How much energy do data centers and artificial intelligence use?

2026-07-20 · Source: Our World in Data · Field: Energy & Utilities — Energy Efficiency & Conservation, Utilities & Infrastructure, Energy Markets & Policy · Depth: Intermediate, long

Summary

Global artificial intelligence (AI) adoption is rapidly increasing, prompting concerns about its energy demand's environmental, local community, and technological bottleneck impacts. In 2025, data centers consumed an estimated 485 terawatt-hours (TWh), representing 1.5% of global electricity generation, with AI-focused facilities accounting for 155 TWh, or 0.5%. Projections indicate data centers could reach 3% of global electricity by 2030, with AI's share growing significantly. Estimates vary, with S&P Global reporting 790 TWh for 2025, partly due to including cryptocurrency mining, which the International Energy Agency (IEA) excludes. While individual large language model (LLM) queries consume small amounts, typically 0.24-0.34 watt-hours, and up to 50 Wh for complex agentic tasks, the overall demand is geographically concentrated, with data centers using 5% of US electricity and over 20% in regions like Ireland. Future energy demand remains highly uncertain, influenced by user growth and hardware efficiency, but the carbon impact is primarily determined by the electricity's generation source.

Key takeaway

For policy makers addressing AI's environmental impact, recognize that while global AI energy consumption is currently a small percentage, its rapid growth and geographic concentration pose significant local grid challenges. Your focus should shift from solely total consumption to the carbon intensity of the electricity supply. Prioritize investments in clean energy infrastructure for data center regions to mitigate emissions and ensure grid stability, rather than restricting AI adoption based on overall energy use.

Key insights

AI's energy footprint is growing and geographically concentrated, with carbon impact tied to grid cleanliness.

Principles

• Inference dominates AI energy demand.
• Data center energy estimates vary by source.
• Local grid strain from concentrated demand.

Method

Data center power demand is estimated either directly from IT equipment and cooling (IEA) or modeled from installed capacity (S&P Global).

In practice

• Individual AI queries use minimal energy.
• Prioritize clean energy sources for data centers.
• Scrutinize data center energy estimate methodologies.

Topics

• AI Energy Consumption
• Data Center Infrastructure
• Electricity Grid Strain
• Carbon Emissions
• Large Language Models
• Energy Demand Forecasting

Best for: CTO, VP of Engineering/Data, Director of AI/ML, AI Ethicist, Policy Maker, Consultant

Related on AIssential

• Officials hugely underestimated impact of AI datacentres on UK carbon emissions - The Guardian — UK government vastly underestimated AI datacentre carbon emissions, raising climate concerns.
• The AI Boom Is Taking Over the U.S. Power Grid, and This Map Shows Where It’s Happening — The AI boom is rapidly expanding data center infrastructure, placing immense strain on the U.S. power grid and energy resources.
• Assessing the Carbon Emissions and Energy Consumption of U.S. Hyperscale Data Centers — US hyperscale data centers disproportionately rely on carbon-intensive electricity, driving significant emissions and energy demand.
• The fight against AI data centers is just beginning — Community resistance to AI data centers is escalating globally due to significant environmental and economic concerns.
• Google’s AI buildout drove 37% increase in electricity use in 2025 — Google's AI growth significantly boosts electricity use, challenging its carbon-free energy goals despite clean energy purchases.
• OpenAI’s AI spending spree has ballooned to $750B — Scaling AI models requires unprecedented infrastructure investment and raises significant environmental and regulatory challenges.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Our World in Data.