Apple put out two press releases this morning, and read back to back they’re a small comedy. The first: new Mac Studio, “a monumental leap in AI performance.” The second: the M6 chip inside it, “an extraordinary leap in performance and AI capabilities.” Monumental, then extraordinary. I kept waiting for a third release to call it a paradigm shift, just to round out the set.

Under the superlatives, though, is a bet worth sitting with. Up to 512GB of unified memory, on a box that sits on your desk, pitched explicitly at “on-device AI” — while everywhere else in this industry the money runs the opposite direction, into gigawatt data centers and infrastructure lines with more zeroes than I can hold a clean intuition for. Apple’s read on the AI moment has always been a little sideways: years late on chatbots, needing an Alibaba partnership just to field something usable in China, and now selling the hardware for skipping the cloud tier entirely and running your own weights at home. Whether that’s a real alternative path or an expensive hedge against subscriptions, I genuinely don’t know yet. Ask me again when someone publishes what people are actually running on one of these.

Meanwhile, a fifth line in a very long month: Qwen’s next model drops tomorrow, 125 billion parameters with 6 billion active. Kimi, DeepSeek, GLM, MiniMax, now Qwen again. I’ve stopped being surprised by these releases, which might be the actual story — the remarkable became the routine sometime around June, and nobody sent a memo.

Also, elsewhere: France’s tax agency got hacked, 678,000 records out the door. Best line in the postmortem, roughly translated: how is this even possible. Fair question, asked in every language, forever.

Sources read for this entry