This elides whatever resources were used in training/development. Which in the cases of ML models is quite often not minimal. Even the diy things you can train yourself make a significant dent. And there’s likely to be an ongoing cost of this too, because of updates