The M5 Ultra Mac Studio will be exactly this. Which is why they will cost $30K.
Apple will sell every one they can make, and new-in-the-box models will go for $50K on eBay.
While I completely agree that those machines will sell for $30-50k new, I don't understand why anyone would ever buy one at those prices.
Let's assume that by canceling your cloud LLM subscriptions, you are saving $500/month. That's just a round number to represent the type of subscription prices out there: OpenAI and Anthropic are up to $200/month, SuperGrok is $300/month, and Google's AI Ultra plan is $200/month. If you have 2 plans that's $400/month. So just for the sake of argument let's say $500.
In that case, if you buy a $30k machine tax-free, it will take
60 months -- 5 years -- for the Mac to get out of the investment hole and start returning monetary value back to you compared to your $500 of LLM subscriptions.
I can think of much better ways to invest $30k to get a bigger payback in 5 years. And you also have to consider that you don't have access to that money during the 5 years after you spent it. Many investment opportunities exist where you can pull your money out any time you need it.
Meanwhile, even if the M5 Ultra Mac Studio has 1 TB of RAM, that won't be very much compared to the demands of the frontier models in 5 years. It'll be a lot, but not anywhere close to what the state of the art is capable of using. We should see state of the art models demanding in the 2 to 5 TB range in 5 years. Maybe even more.
So you are locking yourself into 2026 hardware for a huge up-front cost, and the only way you ever see any financial benefit is if you keep the machine longer than 5 years.
Again - I know people will buy this. But tell me how this makes sense.
Edit: I also realized that you will have a difficult time running multiple models in parallel with your purchase of a single 1 TB Mac Studio. Even if you bought
two of these beasts, you're looking at just under 2 TB of usable VRAM. If each model takes, say, 800 GB, that means you can run two in parallel.
Meanwhile, a real workflow in Claude Code or Codex can use numerous LLMs in parallel. For some of the more demanding complicated knowledge work and coding, you might benefit from 5-10 LLMs running in separate contexts at the same time. This is not only to save time, but to enable agentic workflows like in Ultracode mode.
Those agentic workflows are really smart because they have parallel researcher LLMs, implementation LLMs, and verifier LLMs that form a directed graph of LLMs working a "flow chart" of orderly and utterly methodical implementation until everything is working 100%. With a huge $60k investment in two 1 TB Macs, you're still limited to a fraction of the productivity of someone paying $200/month for Claude Code (or $400 for two subscriptions).
It really doesn't make financial sense at all to me.