NVIDIA just cut the DGX Spark’s memory in half. It cut the price by about 28%. Those two numbers don’t belong in the same sentence, and yet here we are.
The new DGX Spark 64GB lands at $4,999, available starting Friday, Oct. 23, through Acer, ASUS, Dell, Gigabyte, HP, and MSI. It carries the same GB10 Grace Blackwell chip and the same software stack as the 128GB model, which sells for $6,950. So you’re giving up 64GB of coherent unified LPDDR5X memory to save $1,951. If you were hoping the entry-level tier would scale down proportionally, that’s not what happened.
What you’re actually paying for
Memory is the whole story with this box, and NVIDIA knows it. The GB10 silicon is identical. The software stack is identical. Nothing about the compute changed. The only variable NVIDIA turned down is the one that determines which models you can actually load, which means the 64GB Spark is less a different product than it is the same product with a smaller ceiling.
That reframes the value question. You’re not buying a cheaper machine with less horsepower. You’re buying the same horsepower with less room to stretch. For some developers that’s a perfectly fine trade. If your work lives in the small-to-midsize model range, fine-tuning adapters, building agent pipelines, testing inference locally before you push to a cluster, the extra memory on the 128GB unit was headroom you weren’t using. Paying nearly two grand for headroom you don’t touch is a bad deal, and NVIDIA is right to offer an alternative.
But the pricing gap tells you how NVIDIA sees its own cost structure. Memory prices have been climbing, and halving capacity apparently doesn’t halve the bill of materials. The chip, the board, the networking, the thermal design, the software licensing: all of that carries over. What’s left is a $1,951 discount on a $6,950 machine, which is real money but not the kind of price break that opens a new tier of buyer. This isn’t a budget option. It’s a slightly less expensive option.
The clustering angle is the interesting part
Here’s where it gets genuinely useful. NVIDIA supports clustering two Sparks together via the Sync Cluster Assistant for larger models or heavier workloads. Run that math and something awkward appears.
- Two 64GB units: 128GB total memory, $9,998
- One 128GB unit: 128GB memory, $6,950
Two of the cheap ones cost about $3,048 more than one of the expensive ones for the same nominal memory pool. You’d be paying a premium for twice the compute, which may well be worth it depending on your workload. But if memory capacity is what you’re chasing, and for local AI work it usually is, the single 128GB box wins on price. The clustering path makes sense as an upgrade story for someone who already owns a 64GB Spark and outgrew it, not as a purchasing strategy from scratch.
Which, to be fair, is probably exactly how NVIDIA intends it. Buy in cheap, discover your models need more room, buy a second unit. That’s a reasonable product ladder. It’s also a ladder where every rung costs five grand.
Who should care
If you’ve been circling the DGX Spark and the $6,950 price tag was the thing holding you back, $4,999 might get you there. You get the same chip, the same tooling, and a machine designed to run models on your desk instead of someone else’s datacenter. For developers who need local inference for privacy, latency, or cost reasons, that proposition hasn’t changed.
If the 128GB model’s memory was what sold you, nothing here should tempt you. Buying the 64GB unit and planning to cluster later means spending more total money for the same memory you could have had on day one. Know which camp you’re in before you order.
And if you were waiting for NVIDIA to produce a genuinely affordable local AI workstation, this isn’t it. A $4,999 entry point is still a serious capital expense for an individual developer and a rounding error for a well-funded team. The people most likely to buy this are the same people who were already considering the 128GB version, just with a different budget conversation.
Honest read
This is a sensible product move dressed up as a price cut. NVIDIA widened the funnel without touching the architecture, which costs them almost nothing and gives buyers a lower bar to clear. The memory math is unflattering if you look at it too closely, and the clustering option is more upgrade path than value play.
Still, more configurations beats fewer configurations. If 64GB fits your models, you just saved $1,951. If it doesn’t, you now have a clearer picture of what the extra memory actually costs. That’s a reasonable outcome, even if it isn’t the dramatic one the headline implies.
🕒 Published: