8 October 2026 Cleer dashboard: How much energy and emissions does AI inference use? This is great: Sustainable AI Group have a dashboard that models AI model energy use and emissions: Cleer, and I like to see chunky calculation and tech reports with it. Closed-model figures are modelled from the nearest open-weight proxies, so they’re an estimate (please just disclose your emissions, AI giants!). A closed model’s performance on a knowledge benchmark is used to match it to the 5 closest open models they’ve tested (Technical report 5.2). Exceptions: a few models, like Claude Fable, know higher than any open model, so they lean on a single proxy. And Gemini is anchored to Google’s own published energy figure (see Technical report Annex A). “Results include accelerator and host-server energy, idle capacity, cooling and power distribution, and embodied emissions from hardware and data center construction. Scenario default: behind-the-meter gas generation at 640 gCO₂e/kWh, combined with US-average data center assumptions for PUE (1.45)” - you might think these defaults a bit pessimistic but there is a lot of planned BTM gas power for data centres, especially in the US where there are orders for ~75 GW of gas generation. And their rule is strict: if there’s public evidence a provider uses behind-the-meter gas anywhere, all your usage with them gets the gas factor (see the calculation report 6.4). But of course, you want to know your actual figures and how you could improve them. Can you assume your tokens run on clean hydro rather than dirty gas? Or that Claude runs on a low-carbon grid, or that Gemini is hosted on one of Google’s 1.09 PUE DCs? The calculation report says no: unless the provider names the actual site, you use the 1.45 default PUE and the US-average grid (410 g), or the gas factor if there’s evidence of behind-the-meter gas (Calculation report 5.4, 6.1, 6.4). And a provider’s general renewables claims don’t lower your market-based figure unless the contracts can be tied to your own usage (Calculation report 6.2). In their worked example, model development (training, experiments, failed runs, research) comes out at 58% of the inventory footprint, taking the total to 2.37 times the inference-only figure (Calculation report 10.2). But that comes from an assumed 1.5:1 development-to-serving ratio, built from spending and compute-split estimates rather than energy measurements (Calculation report 9.3). They think the energy ratio is probably higher. I’d assumed inference was the bigger share by now, so this is the number I most want an AI provider to disclose. The dashboard shows the inference footprint (it’s in the title, and the per-session figures are operational + embodied), which suggests that for your own inventory you’d multiply the inference footprint by a factor to add development. x2.37 if you’re following the worked example (Calculation report 10.2, 10.4). However, not everyone adds development, eg the Green Software Foundation’s SCI for AI leaves it out of the user’s score (Calculation report Annex C). Good table translating AI terms into GHG terms. One day your AI will say “this conversation used X tokens, which on today’s energy means emissions of Y gCO2e, and tbh if you’d had this conversation on our older model Z instead, you’d have got the same result for half the energy and emissions”.