The 256GB M5 Ultra is already ~$18,000 pre-tax at 16TB storage.
I expect the 512GB to sit around ~$30,000.
For $30,000 I’d prefer investing in 4 more 128GB DGX Sparks.
Needless to say the 256GB wont be available before Feb 2027 so 512GB may be available late Q1 2027.
Apple is the only one currently offering a unified 512GB RAM in mini PC form; but at potentially $30,000+, if portability is not critical, the NVIDIA GB300 DGX Station with 748GB unified/coherent memory starts at $85,000 and some providers offer LTO plans.
It’s funny that in another thread people didn’t believe me when I mentioned the price starts at $85k…
Just a quick clarification: the GB300 has 768GB of memory, where ~250GB runs at 7 TB/s, and the remaining ~500GB runs at 400 or 500 GB/s. In contrast, the M5 has all of its 512GB running at a uniform 1.2 TB/s.
Just a minor clarification.
The 256GB M5 Ultra is already ~$18,000 pre-tax at 16TB storage.
I don’t think many people would buy a 16TB Mac Studio. They’ll likely go with either 4TB or 8TB models, and probably use their Thunderbolt external drives. Here’s (IMHO) a more fair comparison:
256GB M5 Ultra with 8TB NVMe SSD - ~$14,300 pre-tax.
2-node DGX Spark FE stack - ~$14,000 (using the current pricing from Microcenter at $6950 for the DGX Spark FE model), plus another ~$100-$150 for the QSFP cable.
For people already comfortable with the Apple ecosystem, the 1.2TB/s memory bandwidth on the Mac Studio may likely be more appealing at this price range. You get ~4x the memory bandwidth of the Spark for the same capacity without requiring clustering and other tinkering (not to mention the onboard TB5 ports, which open a number of possibilities for storage expansion, etc).
If I didn’t already have a 2-node Spark cluster, I’d probably be spending some quality time deciding how to spend $15k.
The RTX Spark, with under 110 GB usable for models, won’t run anything smart. Too much Windows bloat, and no ConnectX-7 to pool memory with other devices. It looks more like a gaming machine.
If I had to bet, I’d put my money on NVIDIA keeping up small investments in DGX. If RAM prices fall, the workstation is ripe for a price cut and becomes a sensible upgrade for this community. The 64 GB version was released to pull in first-timers.
The prices look ugly today. But picture the day the bubble pops and RAM prices drop: a DGX range from 64 GB to 768 GB is a good product line for the persistent-agent personal computer with local intelligence.
The new rate jumping does make the 2-node DGX Spark Bundle comparable to a 8TB 256GB MAC M5 Ultra.
If your work doesn’t need local NVIDIA OS, Mac’s 1.2TB/s memory bandwidth and its 256GB unified memory has an edge over the DGX Spark.
Edit: 512GB M5 Ultra with 4TB launching later this month may be at the same rate of the 2 DGX Sparks with massive RAM
Apple also offers SMB LOC + trade-in upgrade options, so the investment maybe more palatable when paid over time, with an upgrade route, for those interested in remaining within their ecosystem.
The RTX Spark launch seems to pivot the Spark family from the original “affordable” 3K dev box pitch we heard from Project DIGITS last year making me more interested in keeping an eye on alternative options like the MAC Studio.
Framework’s $6,799 192GB Gorgon Halo launch is interesting and I’m keen to hear interdependent impressions when users get their hands on units in November
Xiaomi’s AI Cube is also interesting if they can pull off the 3-chip architecture and keep their prices true to the 5% profit margin cap.
Personally, I’m grateful to have my bundle and the second bundle on the way. Until the Vera Ruben Spark launches with the next gen Blackwell I’m not investing to cluster further with current DGX Sparks at the newly inflated price.
Now that Exllamav3 added support for ROCm, Strix/Gorgon Halo is a going be a threat too. GLM 5.3 Flash would fit on a single Gorgon Halo 192GB at 3.5bpw EXL3. Yes, not super fast, but GLM 5.3 Flash 2.77bpw EXL3 with DFlash2 on a single Spark is quite usable. 3.5bpw EXL3 would give a run for the money to the NVFP4 quality wise.
MINISFORUM MS-S1 (both 395 and the 495 version) can take PCIe x16 card, so throw some cheap RDMA card and you have yourself a cluster capable “ghetto spark”. The traffic shouldn’t be that high for inference, so you might not even need anywhere near 200Gbit, ConnectX 5 or 6 might work fine. As long as tensor parallelism via RDMA works on the AMD platform, this probably the best bang for a buck now. Watch prices to skyrocket on these. Strix Halos are still relatively cheap.
It’s easy to say this in retrospect but the truth is that Nvidia did not have a runaway hit with this platform out of the gate. Reception a year back was tepid. In fact, I honestly believe we coasted on boards produced with LPDDR5x RAM reserved in early to mid 2025 for about a year.
Models got better, drafting went from a toy to standard allowing us to break beyond what 273 GB/s could deliver on paper, harnesses got better too, and everyone else’s prices scaled harder because GB10 still had old component prices baked in.
Now the other shoe is dropping. It was always going to. For a while I think Nvidia and partners might have believe they overproduced GB10, not underproduced.
How do you cover a journey of thousand miles? Just making one step at a time again and again. Agents help but don’t fail to learn alongside and just spectate.