How much energy does generating an AI image or video use?

MIT Technology Review reports about 1,141 joules of GPU energy for one standard 1024 by 1024 image from the open Stable Diffusion 3 Medium model, and an estimated 2,282 joules, or 0.63 Wh, once doubled for the rest of the system. For a five-second CogVideoX video it reports about 3.4 million joules, which is 0.94 kWh. Both start from measurements on open models by outside researchers, from 20 May 2025, and neither is a figure for a commercial service. SOMA AI (somaai.earth) does not price image, speech or video at its text factors, because they sit outside the current version of the SOMA AI Emission Factors (SAEF).

The same article says the newer video model uses more than 30 times the energy of the older one per video, and more than 700 times the energy of generating a high-quality image. We give the second claim as MIT's own, since it depends on which image figure is the base.

On these figures one image costs about 20 times a response from the small Llama model and about a third of a response from the 405B model, and a video is a different order of magnitude. At 0.94 kWh, three video attempts come to 2.83 kWh, which is 98% of the 2.9 kWh in MIT's example day. In that illustration the video attempts account for almost all of the day's energy. The ML.ENERGY row is GPU energy only, so compare it with the GPU rows and not with the doubled figures. For that reason the ML.ENERGY ratio uses the image at 1,141 J of GPU energy. Text figures from other sources are set out on the ChatGPT query energy page.

SAEF covers token-based chat and completion inference. Embedding, speech, image, video, moderation and reranking models do not work in tokens in the same way, their energy per unit behaves differently, and the text tables would be the wrong instrument for them. SOMA reports those workloads on their own line rather than pricing them at a chat factor. The check is by model name, run here on the dataset's own function.

For scale, a SAEF text exchange of 1210 tokens uses 0.03 Wh for class A (small), 0.099 Wh for class B (mid-size), and 0.37 Wh for class C (frontier), facility energy, as an estimate. That figure must not be multiplied by a count of images.

• A unit that fits the workload, such as per image at a stated resolution and step count, or per second of video, instead of per token.
• A statement of whether the figure is GPU only or includes the rest of the system. MIT's image figure applies a doubling assumption, and its video page does not say.
• The serving region and grid, as for text.
• The count of items generated, from the provider's usage export, including failed and repeated attempts.

The published figures above are for open models run by researchers. They say little about a commercial service whose model, hardware and settings are not disclosed, so any number for one would be an assumption and should be labelled as one.

Keep image, speech and video generation as a separate line from the text inventory, with the count of items and the source of any energy figure used. Do not convert images to tokens. The Scope 3 guide explains how the text lines are built. The free calculator handles token-based text use, and the methodology page states what is out of scope. The figures guide shows why published numbers differ.

Figures per query are one plain exchange, before the Tier 2b reasoning uplift. Electricity term. GPU energy per token from the ML.ENERGY Leaderboard v3, a facility overhead of 1.20, and grid carbon intensity from EPA eGRID 2023 and Ember 2023. Embodied hardware, data-centre construction and amortised training terms are documented on the methodology page. Water combines on-site cooling water (Li et al. 2025) with the water embedded in grid electricity (Reig et al. 2020). The modelled range of the models inside a class, around the class value, is −41% to +48% for class A, −39% to +24% for class B and −67% to +55% for class C.

Dataset: AI Inference Emission and Resource Factors for Corporate GHG Inventories (SOMA AI Emission Factors (SAEF)), Llopis-Garcia, G., SOMA AI, 19 October 2026, CC BY 4.0, concept DOI 10.5281/zenodo.20443585, which resolves to the latest version (this version 10.5281/zenodo.23371447, earlier version 2 values at 10.5281/zenodo.22767475). Method: Accounting for Purchased AI Inference in Corporate GHG Inventories: A Token-Based Method, arXiv:2606.10660, under peer review.

Partners can fetch any row from the AI factors API, and the whole table is a CSV download.

• How much energy does a ChatGPT query use? Text figures to set against image and video
• AI environmental impact statistics: energy, carbon and water, with sources. Energy, carbon and water in one place
• AI carbon footprint tools compared. What each tool covers