alt Hacker NewsTheir point is (and I tend to agree), that LLM AI is not different at the fundamental level. The idea is that e.g. coding capability will be a standard feature of future computing hardware, not a very expensive service you rent. Anthropic is the current-gen mainframe operator selling time-shared access. In the future, that capability, too, will just be a commodity, i.e. a part of the silicon in your notebook or smartphone or whatever will replace those.
Yes, to a large extent, for current real world mass market use cases at least, the only thing keeping the biggest players relevant is that they (and their pals) have monopolised access to GPUs by buying them all and making deals to buy almost everything that Nvidia can produce (even more than they can plug in).
If SMEs were able to buy and host their own GPUs, they would not hesitate to avoid paying a subscription for access to AI tools. They're already starting to do this.
Models already are a commodity, the massive bump in access prices to profitable levels is not going to happen, because businesses will not pay it.