Jose Romero
← All letters
Prefer email? Read this on Substack

I built a decision tree instead of buying a Mac Studio

The new Mac Studio with the M5 Ultra is out, and I am torn. It is a much better machine than what I run everything on today. The 512 GB tier is tempting for the biggest local models. And it would be a dumb financial decision. I catch myself opening Apple's configurator, building the dream config, and closing the tab before the reality check finishes. So this time I did the reality check on camera, and then I built a tool so the next person can do it in two minutes.

The config I keep building

M5 Ultra, 80-core GPU, 256 GB, 2 TB: $11,299. That is more than my daily driver. The 512 GB tier is coming late October with no price; at Apple's flat $25 per gigabyte it lands somewhere around $17,000. One terabyte of storage on a machine this price is sad, but that is Apple.

Why the pull is real: unified memory. Getting to 256 GB any other way means stacking consumer GPUs, and the 5090s are about $5,000 each right now. Four DGX Sparks get you to 512 GB for about the same money as one Mac, with less bandwidth than the Mac advertises. It is an apples to oranges comparison, because CUDA gets priority for everything, and if you train or serve models the Sparks and the NVIDIA cards have obvious advantages. But you get more when you buy a Studio: the ecosystem, the accelerators, and MLX works well in a lot of scenarios now.

One more data point. The 512 GB and 256 GB M3 Ultras got pulled from Apple's store, and as of September 12 they were selling on eBay between $16,000 and $20,000. There is something to be said for buying at launch and being on the right side of that curve. That is speculation, and it depends entirely on how Apple prices the new tier.

The box I actually run

Everything you have seen on this channel runs on a GMKtec EVO-X2: Ryzen AI Max+ 395, 128 GB of unified memory. I paid about $2,000 a year ago, before the RAM prices went crazy. It runs CachyOS, I game on it, and every benchmark video came off it. A 192 GB version of the same platform is about to ship, and if you are buying today I would look at the EVO-X3 over the X2: same chip, about $100 more, plus an OCuLink port for an external GPU. Both are in the gear list below.

Against that box, the M5 Ultra is roughly a 5.5x speed improvement on the same benchmarks I run. That is a real productivity jump for local AI work. It is also $17,000, and you know what you can do with $17,000.

So I built a tree

I told the AI what I was debating, asked for the numbers, and had it draw a flow chart in Excalidraw. Then I hosted it as an interactive page. It asks the questions in order: money first (cash, Apple's 0% plan, or a lease), then whether you actually want better models than your machine runs today, whether the machine you are looking at can run the model you want, whether you would cluster more than one, whether you will train or fine-tune and whether MLX is enough for that, speed, use, the state of the world, and timing. Every path ends in buy at launch, wait, avoid, or a CUDA route with the boxes that hold your model and what they cost in electricity.

The page is a snapshot. Prices, model file sizes, and speed references were pulled on September 16, 2026, and the 512 GB tier is unpriced until late October. It says so at the top, and it will go stale.

My walk

Can I pay cash without touching emergency savings? No. Can I put it on Apple's 0% plan and pay it off without stress? About $1,400 a month; for the sake of the video, yes. Do I want better models than my box runs? Yes. Can the 512 GB machine run GLM-5.3-Flash at Q8, at a quality I would accept? Yes, at an estimated 20 to 30 tokens a second. Will I train on it? No. Would I trade memory for CUDA? No, the M5 Ultra without the price attached is a sweet deal. Has anyone measured the M5 Ultra? No; units ship September 22. Am I okay buying an unmeasured chip with a 14-day return as the only hedge?

No. And that is where it stops: wait for numbers.

I kept going on camera to see how far the tree makes you go to reach a buy. The honest answers after that point were mixed: it does replace things I pay for, it does not pay for itself in three years, it is worth it as a tool beyond local AI, I do not need frontier-model quality for my hobby work, hosted limits have bitten me, the 512 GB price is unknown, I do not expect memory prices to fall within a year, and waiting six months costs me very little because the box I have is running the channel fine.

Where I land

Wait. Not because the machine is bad; it is the most interesting desktop for local AI under $20,000. Because two things I need are missing: a measured number and a price. Both arrive within six weeks. Am I willing to spend an unknown amount of money on an unmeasured chip? The short answer is no.

If you are into local AI and looking at this machine, run your own answers through the tree and tell me where you landed. If you bought one, which config and what runs on it. If you skipped it, what you bought instead.

Gear in this one

Some links on this page are affiliate links. If you buy through them I may earn a small commission at no extra cost to you. It helps support the work. As an Amazon Associate I earn from qualifying purchases.

Sources

Get the letters

Every post goes out free on Substack. What I build, break, and fix, written up in plain English. No spam, unsubscribe anytime.

Subscribe on Substack