Comment by edmundsauto

My guess is the subscription plans provide a way of controlling the base load better, because of how the "session duration" and token quota per session and week work. People can get tremendous value, but a lot of that value is only available at night (theoretically when their GPUs are most likely to have less contention). So it's kind of a pricing strategy around their excess capacity. (For comparison, when first announced, AWS spot instances were like 10x cheaper than reserved.)