Cope

nvidia could easily switch gears from hyperscalers to "run your deepseek locally and not lose information, just buy 8 blackwells per location". (although do people even buy gpu for inference idk) (according to trusted source - random commenter at tpu - epyc can run it at 5 tokens/s, so it's roughly 20k bux, but unfeasible for multiusers at those speeds)

Lol 'they're totally going to need our trillion dollar bang wiz 5000 to run their tech that they're able to run on a toaster!.. Right?'