Comment by dhruvdh

And yet Meta is using MI300X exclusively for all live inference on Llama 405B.

Clearly there are workloads AMD wins at, and just going Nvidia by default for everything without considering AMD is suboptimal.