I think you are overestimating power usage by order of magnitude or two. For a much bigger Deepseek model, you can get <0.1J/token[1], or 36M output token for 1 kWH.
Or more than 500 book worth of token for 1 hot shower.
[1]: https://inferencex.semianalysis.com/infx/inference/deepseek-...