Few AI questions generate as many conflicting numbers as its energy use. Headlines range from “AI will break the grid” to “a chatbot query uses less energy than a light bulb for a minute”. The truth depends on what you measure. This guide summarizes the most credible evidence.

The big picture: data centers

AI runs in data centers, which also host cloud services, streaming, storage and much more. The International Energy Agency (IEA), in its 2025 report Energy and AI, estimated that data centers consumed about 415 terawatt-hours (TWh) of electricity in 2024 — roughly 1.5% of global electricity consumption. Its base-case projection has data-center consumption more than doubling to around 945 TWh by 2030, with AI the most important driver of growth.

For context, that 2030 figure is comparable to the current electricity consumption of a large industrialized country. Yet even then, data centers would remain a modest share of total global electricity demand, which is also growing because of electric vehicles, air conditioning and industry.

Why it matters locally

Global averages hide local strain. Data centers cluster in specific regions where land, connectivity and power are available. In those places, they can account for a large share of electricity demand, requiring new generation, transmission lines and grid upgrades — and raising questions about who pays.

Training vs. inference

  • Training a frontier model uses a large amount of energy in a concentrated period, on thousands of accelerators running for weeks or months.
  • Inference — answering user requests — uses far less energy per request, but happens billions of times. As AI becomes embedded in search, apps and work tools, inference is expected to account for a growing share of AI’s energy use.

Per-query estimates vary widely

Estimates for a single chatbot response differ by orders of magnitude because they depend on:

  • the model size and efficiency;
  • the length of input and output;
  • whether reasoning or long “thinking” is used;
  • the hardware and how fully it is utilized;
  • data-center efficiency, including cooling.

Research such as Power Hungry Processing (2023) by Sasha Luccioni and colleagues found that generative tasks, especially image generation, use far more energy than classification tasks, and that general-purpose models used for narrow tasks are less efficient than specialized ones.

Factors pushing in both directions

Increasing demand: more users, more AI features, larger models, longer reasoning, video generation.

Increasing efficiency: better chips per watt, smaller distilled models, quantization, improved cooling and better scheduling of workloads.

Whether efficiency gains outpace demand growth remains one of the key uncertainties.

Beyond electricity

Data centers also use water for cooling in many designs, and building chips and facilities has its own environmental footprint. The carbon impact depends heavily on the local electricity mix.

Common misconceptions

  • Quoting one per-query number as universal. Energy per request varies enormously with model, task and hardware.
  • Confusing data centers with AI. Data centers serve many workloads; AI is the fastest-growing one but not the only one.
  • Ignoring location. Local grid strain and carbon intensity matter more than global averages for real impact.

Putting individual queries in context

Per-query energy estimates are small compared with everyday activities, but billions of daily queries add up. The bigger drivers are the growth of total usage, longer reasoning outputs and training ever-larger models.

Related: AI data center capex.

Key takeaways

  • The IEA estimated data centers used about 415 TWh in 2024, roughly 1.5% of global electricity.
  • Its base case sees data-center consumption reaching around 945 TWh by 2030, driven mainly by AI.
  • Impacts are concentrated in regions where data centers cluster.
  • Inference is growing as a share of AI energy use as AI spreads into everyday products.
  • Per-query estimates vary by orders of magnitude depending on model, task and hardware.

Frequently asked questions

Does asking a chatbot a question use a lot of energy?

A single typical text query uses a small amount of energy, but estimates vary widely. Long reasoning, image and video generation use considerably more.

Will AI cause electricity shortages?

Globally, data centers remain a modest share of demand. Locally, rapid growth can strain grids and requires new generation and transmission.

Is AI getting more energy-efficient?

Yes, per unit of work: chips and models are becoming more efficient. Total consumption still grows because usage grows faster.