If your solution to the externalities of LLM-use is "open models run on local hardware," then you need to be thinking hard about how that works in a scenario where the cost of RAM continues to climb in tandem with the amount of memory required to run successive models effectively, because there's every indication that's how the industry wants it.