Hosting the model is cheaper per token, the more batched token you get. So they have big advantage here.