I would also add AMD's investment and lead in chiplet technology for its GPUs.
Chiplets have higher yield and lower manufacturing cost, better heat dissipation, more flexibility and scalability and if designed well, doesn't reduce connectivity.
MI300X is chiplet based while H100 and H200 is still monolithic. Nvidia is working on Blackwell, which is chiplets but not sure when they can productionize it.
The two Seeking Alpha comments that try to calculate MI300X production based on TSMC CoWoS capacity:
Original: https://seekingalpha.com/article/4640182-ai-gpu-market-disruptor-how-amd-mi300-change-game#comment-96340677
Update: https://seekingalpha.com/news/4044031-amd-takes-wraps-off-nvidia-competitor-updates-ai-forecast?source=content_type%3Areact%7Cfirst_level_url%3Aauthor%7Cbutton%3Aview_comment#comment-96742671
TLDR: base case is 400k MI300X's produced but possibly higher.
I would also add AMD's investment and lead in chiplet technology for its GPUs.
Chiplets have higher yield and lower manufacturing cost, better heat dissipation, more flexibility and scalability and if designed well, doesn't reduce connectivity.
MI300X is chiplet based while H100 and H200 is still monolithic. Nvidia is working on Blackwell, which is chiplets but not sure when they can productionize it.