The Neural Engine deployed in the iPhone X that has since become a commodity, can train LLMs.
Think out of the box for a moment. Powerful processing doesn't have to happen on an individual device level. Apple's strategy shifted some time in the early 2010's when they found it necessary to design their own silicon, integrate machine learning into each of its chips in a neural engine, as well as low power ultra wide band chips for nearby networking and has begun deploying all of these components even in the smallest, most affordable devices that can't even use them to their fullest potential.
The more you back up and look at the greater picture, the more it's becoming apparent that Apple is building a mesh neural network. Information relevant to the individual user is processed on device for privacy, but a collective model can be processed in a mesh, first in users that know one another and perhaps share iCloud Photo albums or exchange a file through AirDrop, to eventually total strangers waiting at the bus stop together.
There's already a precedent for this where our iPhones are being used to locate other people's AirTags. There are billions of iPhones, Macs, iPads, and Apple TVs out there (and they'll no doubt make their way into Apple Watches, AirPods and Vision Pros) Entire large cities will become LLM processors.
