The Hidden Cost of Thinking: Energy Use and Environmental Impact of LMs Beyond Pretraining
Jacob Morrison, Noah A. Smith, Emma Strubell · 2026
Read The Hidden Cost of Thinking: Energy Use and Environmental Impact of LMs Beyond Pretraining by Jacob Morrison, Noah A. Smith, Emma Strubell (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-01158.txt.
Description
Jacob Morrison and two co-authors provide a detailed breakdown of the environmental cost of a full language model development pipeline, beyond pretraining alone. Green reporting usually focuses on training one final model. Their case covers it through supervised fine-tuning, preference optimization, and reinforcement learning for the Olmo 3 family, in instruction-following and reasoning variants at two sizes.
Those models are far more expensive to post-train than tuned counterparts, driven by reinforcement learning rollout generation. Project work including experiments, failed runs, and ablations accounts for the large majority of total compute. Total estimated energy, emissions, and water use are reported, with water driven by power production rather than cooling. The authors call for fuller disclosure standards.
Catalog record
| CREATOR | Jacob Morrison, Noah A. Smith, Emma Strubell |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 34 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.01158