The Environmental Cost of AI: Training vs Inference Energy, Data Center Electricity, Water and What the Published Estimates Show
What is actually known about AI's environmental footprint: published training estimates for GPT-3 and BLOOM, Google's per-prompt inference figures, IEA projections for data center electricity, company emissions trends, efficiency gains, water use, and why standardised disclosure is still missing.
AI’s environmental footprint is real and growing, but most of the numbers circulating online are estimates, and many are outdated. The best evidence suggests three things. First, a single training run is no longer the main concern; aggregate inference across billions of queries matters more. Second, data center electricity demand is rising fast: the International Energy Agency (IEA) estimates data centers used about 415 TWh in 2024, around 1.5% of global electricity, and projects that demand will more than double to around 945 TWh by 2030, slightly more than Japan’s entire electricity consumption today, with AI a major driver (IEA report, IEA press release). Third, per-query efficiency is improving quickly, but total demand is outpacing it. The biggest gap is transparency: companies disclose little, and their methods differ.
Published estimates of training costs
Only a handful of peer-reviewed or detailed estimates exist, mostly for models from 2020 to 2022. Developers of current frontier models generally do not publish training energy.
| Study | Model or scope | Estimate | Notes |
|---|---|---|---|
| Strubell et al., 2019 | NLP models, including neural architecture search | The architecture search case was compared to the lifetime emissions of about five cars | Widely cited, but a worst-case search scenario rather than typical training |
| Patterson et al., 2021 | GPT-3 | About 1,287 MWh and 552 tonnes CO2e | Authors stress hardware, data center efficiency and grid mix change emissions by orders of magnitude |
| Luccioni et al., 2022 | BLOOM (176B parameters) | About 25 tonnes CO2e from dynamic power; roughly double including manufacturing and idle power | Trained largely on low-carbon French electricity |
The “five cars” figure from 2019 shaped public perception of AI as an environmental villain, but it described an unusually expensive search procedure. Later work showed the same amount of computation can produce very different emissions depending on where and how it runs.
Inference: small per query, large in total
Training happens once per model; inference happens every time someone uses it. In August 2025, Google reported that a median Gemini Apps text prompt used about 0.24 Wh of energy and about 0.26 mL of water, which it compared to watching television for less than nine seconds (Google Cloud blog). Google also said the energy and total carbon footprint of the median prompt fell by 33 times and 44 times respectively over a recent 12-month period.
These are company-reported figures for a median text prompt. Image and video generation, long reasoning chains and agentic tasks can use far more energy per request, and totals depend on how many requests are made. Small per-query numbers can still add up to large aggregate demand.
Data centers, grids and water
| Indicator | Figure | Source |
|---|---|---|
| Global data center electricity use, 2024 | About 415 TWh (around 1.5% of global electricity) | IEA |
| Projected data center use, 2030 | About 945 TWh in the IEA’s base case | IEA |
| Google fleet power usage effectiveness (PUE), 2025 | 1.09 (about 9% overhead beyond IT equipment) | |
| Google greenhouse gas emissions, 2019–2023 | Up 48%, attributed largely to data center energy demand | The Guardian |
Global averages hide local effects. Data centers cluster in particular regions, where they can strain grids, delay coal plant retirements or compete for water used in cooling. Water use depends heavily on cooling design and climate; some facilities use evaporative cooling that consumes significant water, while others use closed-loop or air cooling that uses less water but more electricity.
What the industry is doing
Hardware and model efficiency. Each generation of AI accelerators delivers more computation per watt, and techniques such as quantisation, distillation, mixture-of-experts architectures and smaller task-specific models cut inference energy. See our guides to quantisation for edge deployment and small language models.
Data center efficiency. Leading hyperscale facilities report PUE values close to 1.1, meaning little energy goes to cooling and overhead; older facilities are typically less efficient. Liquid cooling is increasingly used for dense AI hardware.
Clean energy procurement. Large cloud providers buy renewable and other low-carbon power and have set net-zero or carbon-negative targets. But rising AI demand has pushed some companies’ reported emissions up, making those targets harder to meet, as Google’s figures show.
The rebound effect. Efficiency gains make AI cheaper, which tends to increase usage. Whether total energy falls or rises depends on that balance, and so far demand growth has dominated.
The transparency problem
Most AI developers do not disclose training energy, inference energy per model or the methods behind their figures. Estimates therefore rely on academic research, partial disclosures and modelling. Regulation is beginning to require more: under the EU AI Act, providers of general-purpose AI models must document the known or estimated energy consumption of their models, which may be estimated from computational resources where it is unknown (EU AI Act, Annex XI). Standardised, independently verified reporting would allow fair comparisons between models and providers.
What individuals and organisations can do
- Use the smallest model that does the job, and avoid long reasoning modes for simple tasks.
- Prefer providers that publish energy and emissions data and run in low-carbon regions.
- Batch and cache repeated requests in applications rather than regenerating identical outputs.
- Include AI workloads in sustainability reporting, using provider data where available.
Common questions
Is using a chatbot bad for the environment? For an individual, a typical text query has a small footprint by Google’s published measure, comparable to seconds of television. The concern is aggregate demand across billions of users and heavier workloads like video generation.
Is training or inference worse? For widely used models, cumulative inference likely exceeds the one-off training cost over the model’s life, though exact ratios are rarely disclosed.
Does AI help the climate too? AI is used for grid forecasting, materials research and weather modelling, but these benefits are hard to quantify and do not offset emissions automatically.
Our analysis
A single training run is no longer the main issue. Aggregate inference at billions of queries, plus local strain on grids and water near data center clusters, matters more. Per-query figures look small and are falling fast, but totals depend on volume, and the IEA’s projections show demand rising regardless. Company-reported numbers also use different methodologies, so they are hard to compare. Standardised, independently audited disclosure, which the EU AI Act begins to require, would make the debate far more grounded than viral statistics have allowed.