October 18, 2026

Research on measuring the energy, carbon, and water footprints of corporate AI use

Companies are using more AI every quarter, through productivity assistants, coding tools, direct model access, and features built into the software they already buy. Data centers are expected to use 9% to 17% of U.S. electricity by 2030, up from about 5% in 2025. Investors, auditors, and regulators are starting to ask what share of that belongs in a company's own emissions inventory.

Our research develops open, transparent methods that let companies estimate AI emissions with the data they have today, improve those estimates as providers disclose more, and find the choices that reduce emissions as AI use grows. You can see a summary of related research and publications below. We'll add to this brief as new research becomes available.

Watershed research on AI emissions

Measurement: A standard method for corporate AI emissions

Estimating GHG emissions from AI use: Framework for corporate-level measurement An open framework for estimating the electricity use and emissions of a company's AI use. It sets a broad system boundary that includes training, idle capacity, and hardware manufacturing; reports results in kilograms of CO2e per million tokens, the unit companies already use to track AI cost; and offers three calculation tiers, from a spend-based estimate to figures reported by AI providers. It also proposes the data AI providers could disclose to make customer estimates more accurate. Developed with contributors from Stanford University and Tsinghua University.

An open framework for AI emissions measurement Watershed blog, July 2026. A shorter guide to the framework: why AI emissions are hard to measure, how the three tiers work, and the practical levers that lower AI's footprint.

Electricity and location: How where AI runs shapes its footprint

Watt's the difference? Your chatbot's carbon depends on its ZIP code Watershed, September 2026. The analysis holds the electricity use of a single AI prompt constant and changes only the grid that powers it. The same prompt's emissions vary 34-fold across U.S. states, and about 100-fold once accounting methods are included.

Cornerstone eGRID 2024 Dataset, Cornerstone, 2026. For two decades, the U.S. EPA's eGRID database was the standard source for regional grid emissions factors. When EPA stopped updating it in 2025, Cornerstone, a collaboration between Watershed, Stanford University, and ERG, took over maintenance and released the 2024 edition. Our AI emissions framework and regional analysis use it to calculate grid emissions.

Key findings

Across our framework and analysis, one pattern holds: the inputs to an AI emissions estimate matter as much as the method. The same AI use can produce estimates several times apart depending on the data available, and the biggest gaps sit with information only AI providers hold.

Why AI emissions estimates disagree

Better data changes the answer by about four times. For an illustrative company spending $100,000 a year on a frontier AI model, the same use produced 13.4 tCO2e with a spend-based estimate, 3.7 to 5.4 tCO2e from token volumes, and 3.2 to 4.5 tCO2e from provider-reported figures. (Bistline et al., 2026)
Lab benchmarks overstate real-world energy use. Benchmarks that test models one request at a time, without the batching and caching production systems use, can overstate inference energy by 4 to 20 times. (Bistline et al., 2026, citing Oviedo et al., Joule, 2026)
Where providers have published data, AI's footprint has come in lower. Google reports that a median Gemini Apps text prompt uses 0.24 watt-hours, and that the energy per prompt fell 33-fold over one year. (Bistline et al., 2026, citing Elsworth et al., 2025)
Training is the largest unknown. In the case study, model training accounts for 32% to 56% of total emissions. Depending on how many tokens a model serves over its lifetime, a number no frontier provider discloses, that share could range from 5% to 83%. (Bistline et al., 2026)

What drives AI's footprint

Where a prompt runs. Holding a prompt's electricity use constant, its emissions are 34 times higher on West Virginia's coal-heavy grid than on Vermont's hydropower-heavy grid. (Bistline, 2026)
Which model handles the task. Reasoning models can use about 30 times more energy than smaller models for equivalent queries. (Bistline et al., 2026)
What kind of work it is. Electricity per task varies more than 100-fold across workloads, from a short chat prompt to reasoning and agentic work, where a single request can trigger 5 to 50 or more model calls. (Bistline et al., 2026)
How data centers are powered. The grids where U.S. data centers operate average about 307 grams of CO2e per kWh today, and would rise to 351, nearly 15% higher, if planned projects are built. About 30% of planned U.S. data center capacity would generate its own power on site, nearly three-quarters of it from natural gas. (Bistline, 2026; Bistline et al., 2026)

What this means for sustainability teams

AI is a small share of most companies' footprints today, and that share is growing. Building the data and provider relationships now makes the number defensible before it becomes material.

Questions to ask your AI providers

In which regions are our requests served?
Can you report emissions per million tokens on both a location-based and a market-based basis?
Are training emissions disclosed, and how are they allocated across usage?
Does any of the electricity come from on-site generation, such as natural gas, rather than the grid?

Where to start

Inventory AI use by channel. List each vendor and how your company reaches it: chat assistants, coding tools, direct API access, cloud platforms, or AI features inside other software.
Estimate at the level your data supports. Use spend as a starting point for everything, and move to token-based estimates wherever you have usage logs.
Ask providers for better data, starting with electricity per token and serving region.
Report electricity alongside emissions. Tracking both shows whether a change came from using AI more efficiently or from cleaner power.

Levers that reduce AI emissions

Match the model to the task. Route simple requests to smaller models and save the largest reasoning models for work that needs them. This also lowers cost.
Choose lower-carbon regions where your provider lets you select where workloads run.
Write efficient prompts. Trim unneeded context, cut redundant calls, and avoid agentic loops that don't add value.
Favor providers with credible clean electricity and published efficiency data.

About the researchers

Dr. John Bistline: Head of Science at Watershed. His research focuses on large-scale energy modeling, corporate decarbonization, and the climate impacts of AI and data center growth. He previously led energy systems research at EPRI, is a Research Affiliate at MIT, and serves as a contributing author to the IPCC Seventh Assessment Report.

Shaena Ulissi: Climate and Data Scientist at Watershed, focusing on standardizing greenhouse gas accounting methodologies and evaluating AI systems in life cycle assessment. She previously spent eight years in environmental and climate consulting at Ramboll and holds a master's degree in atmospheric science from MIT.

Dr. James Joyce: Sustainability Science Lead at Watershed, responsible for corporate GHG, supply chain, and product carbon footprint methodologies. He previously led corporate GHG tracking and science-based targets at Unilever and holds a Ph.D. in Environmental Strategic Analysis from KTH Royal Institute of Technology.

Dr. Alex Smith: Leads carbon modeling methodologies and AI products at Watershed. He previously led financial data science teams at Stripe and worked on machine learning and inference systems at the World Bank and Salesforce. He studied at UC Berkeley.

Dr. Mo Li: Climate Science Lead at Watershed, leading environmental modeling systems, Scope 3 supply-chain emissions databases (CEDA), and open data collaborations (Cornerstone). She previously developed environmental input-output frameworks for the U.S. EPA and holds a Ph.D. in Natural Resources Science and Management from the University of Minnesota.

Dr. Steven J. Davis: Professor of Earth System Science at Stanford University and Director of the Sustainable Solutions Lab. His research covers global energy scenarios, net-zero transitions, embodied carbon in trade, and corporate climate strategy.

Dr. Sangwon Suh: Professor at Tsinghua University's School of Environment and Research Advisor to Watershed, where he previously served as Head of Science. He has over two decades of experience in life cycle assessment and GHG accounting, including roles as an IPCC Coordinating Lead Author and GHG Protocol working group member.