The moat is building GPU infrastructure to serve inference at scale.

Why? Almost all of the open source models end up actually sucking for general purpose use, and require a decent bit of effort to even make remotely usable for specialized use cases.