Apple’s marketing deserves skepticism, but the technical argument is separate. Local inference does not eliminate giant training clusters, it mainly cuts inference costs, latency and improves privacy. Apple still uses cloud models when needed.