This is a matter of consequence, and the arithmetic fits on a napkin.
I pay about $1,000 a month for access to AI, and I still ration myself. For scale — the top Claude plan alone is $200 a month, for 20x the Pro rate limits, and a few subscriptions like that get to a thousand quickly.
Or I can spend $40,000 once — four Mac Studios — and have all the AI I want, in complete privacy, with no limits, until the machines die years from now.
The four aren’t arbitrary. Apple’s current Mac Studio takes up to 512GB of unified memory, “the most unified memory ever in a personal computer,” and Apple says the machine is “capable of running large language models (LLMs) with over 600 billion parameters entirely in memory.” That configuration isn’t cheap: the 512GB upgrade adds $5,500 and brings the price to $9,499 before storage. Four of them, with some room for drives, is about $40,000.
It’s hard to read the 600-billion figure without thinking of DeepSeek-R1, which is 671 billion parameters in total, 37 billion active per token, with the weights released under the MIT license. The Register’s judgment on the day the Mac Studio launched was that “running something like DeepSeek R1 at home is totally feasible, at 4-bit precision at least.” A very large model, free to download, on a machine that sits under a desk.
So the choice is about $12,000 a year for access, metered by somebody else’s limits — or $40,000 once, for hardware I’d own outright. A little over three years of subscriptions pays for the machines, and the machines keep running after that.
The other side of the ledger is real — the machines need tending, and the best closed models will likely stay a step ahead of the open ones. None of that changes the shape of the sum.
And the part that matters most: the machines take advantage of every new open-source model delivered in the years ahead.
