I Was Going to Buy the 512GB Mac Studio M5 Ultra. Not Anymore!

I've written in several previous posts that when Apple released the Mac Studio M5 Ultra with 512GB of unified memory, I would upgrade from my current Mac Studio M3 Ultra with 256GB. I wrote that I wasn't going to make the same mistake again, buying a Mac with too "little" memory and regretting it later.

Well, Apple finally made the 512GB version available yesterday, and I've changed my mind.

The reason? 149,500 SEK, Yes just for the extra jump from 256gb mem to 512gb of mem :)

Not for the entire computer, just those additional 256GB. I knew it would be expensive, but I didn't expect quite that much. Apple has a gold star in pricing, I was expected it to cost more than reasonable somewhere around half of what they are asking today.

I don't mind spending money on computers, camera gear, or whatever makes the job easier or the experience better. Since I was around 20, I've had this mindset that I call "Work Hard, Play Harder."

What I mean by that is that I work a lot. I'm often up around 04:15 in the morning, going straight to my desk, and I usually work until it's time for bed, or we round off the evening with a movie. Weekends? Pretty much the same.

So when I finally get some time to "play", I want the best experience I can get. I don't want to spend that time being annoyed by equipment that isn't quite good enough.

Take photography, for example. I could buy a mid-range camera for a fraction of what a flagship costs, and it would probably do a perfectly good job most of the time. But then there's that one moment when the autofocus misses a shot because it couldn't quite keep up. A moment I might have travelled halfway around the world to photograph, and one that might never happen again.

Of course, even the most expensive camera can miss a shot (Hi Canon R1 in Pantanal overheating at the moment of an jaguar attacking a caiman), but I'd rather know that I've given myself the best possible chance of getting it.

The same goes for computers. I spend an enormous amount of time in front of one, and I don't want to sit there waiting for things to happen because I saved some money on a slower machine.

That's what "Work Hard, Play Harder" has always meant to me. Work as hard as you can, and when there's finally time to play, don't compromise unnecessarily. At least not if I can afford to avoid it.

Back to the Mac Studio and 512gb of mem. Anyone who has followed my recent adventures with local AI knows that. I've bought four GB10 Sparks over the past few months (when they where "cheap"), and my Mac Studio M3 Ultra with 256GB of memory wasn't exactly cheap either (well in retrospect it was a steal too).

But something has changed since I first decided to wait for the 512GB M5 Ultra.

Back then, my thinking was that I wanted as much unified memory as possible to run larger AI models locally on the Mac Studio. I had already experienced the limitations of 256GB, and the idea of having 512GB in a single Mac was very appealing. I was pretty much convinced that this would be my next machine.

Until yesterday, I was running my four DGX Sparks as two separate pairs, with two primarily used for coding and the other two for local AI inference and ComfyUI.

Yesterday, I changed that setup (again). I connected all four DGX Sparks in a ring and configured them to work together as one 512GB cluster, currently running the new DeepSeek 4.1 Flash 1017, it's early, but the tokens per second I'm seeing is just Wow, I have barely used it for coding so far, but from what I read from others on X it's very close to the frontier models.

Technically, of course, it's still four separate machines with their own memory, connected over a QSFP network (I measured 98gb/s per link). It's not the same as having 512GB of unified memory in a single Mac Studio. But for running large AI models distributed across the four machines, I can now make use of their combined memory capacity.

And suddenly, the main reason I wanted a 512GB Mac Studio isn't quite as compelling anymore.

The DGX Sparks have also improved quite a bit on the software side since I bought the first two GB10s. Between vLLM, quantized models and ongoing optimizations, I'm getting more useful work out of them than I did initially. They're not perfect, and I still want more performance, but they're now good enough (better than good enough) for what I'm doing.

My Mac Studio M3 Ultra is still an excellent desktop computer. With 256GB of unified memory, it's hardly struggling with my normal work. Most of the things I do on it, from software development to photography and video editing, don't require anything close to that amount of memory.

So what would I actually gain by replacing it with an M5 Ultra with 512GB?

A faster Mac, obviously. And having half a terabyte of unified memory in a single machine would be nice. There are certainly AI workloads where that would be useful, especially when a model needs to fit into one memory pool rather than being distributed across multiple machines.

But nice and necessary are two different things. And at these prices, I need a better reason than simply wanting the latest hardware.

It's a little funny, considering how certain I was in my previous posts. I was waiting for Apple to release exactly the machine I wanted, promising myself that this time I would go all the way to 512GB. Now it's here, and I'm keeping my M3 Ultra.

I guess Apple managed to cure my upgrade itch with a price tag. :)

Let's see in about three years what the Mac Studio M6 Ultra 512gb (or even 1tb) of memory will cost, maybe the hardware prices has come down to previous level if the shortages from AI data center boom has chilled down.