uncertainty range runs from 521 to 843 TWh.
The range represents 9.5% to 15.3% of US electricity consumption. The model varies equipment shipments, AI chip life, utilization, and idle power.
Earlier LBNL work measured 176 TWh in 2023. US data centers then represented about 4.4% of national electricity consumption.
These projections cover all data-center work. AI servers drive much of the increase, but the total is not AI-only.
What Physical Limits Could Slow Growth?
Electricity projects need more than generation. Data centers also need substations, transformers, transmission, cooling, permits, and suitable land.
The IEA reports tightening supplies for transformers, gas turbines, advanced chips, and other equipment. Grid queues can delay planned sites.
Around 20% of planned projects could face delays without faster grid integration. This constraint appears in the agency's 2025 analysis.
Capital requirements are also large. Five major technology companies spent more than $400 billion during 2025, according to the IEA.
Does Renewable Procurement Make AI Carbon-free?
Renewable purchases can reduce annual emissions. They do not mean every facility receives carbon-free power during every hour.
Location and timing affect the grid mix. A data center can buy renewable energy while drawing from a mixed local grid.
Energy use and emissions are separate measures. A facility may consume more electricity while reducing emissions per kWh.
Readers should inspect both totals. Corporate claims should state whether they use annual matching or hourly supply.
How Do PUE and Utilization Affect AI Energy Use?
Power usage effectiveness compares total facility energy with IT equipment energy. A PUE of 1.4 means 40% extra facility energy per unit of IT energy.
LBNL reported an industry average near 1.4 during 2023. The average had improved from around 1.6 in 2014.
Better PUE reduces overhead. It does not reduce computing demand inside the servers.
Server utilization also changes efficiency. An idle server still consumes power. Higher utilization can spread fixed demand across more completed work.
Stack Row's cloud infrastructure coverage examines how architecture choices affect capacity, cost, and operations.
What Should Buyers Ask AI Providers?
Procurement teams need comparable disclosures. Useful questions include:
- What period and system boundary does the figure cover?
- Does it include training, inference, or both?
- Is electricity measured or modeled?
- Which model and hardware version were tested?
- Does the emissions claim use annual or hourly matching?
Provider answers may remain incomplete. Consistent questions still improve contract review and future reporting.
Our article on how much electricity data centers use separates sector totals from AI-specific growth.
Methodology
This article uses evidence published through 8 November 2026. It prioritizes IEA, US Department of Energy, LBNL, and first-party disclosures.
We excluded universal per-prompt claims without current system assumptions. Forecasts remain labeled and separate from observed use.
FAQs
How Much Electricity Does One AI Query Use?
There is no universal figure. Model size, task length, hardware, utilization, and internal reasoning steps change the result.
How Much Power Do Data Centers Use Globally?
The IEA estimated 485 TWh during 2025. Its central projection reaches about 950 TWh in 2030.
Will AI-focused Data-center Demand Keep Growing?
The IEA projects a threefold increase from 2025 to 2030. Grid, equipment, financing, and permitting constraints could change the pace.
Does Better Hardware Reduce Total AI Electricity Use?
Better hardware can reduce energy per task. Total use may still rise when adoption and task complexity grow faster.
Is AI the Only Cause of Data-center Growth?
No. Cloud services, streaming, storage, and enterprise computing also contribute. AI is the fastest-growing driver in recent projections.
