In case anyone haven't seen it yet, here is the experience of running a local model (Gemma 4 26B A4B QAT) using LM Studio with a web search MCP.
That model is 15.6 GB on the disk and pushes the limits of a M4 Pro with 24Gb RAM. In those cases, the best option is to tweak: primarily by loading the model with a reduced context size.
I think it is more likely we will have a period of high energy costs that will lead to AI services becoming less affordable and/or being throttled or blocked by government intervention.