New Studio M5 Ultra ready for local home AI lab? I compared it to OpenRouter prices. 512GB will be available from October, and I guess that the price will be around 15-17k. Let's see how much it will cost to run a proper openweight model at home: DeepSeek V4 Pro: 2 Macs, $32–34k Kimi K3 and GLM-5.3: 4 Macs, $64–69k At 100 output tokens/sec, 24/7, compared with OpenRouter prices today, the hardware payback would be: Kimi K3: 16–18 months DeepSeek V4 Pro: 36–39 months GLM-5.3: 56–60 months K3 makes sense for a privacy first fully local setup.
The 512GB M5 Ultra will ship in October with a massive 1.2TB/s memory bandwidth. 50% more memory bandwidth than M3 Ultra, 4.4x more than DGX Spark and M4 Pro. Stacking 4 x M5 Ultra, I expect you'll be able to run Kimi K3 / GLM 5.3 faster than the API (>100 tok/sec).