Why should AI providers say where your prompt runs?

The same AI request can carry more than seven times the carbon depending on which EU grid serves it. The companies that must report those emissions cannot see it, and current EU rules stop one step short of letting them.

A company in Barcelona that sends a request to a cloud language model rarely knows where the answer is computed. It might be a data centre in Sweden. It might be one in Poland. For that company's emissions report this is not a detail. It decides most of the number.

We estimate that one exchange with a mid-size language model emits about 0.009 grams of CO₂e on Sweden's average grid and about 0.070 grams on Poland's, a gap of more than 7 times inside a single market. The EU-27 average is 0.029 grams. Across the 143 cloud regions operated by Amazon, Microsoft and Google, the spread is about 12 times, from 0.008 grams on the Quebec (Canada) grid, where the dataset maps AWS ca-central-1, to 0.095 grams on the Bahrain grid.

These are estimates. They combine 2023 national grid averages, mostly from Ember, a fixed request of 1,210 tokens, a standard allowance for cooling and other facility overhead, and a fixed share for server hardware, the data-centre building and model training that is the same on every grid. They are not readings from any provider's servers, and a country's average is not the mix of a particular data centre, which may sit on a cleaner or dirtier part of the grid or buy renewable electricity. That limitation is the argument of this note. Nobody outside the providers can do better, because the providers do not say where requests are served or how much energy each one uses.

A fraction of a gram per request sounds trivial. Multiplied across a workforce, customer-facing chatbots and automated agents, it becomes a line in a carbon inventory. What matters more is who has to count it.

The EU's Corporate Sustainability Reporting Directive was narrowed this year to the largest companies, those with more than 1,000 employees and more than €450 million in net turnover. The simplified reporting standards the Commission adopted on 3 July 2026 still ask those companies for their value-chain, or Scope 3, emissions. Purchased cloud and AI services belong in that value chain. A sustainability team that wants to account for AI today has two choices. It can apply a global average and accept an error that can run several-fold in either direction, or it can ask the vendor and usually receive a corporate sustainability page with no figure for its own usage. Our Scope 3 guide sets out the method.

Two EU instruments touch this problem. Neither reaches the customer.

The first is the AI Act. Article 53(1)(a) requires providers of general-purpose AI models to keep technical documentation containing at least the items in Annex XI, which include the "known or estimated energy consumption of the model". That obligation has applied since 2 August 2025. But the documentation exists "for the purpose of providing it, upon request, to the AI Office and the national competent authorities", and Article 53(7) places it under the Act's confidentiality rules. The separate list of information providers must give to companies building on their models, Annex XII, does not mention energy at all. The figure also describes the model, chiefly its training, not where or how efficiently each request is served. Our guide to AI Act energy documentation goes through the text.

The second is the Energy Efficiency Directive. Under Article 12 and Annex VII, data centres with an installed power demand of 500 kW or more report energy consumption, power usage effectiveness, renewable share, water use and waste-heat recovery to a European database, which the Commission publishes in aggregated form. In September 2026 the Commission proposed a common rating scheme and opened a consultation on minimum performance standards for data centres, which closes on 14 December 2026. This is useful work. It tells regulators how efficient a building is. It does not tell a tenant which building served its requests, or what share of that building's energy its workload used.

So the EU holds model-level energy data that stays confidential and facility-level data that is published in aggregate. What is missing is the level a buyer can act on, meaning this service, in this region, this month.

The first objection is about accounting method. Providers will point out that they buy renewable electricity, and under market-based accounting a request served in Poland may count as close to zero. The analytics firm Watershed made this point well in a September 2026 analysis. Starting from Google's disclosed median of 0.24 watt-hours for a Gemini text prompt, Watershed estimated about 0.006 grams of CO₂ on Vermont's hydro-heavy grid and 0.212 grams on West Virginia's coal-heavy grid, a 34-fold difference, and showed that market-based accounting can take the same prompt to near zero. Its conclusion was that "any per-prompt or per-token number without a stated grid method is incomplete." We agree. Watershed reached that conclusion with US state data and we reached a similar one with national data for Europe and beyond. Both methods need the same missing input, which is where the electricity was drawn. Disclosure of the serving region lets an auditor compute either.

The second objection is that electricity is not the whole story. Robinson Meyer at Heatmap reported that capital goods, the concrete, steel, server racks and silicon of new data centres, made up more than 40% of Microsoft's total emissions, over 9 million tonnes, with Google reporting almost as much in a broader category. The figures above carry only a modelled average share for hardware and buildings, the same in every region, not any provider's own capital goods. That is a real gap. But electricity is the part a buyer can change today by choosing a region, and per-customer usage data is also what any fair allocation of embodied emissions to customers would need.

The third objection is that estimates like ours may simply be wrong. They may. Real energy per request depends on hardware, batching, model size and load, none of which outsiders see. That is the strongest case for disclosure, not against it. Google's per-prompt figure shows a provider can publish such a number when it chooses to.

The AI Act already gives the Commission the tools. Article 53(5) lets it adopt delegated acts detailing measurement and calculation methods for the energy item in Annex XI, and Article 53(6) lets it amend Annexes XI and XII as technology evolves. The data-centre consultation offers a second route. Three changes would close the gap.

• Serving region for every business customer. Providers of general-purpose AI models and the cloud services that resell them in the EU should report to each business customer, at least monthly, the share of its requests served in each cloud region or country. Adding this to Annex XII would make it part of what downstream companies already receive.
• Energy per request, with the method stated. The same report should give estimated energy per request or per million tokens, by model and region, including facility overhead, and say whether the figure is measured or estimated. Methods set under Article 53(5) would make figures comparable across providers.
• A tenant clause in the data-centre standards. The minimum performance standards now under consultation should oblige operators above the 500 kW threshold to give each tenant its share of energy use and the grid it was drawn from, so the facility data already collected reaches the companies that must report it.

None of this caps AI use or picks a winning accounting method. It gives European buyers the two facts they need to count AI honestly, where it ran and how much energy it used, and it gives providers with clean regions a reason to say so.

SOMA AI submitted public feedback to the Commission's call for evidence on minimum performance standards for data centres in Europe on 11 November 2026. It asks that operators above the 500 kW threshold give each tenant its share of the facility's energy use and the grid it was drawn from, that cloud and AI providers tell business customers where their usage is served, and that figures state whether they are metered or estimated. The submission is on the Commission's Have Your Say portal.

Ask each AI provider where your usage is served, record the answer and use that region's factor, not your office's country. The free calculator applies the region for you. This note does not claim

• That the per-request figures are readings from any provider's servers. SOMA AI estimates them from the SAEF factors and a typical exchange.
• That a country's average grid is the electricity mix of a specific data centre, or that it reflects renewable contracts.
• That any provider serves any given model from the regions named. Regions are dataset inputs.
• That the hardware, building and training share in these figures reflects any provider's actual capital goods. It is a modelled average, the same in every region.

SOMA AI publishes the dataset behind these estimates under an open CC BY 4.0 licence (DOI 10.5281/zenodo.20443585). The method was accepted for publication in Sustainable Production and Consumption, and the equations are on the methodology page.

• Where does an AI query emit the least? The carbon footprint of AI by country. The grids of 210 countries and the EU-27 ranked
• How much CO₂ and water does an AI query served from a data centre in Spain use? The Spanish grid against the EU, with sources
• How can a company cut the carbon footprint of its AI use without losing quality? Six levers, with the region among them
• EU AI Act energy documentation. What Article 53 and Annex XI require