Training an AI model is the easy part. Building it is the expensive one.
Researchers at LORIA's Multispeech team used activity logs from the Grid5000 shared computing platform to measure the energy cost of developing four audio deep-learning projects, not just training the final models. They found the development phase - architecture prototyping, failed experiments, hyperparameter sweeps - consumed 3 to 256 times more energy than training the single best-performing model that came out of it. That gap held across all four projects studied. The team built this measurement approach specifically because most energy accounting for AI stops at the training run and ignores the months of trial and error before it.
Every corporate carbon disclosure about AI tends to cite the cost of the final training run, which this study suggests is the smallest number in the pipeline. If development actually consumes most of the energy, a company reporting "our model cost X to train" could be undercounting its real footprint by two orders of magnitude. That makes a lot of voluntary AI energy disclosures look more like marketing copy than accounting.
It is one case study from one lab's compute cluster, not an industry-wide audit - but if that ratio holds elsewhere, the sustainability conversation has been measuring the wrong stage of the process.