Amazon's Trainium2 AI Accelerator Features 96 GB of HBM, Quadruples Training Performance

It seems to me like this approach to custom-built ML hardware which can only be leased as a service, not bought outright (such as this or Google's TPUs), is not very promising for research. Sure, this might be a cost-effective way for companies to deploy LLMs and maybe AWS can squeeze out a few interesting papers but since nobody else can do this without paying Amazon obscene amounts of money I don't see this leading to the next great innovation in the field. It's at least interesting that Google's TPUs barely had an impact, I'm curious to see if this will also be true for Amazon.