Each new version is trained on far more compute than the last, so a release schedule is also a datacenter and chip order schedule. Usage numbers matter for the same reason: running the model costs GPUs and power every day, not just once at training time.