With all the hype, attention and money going into AI specific clusters - HPC for science needs to reinvent itself to have a seat at the architectural trade off table.

Code ninjas who speak the language of perf create the biggest moat that the HPC community has at this point.

The cost of software co-design can be expensive to the academic/govt partner if the architecture gets canned by the vendor.

I agree, vendors can benefit from that.

how I wish that this 👆 was just sarcasm, this viewpoint is hyper prevalent in academia.