AI & ML · Jun 13
Apple's Most Capable Cloud AI Runs on Google's Servers. Apple Is Fine With That.
At WWDC26, Apple announced five new foundation models. One of them lives on Nvidia GPUs in Google's infrastructure, and understanding why tells you a lot about where on-device AI actually hits its ceiling.
- Apple's five-model AFM lineup routes requests across on-device, Private Cloud Compute, and Google Cloud tiers based on task complexity, a tiered inference pattern worth studying for any multi-environment AI deployment.
- AFM 3 Core Advanced packs 20 billion parameters into on-device inference by activating only 1 to 4 billion at a time; the sparsity is the engineering story, not the headline parameter count.
- Even Apple's privacy-first architecture depends on third-party cloud infrastructure at its capability ceiling, a practical reminder that vertical integration has limits when frontier-level inference is the target.