With no warning, Apple just released new Mac mini and Mac Studio models, both targeted squarely at AI inferencing.
Three weeks ago, in "Four Watersheds", I noted that Apple had quietly withdrawn its 256GB and 512GB Mac Studio configurations - retreating, one memory option at a time, from precisely the product that would have owned this moment. I asked then whether Cupertino understood the opportunity. Apparently someone did. The door Apple closed reopened this week, with AI inference printed on the invitation.
As I explored in "Apple's Difficult Calculus", Apple has to balance its need for RAM for the upcoming iPhone 18 Pro models - shipping next month - against these desktop workstations. But let's be clear about the scale: Apple shipped only around a million of these two desktop models worldwide last year, compared with a quarter billion iPhones. Unless desktop sales increased ten-fold in a single year - an unprecedented jump - they would not significantly affect RAM availability for iPhones. As it stands, they're little more than a rounding error on Apple's annual RAM purchasing.
Yet that rounding error offers exactly the kind of compute that enables both the business and home watersheds.
At AUD $17,500, a Mac Studio with an M5 Ultra and 256GB of RAM is an ideal platform for serving powerful models like GLM 5.3 Flash (aka 'Ox Alpha') and Qwen3.8-Flash-Next, with enough headroom to run a department's worth of agents.
That matters little today, but the way we use AI today is not how we'll use it within a year. Right now we have a chatbot, or a single agent; only rarely does either spawn subagents to help complete a task. As confidence grows, the agents multiply - and we'll think nothing of half a dozen of them working for us constantly, supporting everything we do.
That kind of capability is what every business is going to need, and it's precisely what Apple now supplies at the high end. The two fit together perfectly. No longer consigned to high-end media editing, the brand can become the one that provides a just-works brain for the office. Silence, reliability, nothing to configure: the product requirement for an office intelligence you buy like furniture. Apple wrote the box's marketing a decade before the box existed.
Is the cost justified? Run the arithmetic. Heavy agentic users are already spending on the order of $1,000 a month in tokens. At that rate, the Studio pays for itself in about eighteen months for a single such user - and for a department with a handful of them, in about a quarter. After payback, the marginal token costs electricity. Just as important for the person who signs: a one-off capital expenditure replaces an ongoing, variable, hard-to-forecast monthly bill. Not every business consumes like that today. Enough do - and the shape of work is headed that way - to keep the M5 Ultra in demand for the foreseeable future.
At AUD $5,000, the new M5 Pro Mac mini with 64GB serves the home watershed: enough RAM to comfortably run the incredibly capable Qwen3.8-27B alongside a clutch of subagents, with headroom left over to be a very good home computer. And the Mac's native virtues turn out to be exactly what unattended overnight work demands: quiet, thermally sane, sipping power, happy to run for months. Nobody at Apple will say it this way, but the new mini is the first computer sold primarily for the hours its owner is asleep.
If the smaller models keep improving - and there's no reason to expect they'll slow - this configuration is all the compute a household needs for its automations and cognition. The business case for the home watershed is not yet as well articulated as the one for the office; this is the realm of the hobbyist, and plenty will buy simply to have high-quality AI at home, even before they're sure what they'll do with it.
I'm sure they'll come up with something. What that something looks like will echo a well-established precedent from a century ago.
That's the topic of my next post.