Google do have an inference advantage with TPUs.

Everyone else needs to pay nvidia margins.

Training is murkier as it’s more about the total performance and scalability of the system.