Right, the gateway layer is the genuinely interesting part. Intercepting every outbound network call before it leaves the sandbox gives you a real enforcement surface, not just "trust the app to behave". The problem is the threat model is still inverted for the security critics in this thread: the agent is the client, so the dangerous calls are the ones going out to your authenticated services (Gmail, Slack, whatever), and a gateway that filters those is only as good as your policy definitions. One misconfigured rule and ure back to square one.
The GPU rental angle makes total sense too. This is basically Nvidia saying "don't buy Mac Mini, rent ours" wrapped in enough infrastructure glue to make it feel like a platform.