I was hoping to buy a competent local model machine later this year but given the prices I’m shelving that for now. Especially because the frontier models are very cheap relative to the cost of building my own setup. Especially because AI specialised hardware and processors are improving very fast, meaning hardware we buy now will become obsolete for this use case much faster than for traditional computer use cases.

In 1-3 years the hardware crunch will be over, local distilled models will provide Opus 4.8 like intelligence, and the hardware will exist to provide usable performance.