I disagree in this context, LLMs raise the lower bound and diminish the relative advantage. Consider the introduction of firearms into feudal Japan, the lower bound is raised such that an untrained person has a much higher chance of prevailing against a Samurai than if both sides fought with swords. Sure the Samurai could afford better guns and spend more time training with them, but none of that would allow them to maintain the relative advantage they once had.

This only holds true for local inference and open source models. LLMs are not truly ours today: comparing a firearm which is totally yours (we can argue about bullets etc, which have a (still low) production barrier) to a big-tech-mega-datacenter-in-texas-run LLM is naïve.

I fail to see why needing to be able to train your own LLMs is any sort of prerequisite. I already made the distinction between different qualities of guns, a lower quality gun is immensely more effective than no gun at all.

No but there's an advantage against small and midsized corps