OpenLLM-France's Lucie 7B training emits 21 tCO2eq, half from hardware manufacturing

A new life cycle assessment (LCA) of pre-training the 7B-parameter Lucie 7B open-source large language model reveals that hardware manufacturing accounts for roughly half of total carbon emissions—a factor often overlooked in prior AI energy studies. Conducted by the OpenLLM-France consortium on the NVIDIA H100 GPU partition of the Jean Zay supercomputer (operated by IDRIS/CNRS), the study logged 574,564 GPU-hours and a total carbon footprint of 21 tCO2eq. The annual footprint of the entire H100 partition was 417.5 tCO2eq, split nearly equally between manufacturing and operational energy, highlighting that embodied emissions from chip fabrication and infrastructure are as significant as runtime electricity use.

The LCA also reports an effective carbon intensity of 36.7 gCO2eq per H100 GPU-hour, on-site water consumption of ~76 m³ for training, and a Water Usage Effectiveness (WUE) of 0.07 L/kWh at IDRIS. Notably, 37% of waste heat was recovered for urban heating (ERF factor of 0.37), partially mitigating thermal impact. Following the AFNOR SPEC 2314 'Frugal AI' standard and the Labos 1point5 methodology, the study decomposes emissions by subsystem (compute, storage, power chain, cooling). This granular, transparent reporting sets a new benchmark for environmental accountability in LLM development and underscores the importance of considering the full hardware lifecycle, especially as data center infrastructure scales.