As hardware improves, local AI models will be much more common. Specific tasks, at TFA points out, don't need powerful LLMs. Local models that are much more task specific and can have defined rules can met the needs of many users; with an added benefit of keeping data local and private. Local AI isn't as fast or powerful, but will suffice for many tasks at a fraction of the cost. Of course, that means the days of free model d/ls may be numbered, since that would mean giving away the IP and reducing the use of paid cloud models; I suspect d/led models will at some point be history and replaced with subscriptions.