Comment by seydor

even if their chips weren't as good, what stops a company from training a large model for a very long time in less capable hardware? Is there a way to overcome memory limitations somehow?