by REEL LCICommunityCarbon and water footprint estimates for AI language models and AI hardware, with sources, low and high values, and the assumptions shown.
Carbon of Compute estimates the greenhouse gas and water footprint of running AI language models, per 1,000 output tokens or per task. Each figure is split into four parts: the electricity the hardware draws, the manufacturing of that hardware spread over its working life, the datacenter facility plant, and optionally a share of model training. Footprint results come with low and high values.
Ask AgentOne for one model's footprint, or compare models under the same assumptions. You can change the grid region, datacenter efficiency (PUE), hardware lifetime, utilization and cooling regime. The extension also returns the manufacturing footprint of AI accelerators (NVIDIA H100, H200, B200, A100, AMD MI300X) split into logic die, HBM memory, package and board, an 8-GPU server line by line, and datacenter facility factors from manufacturer EPDs.
Footprint and ranking answers link to the same view on carbonofcompute.com and to the method report. Hardware manufacturing is built on REEL LCI, a life cycle inventory of electronics manufacturing whose foreground inventory comes from public sources, linked to licensed ecoinvent 3.12 background data.
Figures are estimates from public evidence for a reference deployment, not a provider's own measurement. The extension only reads published data. It needs no account and takes no actions.