They’re leaving out two big pieces of the story:

1. Apple is switching to CXMT, whose IPO just made them China’s most valuable company.
2. Open weight models are putting a damper on centralized frontier models, making it look like efficient local inference rather than next-gen training may be the winning investment. And what do you need for efficient local inference? Why, a heterogenous SoC architecture with unified memory. And what has Apple bet their entire product line on, for the past 6 years? Exactly that!