Guillermo Llopis
GuillermoLlopis
AI & ML interests
None yet
Recent Activity
posted an update about 17 hours ago
How much carbon, electricity and water does your AI use? Nobody outside a provider can meter a prompt, but you can estimate it openly, with a range.
Published per-prompt figures count different things. Google reports 0.24 Wh for the median Gemini text prompt, Epoch AI about 0.3 Wh for a GPT-4o query, and SAEF estimates 0.03 to 0.37 Wh across small to large model classes for a 1,210-token exchange. Length matters most, a 10,000-token input is about 2.5 Wh.
Water depends on where the model runs, 0.15 to 0.94 mL for a mid-size chat exchange across 14 cloud regions.
Your own footprint in three steps. Count messages or tokens, pick a model class, apply the grid of the region where it runs. 20 chat messages a day for a year come to about 0.4 kg CO2e on a US East grid (band 0.2 to 1 kg), about 0.1 kg on a Swedish grid.
Try it in the browser https://huggingface.co/spaces/GuillermoLlopis/saef-ai-footprint-estimator or pip install saef. Data https://huggingface.co/datasets/GuillermoLlopis/ai-inference-emission-factors (CC BY 4.0). The method is now peer reviewed and published in Sustainable Production and Consumption, https://doi.org/10.1016/j.spc.2026.11.003
These are estimates, not measurements. I build SAEF at SOMA AI, which also sells an AI footprint report. The data, package and Space are free. updated a dataset about 18 hours ago
GuillermoLlopis/ai-inference-emission-factors updated a Space 1 day ago
GuillermoLlopis/saef-ai-footprint-estimatorOrganizations
None yet