It's unclear why this drew downvotes, but to reiterate, the comment merely highlights historical facts about the CUDA moat and deliberately refrains from assertions about NVDA's long-term prospects or that the CUDA moat is unbreachable.

With mature models and minimal CUDA dependencies, migration can be justified, but this does not describe most of the LLM inference market today nor in the past.