So, my point isn't that every model has to run locally, but if what you use is pay-as-you-go, and you only use pay-as-you-go for the long-horizon work we talked about, in the end you're still not necessarily spending more.