On 25 August, Apple opened pre-orders for a new Mac mini and a new Mac Studio, both shipping from 22 September. Read past the keynote language and the announcement is really about one thing: how much AI a machine on a desk can now run.
The numbers make the point on their own. The Mac Studio's new M5 Ultra chip pairs an 80-core GPU with up to 512GB of unified memory and 1.2TB/s of memory bandwidth; Apple's own framing is that it runs models "with hundreds of billions of parameters entirely on device". One tier down, a Mac mini with the M5 Pro takes 64GB of unified memory, comfortable territory for the mid-size open-weight models that handle real document work, in a box the size of a lunchbox.
Why this matters more than a spec sheet
When we compared cloud and local AI under the AI Act, we argued that architecture decides one column of your risk table: where client data lives. The standing objection to the local answer has always been economic. On-premise AI sounded like data-center thinking: racks, cooling, six figures, a hallway you do not have.
That objection is aging fast. In the US, the new Mac mini starts at $899 and the Mac Studio at $2,499, with the M5 Ultra version at $5,499. The hardware that runs serious AI models now costs what a firm already spends on a partner's workstation, is bought in a store rather than procured as a project, and runs with no per-token meter in the background.
There is a second signal here, easy to miss. Apple is not positioning these machines for hobbyists; its release notes talk about running "frontier AI models on device" and about "secure and private agentic tasks". When Apple builds its entire desktop line around the premise that AI should run where the data sits, local AI is no longer the contrarian option. It is a product category.
The part that concerns us
This is the trend Granid is built on. Legal Intelligence runs on Apple Silicon, on exactly this class of machine: a Mac in your office, running models over your documents, with nothing leaving the building. Announcements like this one do the part of the work we cannot do, making the hardware cheaper and more capable every year. Our job is everything the chip does not do: the models, the Swiss legal corpus, the verified citations, maintained as a product for a flat licence.