Open Source?

Much of the user-space code is proprietary and closed and always has been (libnvidia-ml, libcuda, libcudart, etc). There are no signs that Nvidia will be willing to even consider opening those.

The open source kernel module has most of the same code that the proprietary module has, except some components that manage and control low-level GPU behavior. Nvidia has moved that low-level management code to a binary blob that is loaded onto the GPU itself and run on a RISC-V management co-processor called the GSP. Along with another co-processor called the FSP, it performs similar functionality to the Intel CSME and AMD PSP. By moving the low-level management code out of the drivers, they're able to open up their drivers, and thus ease code maintenance and bug reports, without revealing anything new about the GPU's inner workings.

This is also why the open driver, despite having feature parity with the proprietary one, only lacks the ability to disable the GSP (the register that disables it is just a no-op on the open driver). It's also why there is no open driver for the Kepler architecture, which lacks a GSP in the first place.