Nvidia unveiled a 64GB-memory version of its DGX Spark desktop AI system on October 2, starting at $4,999. The unit goes on sale October 23 exclusively through six partners: Acer, Asus, Dell, Gigabyte, HP, and MSI.
The core specs match the 128GB model: the same GB10 Grace Blackwell Superchip, ConnectX-7 networking, DGX OS, and the full Nvidia AI software stack. The difference is unified memory, cut in half — Nvidia says a single unit can run models up to 100 billion parameters, versus 200 billion for the 128GB version.
Introducing Sync
Launched alongside the new hardware is NVIDIA Sync, a clustering assistant. Two 64GB units connect directly over a QSFP cable on a 200GbE link, and Sync detects the paired unit, validates the configuration, and automatically sets up ConnectX-7 networking. Once linked, the combined memory reaches 128GB, memory bandwidth doubles, and the cluster can run models up to 200 billion parameters.
Nvidia provided its own benchmark: running Qwen3.8 27B, a cluster of two 64GB units delivered performance up to 1.7 times that of a single 128GB unit. The figures come from Nvidia internally — the blog post did not disclose the test workload, batch size, or precision settings — and independent testing will have to wait until the October 23 launch.
Doing the Math
When DGX Spark launched last October, the 128GB model was priced at $3,999 — about $31 per gigabyte. The new 64GB model costs $4,999, or roughly $78 per gigabyte, about one and a half times higher on a per-gigabyte basis.
The 128GB model's price has also climbed over the past year. Reports say Nvidia's own 128GB Founders Edition has risen to $6,950, and Tom's Hardware found partner retail prices ranging from $7,000 to $9,000. Nvidia's blog post did not disclose a current official price for the 128GB model; the figures above come from retail channels and fluctuate significantly.
Side by side:
| Configuration | Memory | Price (USD) | Max model per unit |
|---|---|---|---|
| Original 128GB launch | 128GB | 3,999 | 200B parameters |
| 128GB partner retail | 128GB | ~7,000-9,000 | 200B parameters |
| New 64GB | 64GB | From 4,999 | 100B parameters |
| Two-unit 64GB cluster | 128GB | ~9,998 | 200B parameters |
Roughly speaking, getting to 128GB by buying two 64GB units costs $1,000 to $3,000 more than buying one 128GB unit at retail, in exchange for Nvidia's claimed performance gain of up to 1.7 times and the compute of two chips. The 128GB model also ships with a 4TB SSD; none of the reports have clearly specified the 64GB model's storage configuration, so the real price gap will depend on the specific partner model.
Several outlets have attributed the price increase to memory costs; Nvidia has not commented on the pricing. Based on the specs and pricing, the 64GB model looks like an "entry ticket" — pulling the starting price back under $5,000 lets developers who want to fine-tune or debug agents with 70-billion-parameter-class models locally get in first, with the option to add a second unit later if they need more memory.
Whether retail pricing for the 64GB model holds at $4,999 after launch will depend on how the memory market moves — the 128GB model's climb from $3,999 to seven-to-eight-thousand dollars at retail is already a cautionary precedent.
Sources: Nvidia official blog, StorageReview, tbreak, CocoLoop; the Nvidia blog verifies the 64GB model's price, sales channels, parameter specs, and 1.7x cluster figure, while retail pricing follows the Tom's Hardware figures cited by tbreak.