Google and Amazon still have to buy tons of Nvidia HW to provide in their clouds. No one writes to their custom chips besides internal teams because the software stack doesn't exist.

PTX assembly. Deepseek used some of it to do a little bit of work that CUDA didn't have APIs for.

Sadly platform specific

Distant cousins, and they had never met until they were both CEOs. This is a stupid trope and people need to stop using it.

The long history of both companies doing a half-assed job of entering the others markets speaks differently.

Every 'run cuda on AMD' project gets bought then cancelled. Every cuda alternative is understaffed and cancelled as soon as it starts to gain any traction, etc.

To me it looks like a clear attempt to not hurt each others businesses whilst being deny-able if anyone accuses them of unfair market practices/collusion.