What this is
Apple this week released the M5 Ultra with 1.2TB/s memory bandwidth—that's 1.2 trillion bytes per second. We've noted this is the first time running large models locally (AI running on your own machine, not the cloud) has crossed the practical threshold.
Bandwidth sounds abstract to most people, but for anyone running LLMs it's the make-or-break spec: every generated token (the unit AI generates piece by piece) requires reading hundreds of GB of model parameters from memory. Without enough bandwidth, output crawls one token at a time. At 1.2TB/s, 70-billion-parameter models can run smoothly.
Technical detail: Reddit users speculate the M5 Ultra still uses LPDDR5X (current mainstream laptop memory) rather than next-gen DDR6. If true, it means Apple is winning on memory controller and packaging rather than raw hardware upgrade—a signature Apple move.
Industry view
Optimism centers on "on-device AI" (running AI locally, no cloud required) finally landing. Over the past two years, vendors have shouted "AI PC," but the only laptops actually capable of running 100B+ parameter models have nearly all been Apple's. The 1.2TB/s figure turns "replace cloud APIs with a Mac" from a geek toy into a viable option.
But sober voices flag three caveats. First, the training market remains Nvidia's monopoly—H100/B200 deliver 5–10× the bandwidth of M5 Ultra, and Apple can't bite into the training side of the cake. Second, a Mac that fully utilizes 1.2TB/s will almost certainly start above ¥40,000 (~$5,500), so "affordable for regular people" is still a stretch. Third, the AI software ecosystem still runs on Nvidia's CUDA (the industry's dominant GPU programming framework); whether developers are willing to rewrite for Apple's Metal framework determines how far this road extends.
Impact on regular people
For enterprise IT: On-premise AI deployment becomes a real option again for data-sensitive industries—finance, healthcare, and law firms can run models on their own machines, with data never leaving the building.
For individual professionals: Over the next year, knowledge workers may start running daily AI processing of documents, meeting notes, and client materials on their MacBooks rather than uploading to ChatGPT or enterprise cloud.
For the consumer market: The high-end Mac's "productivity tool" positioning is being rewritten—more white-collar workers will see the Mac as the only device that runs AI locally, not just video editors' choice.