The tok/s stat is interesting. Since the dominant constraint on inference speed is hardware, it suggests X purchased far more compute than was really needed to serve the demand for their models.

Expensive miscalculation.