gpt-oss-120b training emissions
OpenAI released gpt-oss-120b on August 5, 2025 (source).
- Our estimate410–870 tCO2e to train
- About as much as96–200gasoline cars driven for a year (EPA)
How we got the range
OpenAI has not published what training gpt-oss-120b emitted. We multiplied the compute it took by what each NVIDIA H100 uses and what the grid emits, at the low and high end of each assumption. The method, in full.
| Input | Value | Source |
|---|---|---|
| Training compute | 4.94 × 10²⁴ FLOP | Epoch AI, Data on AI models, gpt-oss-120b (from OpenAI's 2.1 million H100-hours, plus 10% for post-training) |
| H100 peak (FP8, dense) | 1.98 × 10¹⁵ FLOP/s | NVIDIA, H100 specifications: 3,958 teraFLOPS FP8 with sparsity, dense is half |
| H100 rated power | 700 W | NVIDIA, H100 specifications: max TDP up to 700 W |
| Utilization, low | 0.3 share of peak | Epoch AI, estimation methods: typical language-model range 0.3 to 0.5 |
| Utilization, high | 0.5 share of peak | Epoch AI, estimation methods: typical language-model range 0.3 to 0.5 |
| PUE, low | 1.1 ratio | Google, data center efficiency: fleet quarterly PUE 1.10, trailing twelve months 1.09 (Q2 2026) |
| PUE, high | 1.4 ratio | LBNL, 2024 United States Data Center Energy Usage Report: U.S. average PUE 1.4 in 2023, p. 47 |
| U.S. grid, location-based | 384.4 gCO2e/kWh | Ember via Our World in Data, carbon intensity of electricity, United States, 2025 |
- OpenAI’s model card says the run took 2.1 million H100-hours. Our method implies 1.4 to 2.3 million, so the range brackets what OpenAI disclosed.
- OpenAI publishes no emissions or market-based electricity figure, so both ends use the U.S. grid average.
Also cited: OpenAI, gpt-oss-120b & gpt-oss-20b model card: 2.1 million H100-hours
Last checked October 17, 2026. Every model