GPGPU across the product line with a stable programming interface isn't contingent on LLM success.

It was a bet that the future would need more streaming parallel compute, and GPUs were well-placed to deliver it.

That's a shrewd bet, and one AMD (and Intel) were stupid not to also make.