If this is true, the hyperscalers are toast

In my regular research (behind a paywall), I have been saying for a while that I think the future of AI is not large language models (LLM), but small language models (SLM) run on local desktop compute...