I'm building a local LLM server myself. For about 8k€ you can have 170tps, ~6000 prefill.

A lot more with Haiku equivalents (Qwen3.8 27B) or gemini. It's not frontier, but not that far away.

MacBooks with large RAM and optimized for local models should be Apple's medium term goal.