Don't have specific sources since this is my napkin math. But Deepseek R1 is at about 2J per token [0], Luna is almost certainly bigger than R1 based on people's gossip in ML circles in SF, but also much more efficient per token, and served at ideal batch size etc, so I assumed it's around 1J/decode token.
From there, I estimated 80% cached, 2% decode, and the rest pre fill from tool calls inputs. That gave me 50MWh total, probably conservative since this kind of project will have very long context length on every pass. then take average energy mix in the US. It's definitely wrong by some factor in either direction, but I would personally feel the same about this experiment if it was 9 tons of CO2 or 36t.
From there, I estimated 80% cached, 2% decode, and the rest pre fill from tool calls inputs. That gave me 50MWh total, probably conservative since this kind of project will have very long context length on every pass. then take average energy mix in the US. It's definitely wrong by some factor in either direction, but I would personally feel the same about this experiment if it was 9 tons of CO2 or 36t.
[0]: https://ml.energy/leaderboard/