That's just not enough dollar volume to pay for all the inference, data center buildout, training, salary costs to compete with Meta etc etc. They also need to pay for the inference of the 98% of users who don't pay anything but use the models, that's essentially a marketing cost.