It’s still a lot of extrapolation but better than before. Researchers studied the energy usage of open-source LLMs, image and video generators. Then, based off this Microsoft paper and interviews with other experts, doubling the GPU’s energy usage roughly accounts for the energy usage of everything else involved. An interesting finding is that stable diffusion image generation is often less energy-intensive than prompting high-parameter LLMs (around five seconds of microwaving).

But there are still so many variables like what kind of model, the amount of parameters, the complexity of a prompt, what data center your prompt goes to and what the energy mix in its area is like… The researchers still make a point to emphasise how opaque companies are about their closed models.