$1,951. That’s the gap between the two DGX Spark configurations now on NVIDIA’s menu — $4,999 for the new 64GB box versus $6,950 for the 128GB one that’s been out in the wild. Same GB10 Grace Blackwell chip. Same software stack. Half the unified memory.
For those of us who build bots for a living, that number is the whole story. It’s not a spec question. It’s a question about what you actually load into memory on a Tuesday afternoon.
What’s actually on offer
The short version, straight from the announcement:
- 64GB of coherent unified memory, down from 128GB
- The same GB10 Grace Blackwell silicon as the existing config
- The same software stack — nothing stripped, nothing special-cased
- $4,999, available starting Friday, Oct. 23
- Sold through Acer, ASUS, Dell, Gigabyte, HP and MSI
- Two units can be clustered with the Sync Cluster Assistant for bigger models or heavier jobs
NVIDIA is framing it as a lower-cost entry point for running models locally. That framing is accurate, and it’s also the most interesting thing about the release, because “entry point” means something very different to a solo builder than it does to a lab with a procurement department.
The memory question is the only question
Unified memory on these machines is the budget you spend on everything at once — weights, context, whatever else you’ve got resident. Cutting it in half doesn’t cut your capability in half, but it does change which models sit comfortably and which ones make you start negotiating with quantization settings.
Here’s how I’d think about it as someone who ships agent pipelines rather than benchmarks. If your working set is a mid-sized model plus an embedding model plus a reranker, all quantized, 64GB is a real workspace. If your habit is loading the largest thing you can find at the highest precision you can afford and then seeing what happens, you were always going to want the 128GB box, and you probably already knew that before you read this.
What I like about the 64GB config is that it removes the excuse. A lot of bot builders have been renting GPU hours to do work that is fundamentally iterative — prompt tuning, tool-calling loops, retrieval experiments, the thousand small runs where you’re debugging your own orchestration logic rather than the model. That work doesn’t need frontier-scale memory. It needs to be local, fast to restart, and free at the margin. Two grand off the price of admission matters a lot for that use case.
Doing the clustering math honestly
The Sync Cluster Assistant support for pairing two units is the detail that’ll get the most speculation, so let’s be careful with it. NVIDIA says two can be clustered for larger models or heavier workloads. That’s the fact. Everything else is arithmetic and judgment.
The arithmetic: two 64GB units run $9,998 and give you 128GB of aggregate memory. One 128GB unit runs $6,950 and gives you 128GB in a single pool. So if memory capacity alone is what you’re chasing, buying two of the cheaper boxes is the worse deal by about three thousand dollars.
The judgment: capacity isn’t the only thing you get. Two units means two GB10 chips, which is a different proposition — more parallel compute, two independent machines when you don’t need them paired, and a growth path that doesn’t require throwing out what you bought first. I haven’t tested the clustering behavior and can’t tell you how a given model partitions across two boxes or what the interconnect does to throughput. Anyone who tells you otherwise right now is guessing.
My read: the 64GB config is best understood as a single machine that happens to have a second-unit option, not as half of a planned pair. Buy it because $4,999 gets you a local AI workstation. Treat clustering as insurance against your ambitions growing, not as the reason for the purchase.
Who should care
If you’re building bots and currently doing all your inference against an API, this is the cheapest serious on-ramp NVIDIA has offered for its own local AI hardware. You get the same stack the 128GB users get, which means tutorials, container images and tooling all apply without translation.
If you already own the 128GB version, nothing here is for you. If you were waiting for a price that made local development defensible on a one-person budget, $4,999 is closer to that line than $6,950 was — not cheap, but a different category of decision.
And if you’re genuinely unsure which one you need, the honest answer is to look at what you loaded last month. Memory requirements are the easiest thing to measure and the easiest thing to lie to yourself about. Check the numbers, then pick the box.
🕒 Published: