Companies increasingly need to report and reduce the environmental impact of their digital activities, including AI. Measuring the footprint of AI is tricky but possible, and many reduction measures also save money.

What contributes to AI’s footprint

  1. Operational energy: electricity used to train models and run inference.
  2. Carbon intensity of that electricity: the same computation emits very different amounts of CO₂ depending on whether the grid runs on coal, gas, hydro, nuclear or renewables, and even on the time of day.
  3. Data-center overhead: cooling and power distribution, measured by power usage effectiveness (PUE).
  4. Water: consumed by data-center cooling and, indirectly, by electricity generation.
  5. Embodied impacts: manufacturing chips, servers and buildings.

What research has shown

A 2019 study by Emma Strubell and colleagues drew attention to the emissions of training large NLP models, especially when extensive hyperparameter searches were included. Later work emphasized that location, hardware efficiency and data-center practices can change emissions by large factors. Research on water, such as a 2023 paper by Pengfei Li and colleagues, highlighted that AI’s water footprint depends strongly on where and when computation happens.

How to measure

  • For your own training and inference: tools such as CodeCarbon estimate energy use and emissions from hardware utilization and local grid data.
  • For cloud services: major cloud providers offer carbon footprint dashboards for customers.
  • For API use of third-party models: precise figures are often unavailable; estimate based on token volumes and any data the provider publishes, and ask vendors for disclosures.
  • Use location-based and market-based accounting consistently, following greenhouse gas reporting standards.

How to reduce the footprint

Use less compute

  • Choose the smallest model that meets quality needs.
  • Use distilled or quantized models for inference.
  • Avoid reasoning modes and very long contexts when not needed.
  • Cache repeated results and batch non-urgent jobs.
  • Avoid unnecessary retraining; reuse pre-trained models and fine-tune efficiently.

Run it in cleaner places and times

  • Choose cloud regions with low-carbon electricity.
  • Schedule flexible workloads when renewable supply is high (“carbon-aware computing”).

Choose efficient infrastructure

  • Prefer providers with efficient data centers (low PUE), renewable energy procurement and transparent reporting.
  • Consider water usage effectiveness and cooling methods when choosing locations.

Measure and govern

  • Include AI in sustainability reporting.
  • Set efficiency targets alongside cost targets — they usually align.

A simple estimate you can make

To get a rough sense of the electricity used by an AI workload, multiply the number of accelerators by their average power draw and the hours they run, then add overhead for cooling and power delivery using the data center’s power usage effectiveness (PUE). For example, 1,000 accelerators drawing 0.7 kW each for 30 days use about 500 MWh before overhead; with a PUE of 1.2, roughly 600 MWh. Multiply by the carbon intensity of the local grid to estimate emissions. Real figures depend on utilization, hardware and location — which is why transparency from operators matters so much.

Related: how much energy does AI use?, liquid cooling in AI data centers and AI and the power grid in the energy sector.

Bottom line

Credible footprint numbers require transparency on energy, location, hardware and usage. Treat single estimates with caution and look for methodology.

What to watch next

  • Standardized methods for reporting AI energy and emissions.
  • Provider disclosures of per-query or per-token energy estimates.
  • Regulation requiring data-center energy and water reporting.
  • Carbon-aware scheduling built into cloud platforms.

Key takeaways

  • AI’s footprint includes operational energy, grid carbon intensity, cooling overhead, water and embodied impacts.
  • Location and timing can change emissions dramatically for the same computation.
  • Tools like CodeCarbon and cloud dashboards help estimate footprints.
  • Smaller models, caching, batching and cleaner regions reduce impact and cost.
  • Include AI in sustainability reporting and ask vendors for disclosures.

Frequently asked questions

Is training or inference worse for the environment?

Training is a large one-time cost; inference adds up over millions or billions of requests. For popular services, inference can dominate over a model’s lifetime.

Does using a cloud region with renewable energy really help?

Yes, grid carbon intensity varies widely by region, so location choices can significantly reduce emissions for flexible workloads.

How much water does AI use?

It depends on data-center cooling methods, climate and electricity sources. Estimates vary widely, which is why location-specific reporting matters.