In short
When you buy AI, ask for energy and emissions per unit of use with the method behind them, location-based as well as market-based carbon figures, water use, and where the work runs. Put the answers into tender scoring and contract clauses, and prefer suppliers whose figures follow published methods and can be checked.
Key points
- Ask for energy and emissions per request or per task, with the method and system boundary, not just company-wide totals.
- Ask for both location-based and market-based carbon figures; they can differ greatly.
- In the EU, providers of general-purpose AI models must document the model's known or estimated energy consumption under the AI Act.
- Large EU data centres (500 kW of IT power or more) report PUE, WUE and renewable energy figures each year under the Energy Efficiency Directive.
- Score energy evidence in tenders, and write reporting, model-choice and opt-out rights into the contract.
Why bring sustainability into AI buying?
Because most of an organisation's AI footprint sits with its suppliers, and procurement is where you have the most influence over it. Few organisations train their own models; they buy services, licences and cloud capacity.
Energy and carbon also track cost. Buyers want cost, speed and quality, and a well-designed AI service that uses less computation usually delivers all three. Asking about energy is a way of asking whether a service is well engineered.
What should I ask AI suppliers?
Ask for figures per unit of use, the method behind them, and where the work runs. The Green Web Foundation's guide to AI's environmental impact suggests questions along these lines; the list below builds on it.
Energy and carbon
- What is the energy use per request, per task or per 1,000 queries for the models we would use, and how is it measured or estimated?
- What does the figure include: accelerators only, whole servers, idle capacity, data-centre overhead, network?
- Can you give emissions on both a location-based and a market-based basis? The GHG Protocol's Scope 2 guidance asks companies with access to contractual instruments, such as renewable certificates, to report both.
- Does the model have a published energy rating, such as a Hugging Face AI Energy Score? Proprietary models can be rated using its container without disclosing their details.
- What were the training emissions, hardware and location, and are they in the model card?
Location, water and hardware
- In which regions and data centres will our workloads run, and what is the grid carbon intensity there?
- What are the sites' PUE (power usage effectiveness) and WUE (water usage effectiveness), and do water figures include water used to generate the electricity?
- What is your hardware refresh cycle, and how are retired servers reused or recycled?
Choice and control
- Which model sizes are available, and can we route simple tasks to smaller models?
- Can bundled AI features be turned off for our users?
- Will you give us usage and energy data for our account, at least monthly?
What evidence should I request?
Prefer figures produced by a published method, with a stated boundary, that a third party could check.
| Evidence | Why it helps |
|---|---|
| Energy per request on a named benchmark (for example AI Energy Score) | Comparable between suppliers |
| An SCI score (ISO/IEC 21031:2024) for the service | Per-unit rate with a defined boundary; offsets and certificates can't reduce it |
| Model cards with training emissions, hardware and location | Shows the training cost you are inheriting |
| Data-centre PUE, WUE and renewable share | Overhead and water at the sites used |
| EU AI Act technical documentation | General-purpose AI model providers must document known or estimated energy consumption (Annex XI) |
| EU data-centre reporting | Data centres with at least 500 kW of IT power report PUE, WUE, energy reuse and renewable energy figures annually under Delegated Regulation 2024/1364 |
| Ecolabel criteria such as Blue Angel DE-UZ 215 | Established criteria for resource- and energy-efficient software |
| Independent verification or audit | Lowers the risk of relying on a vendor's own claim |
Treat company-wide totals, net zero pledges and offset purchases as context, not as evidence about the service you are buying. Vendors often treat energy data as commercially sensitive; record refusals and score them.
How do I put this into a tender?
Make sustainability a scored criterion with a clear evidence standard, not a tick box.
- Define the functional unit you care about: per document processed, per conversation, per user per month.
- State the evidence you'll accept and how you'll score it (for example measured and independently verified scores higher than estimated and self-reported).
- Weight it alongside price, quality and security. Even a modest weighting changes supplier behaviour.
- Ask for a reduction plan: what the supplier will do to cut energy per unit over the contract.
