If my calculations are right, and Deepseek is being truthful about its model's energy usage being only about 2% that of ChatGPT's, one Deepseek #
AI response takes about the same energy, give or take, as a 60W light bulb being on for 3 seconds.
I don't find this entirely unbelievable. Personal computer energy is often greatly overestimated; I remember my chemistry prof having us do the calculations as an exercise in class to find that basically leaving your desktop computer on all the time costs next to nothing. (Your refrigerator, on the other hand...)
These huge datacenters, especially for generative AI, are still a massive problem. Apparently, Deepseek's efficiency is mainly because of its Mixture of Experts model that allows for about 90% of the model's total parameters to be entirely inactive for any given prompt.
If these numbers are correct, then a Deepseek query is actually 60-70% more efficient than a Google search. So I do think that LLMs may have a sustainable future in terms of information seeking and critique; I'm still admittedly very impressed with how useful this model has been as an 'audience member', but generative AI still has a long way to go — and I'm not at all convinced that we
should go. Deepseek has released its own generative AI model, Janus, but it's self-admittedly largely the same in terms of energy usage as its western counterparts.
