It’s easy to defend this position.

It is safer to operate an AI in a centralized service, because if you discover dangerous capabilities you can turn it off or mitigate them.

If you open-weight the model, if dangerous capabilities are later discovered there is no way to put the genie back in the bottle; the weights are out there, anyone can use them.

This of course applies to both mundane harms (eg generating deepfake porn of famous people) or existential risks (eg power-seeking behavior).