I'm happy to start.
Yes, I think it's absolutely feasible. As David suggested, AI is only made possible through huge amounts of computation done by very specialized chips. These chips are built essentially by one company, using machines built by one company and designs built by a handful of companies.
We've seen a lot of discussion of this compute capability in terms of export restrictions and controls, but these chips have hardware-level security. It isn't just about where the chips go but also about controls that can be on them. These chips have hardware-level security capabilities that enable verification of their level and type of use. Just as your phone can be remotely bricked if someone steals it, AI hardware can and should be configured so that it could be shut down at the hardware level. I think the more powerful the AI, the more it needs a reliable off-switch.
We can use the capabilities of the hardware, the base layer at which these AI systems are operating, both to add a layer of control and to add a layer of verification if we institute red lines that should not be crossed in their development and deployment.