How much energy do data centers and artificial intelligence use?

Artificial intelligence (AI) has been adopted faster than nearly any other technology in history. This rapid growth raises questions about its energy use, which stems from the electricity needed to train AI models and run them for users (a process called inference). Training AI models involves feeding vast amounts of data into computers to teach them patterns, while inference is when the trained model makes predictions or generates responses based on new inputs. Most of AI's energy consumption comes from inference, not training, though exact figures are rarely published by tech companies. For example, in 2025, AI-focused data centers likely used about 0.5% of the world's electricity, while non-AI data centers used twice that amount. The International Energy Agency (IEA) projects that by 2030, AI and non-AI data centers will each account for about 1.5% of global electricity use, totaling 3%.

Data centers, which host servers for digital services like email, streaming, and AI, consumed about 485 terawatt-hours (TWh) of electricity in 2025, equivalent to the annual electricity generation of Germany. This represents 1.5% of the world's total electricity. AI-focused data centers accounted for about one-third of this, or 155 TWh (0.5% of global electricity), while non-AI centers used the remaining 330 TWh (1%). Electricity makes up roughly one-fifth of the world's primary energy, so AI's share of total primary energy in 2025 was likely around 0.1%. The IEA projects data center electricity use to nearly double by 2030, reaching 945 TWh (3% of global electricity), with AI-focused centers growing faster than non-AI ones.

While data centers use only 1.5% of the world's electricity, this demand is highly concentrated in specific regions and countries. For example, 5% of electricity in the United States is used by data centers, with some states like Virginia consuming over 25%. In Europe, data centers account for 1.6% of electricity use, but Ireland's share exceeds 20%. This concentration can strain local power grids, even if the global impact remains small. The rapid growth of AI could exacerbate this issue, as AI-focused data centers are expected to drive most of the future demand increase.

The energy used by an individual AI query, such as asking a chatbot a question, is relatively small. Estimates from companies like Google and OpenAI suggest a typical text-based query consumes around 0.24 to 0.34 watt-hours (Wh) of electricity. For context, this is roughly the energy a microwave uses in less than a second or a television for ten seconds. More complex queries, such as those involving long inputs (e.g., 7,500 words) or AI agents with reasoning, can consume up to 2.5 Wh or 50 Wh, respectively. These figures are still small compared to daily electricity use in high-income countries, where the average person consumes about 17,000 Wh per day in the EU and 34,000 Wh in the US.

Predicting future energy demand for AI is highly uncertain due to factors like user growth, technological efficiency, and infrastructure development. The International Energy Agency (IEA) has outlined four scenarios for future demand, ranging from high user growth to rapid efficiency gains. For instance, one scenario projects much higher demand growth, while another assumes faster improvements in hardware efficiency. The IEA's base-case projection suggests data centers could use 3% of global electricity by 2030, but this could vary significantly. Regardless of the exact figure, the article emphasizes that how electricity is generated (e.g., renewable vs. fossil fuels) may be more critical for climate impact than the total amount used.