- Score right-sizing: credit suppliers who offer smaller models, routing and opt-outs.
Which contract clauses help?
Write the commitments into the contract so they survive after the sales conversation. Adapt these with your legal team:
- Reporting: the supplier provides energy and location-based and market-based emissions for the customer's usage at least quarterly, with the method and boundary stated.
- Method changes: the supplier notifies the customer of any change in measurement method and restates prior figures on the new basis.
- Location: workloads run in named regions; moves to higher-carbon regions need notice.
- Model choice: the customer may select smaller or more efficient models where they meet the agreed quality.
- Defaults and opt-outs: AI features not in scope can be disabled for the customer's users.
- Claims: the supplier will not describe the service to the customer as "green", "carbon neutral" or similar without the evidence set out in the contract. See honest green claims.
- Review: sustainability performance is reviewed annually, with the right to renegotiate if reported figures rise per unit of use.
- End of life: data deletion and hardware disposal follow stated environmental standards.
Checklist
- [ ] Functional unit defined for the service
- [ ] Energy and emissions per unit requested, with method and boundary
- [ ] Location-based and market-based figures both requested
- [ ] Regions, PUE and WUE of hosting sites requested
- [ ] Model sizes, routing and opt-outs asked about
- [ ] Sustainability scored in the tender with an evidence standard
- [ ] Reporting, location, model-choice and claims clauses in the contract
- [ ] Annual review date set
What to do next
- Add three questions to your next AI purchase: energy per unit with method, location-based emissions, and available model sizes.
- Draft a standard sustainability annex for AI contracts using the clauses above.
- Agree an internal measurement method so supplier figures can be compared with your own.
- Check hosting claims with hosting and data centres.
- Write procurement rules into your green AI policy.
- Find suppliers and advisers in the products and consultants directories, and see the business guide.
Frequently asked questions
Will AI suppliers give me energy figures?
Some will; many treat them as commercially sensitive. Asking still matters: it signals demand, and a refusal is itself useful information when you score bids.
What is the difference between location-based and market-based emissions?
Location-based figures use the average carbon intensity of the grid where the electricity was used. Market-based figures reflect contracts such as renewable certificates. The GHG Protocol asks companies with access to such contracts to report both.
Is a supplier's net zero pledge enough?
No. Ask what the service uses now, per unit of work, how that is measured, and what the supplier is doing to reduce it, not only how it is offset.
Do small organisations need to do this?
A short version helps any buyer: three or four questions in every AI purchase, and a preference for suppliers who answer them well.
Sources
- Thinking about using AI? Here's what you can and (probably) can't change about its environmental impact, Green Web Foundation (2024)
- AI Act, Annex XI: technical documentation for providers of general-purpose AI models, European Commission AI Act Service Desk (2024)
- Commission Delegated Regulation (EU) 2024/1364 on the first phase of a common Union rating scheme for data centres, EUR-Lex (2024)
- GHG Protocol Scope 2 Guidance, GHG Protocol (2015)
- AI Energy Score, Hugging Face (2025)
- Software Carbon Intensity (SCI) specification, Green Software Foundation (2024)
- Blue Angel for resource and energy-efficient software products (DE-UZ 215), Blauer Engel (2024)
- Displaying carbon emissions for your model, Hugging Face Hub documentation (2025)
- Web Sustainability Guidelines, W3C (2025)
Related
Green AI hosting and data centres
How to choose where AI runs: grid carbon intensity, PUE and WUE, renewable energy claims, water use and siting, with questions to ask hosts.
Honest green claims about AI
How to avoid greenwashing about AI: what you can and can't claim, rebound effects, honest reporting, and the UK, EU and US rules on environmental claims.
Green AI for business
A practical guide for businesses on cutting the energy, cost and carbon of AI: model choice, suppliers, measurement, procurement and honest claims.
Energy labels and benchmarks for AI
Where AI energy ratings stand in 2026: AI Energy Score, MLPerf Power, EU AI Act documentation and data-centre rating, and what labels can show.