That's an interesting take. I had a similar thought about safety and how AI providers propagate that using cloud options is safer because hosting a server is open to attack vectors, and who would manage that best if not the big providers themselves.

The fact that Hugging Face was attacked not by a small amount of compute but such a vast amount of computation. It was only possible because its already so centralized by Open AI.

It feels a bit like cyberpunk where only the big companies in reality will have the resource to pull of such an attack.