Never - data centers will always offer more power if you only care about raw inference speed. HOWEVER I think that we'll reach the 'good enough' bar super soon. In 2-3 years I expect apple macs to be able to run a model as 'good' as Claude 4.6 sonnet at 90% of the inference speed we're used to from a cloud API.

Yes, I'm sure by then there will be better models on offer via cloud providers, but idk if I'll even care. I'm not doing science / research or complex mathematical proofs, I just want a model good enough to vibe code personal projects for fun. So I think at that point I'll stop being a OpenAI / Anthropic customer.

The answer is "whenever Apple decides to sign CUDA drivers" at this point.