You'd still need those giant data centers for training new frontier models. These Taalas chips, if they work, seem to do the job of inference well, but training will still require general purpose GPU compute