Apple announced the M6 and M5 Ultra on August 25, pitching both parts on "a big leap in performance and AI compute" — the second phrase doing most of the work in the headline.
The pairing is the notable part. M5 Ultra caps the current generation, presumably as the top-end desktop part, while M6 starts the next one; Apple is now shipping the tail of one silicon generation and the head of the next in the same announcement window.
At the time of writing the only source is Apple's own newsroom post. No independent benchmarks, no pricing analysis, no memory-bandwidth teardown — and on-device inference throughput on Apple silicon has historically been bounded by memory bandwidth far more than by raw neural-engine TOPS. Treat the "AI compute" claim as unverified until someone runs real local-inference workloads on both parts.
Sources: Apple Newsroom, HN discussion