My new theory is the Nvidia driver can't be GPL and in the Linux kernel, because then they couldn't ban datacenter usage of their GeForce cards by not licensing the driver for datacenter use. The upcharge on the Tesla series of cards is huge compared to GeForce for mostly the same chips. (For those not aware, see if you can find and GTX 2080 or 3080 from a cloud provider. It's not a thing. This is actually a huge deal for the machine learning industry, massively increasing costs. I doubt Google would have made the TPU if not for this.)
Also, their driver is very complex, and they are constantly improving their hardware. They don't want to be dependent on getting new features and performance improvements upstreamed.
Also, their driver is very complex, and they are constantly improving their hardware. They don't want to be dependent on getting new features and performance improvements upstreamed.