NVIDIA has added a 64GB DGX Spark at $4,999 while the 128GB model now sits at around $6,950, against $3,999 at launch a year ago. The cheapest way into NVIDIA's local AI stack has become a lesson in how memory pricing now drives the whole bill of materials.
What was announced
- NVIDIA announced on 2 October 2026 a DGX Spark configuration with 64GB of unified memory, sold only through partners Acer, ASUS, Dell, Gigabyte, HP and MSI from Friday 23 October, starting at $4,999.
- The 64GB model keeps the GB10 Grace Blackwell Superchip, DGX OS, the NVIDIA AI software stack and the built-in ConnectX-7 networking of the 128GB version. NVIDIA says it supports models of up to 100 billion parameters on device.
- Two 64GB units can be linked directly with a QSFP cable and configured by NVIDIA Sync Cluster Assistant, pooling 128GB of memory for models up to 200 billion parameters. In NVIDIA's own Qwen 3.8 27B test the pair delivered up to 1.7x the performance of a single unit.
- The original DGX Spark with 128GB and 4TB of storage went on sale in October 2025 at $3,999. ServeTheHome reports that the 128GB model now sits at about $6,950, with individual OEMs pricing differently, which puts a two-unit 64GB cluster at around $8,000.
- ServeTheHome says it is now paying around $2,000 for a single 64GB ECC RDIMM, and that it still advises buyers to take one 128GB node rather than two 64GB nodes, preferring to scale up within one system before scaling out.
The ETON view
The interesting part of this launch is not the new SKU, it is the price of the old one. A 128GB GB10 box that cost $3,999 a year ago now sits near $6,950, and NVIDIA's answer is to sell less memory rather than hold the price. That is the DRAM shortage arriving at the smallest system in the AI range. If a desk-side developer machine has moved this much, every quote with large memory footprints in it, from GPU servers to dense virtualisation hosts, deserves a fresh look before a budget is signed off.
For teams buying for developers, the 64GB unit is a sensible entry point for agents, fine-tuning and testing smaller models without cloud spend. Be careful with the two-box route to 128GB, though: on NVIDIA's own figures the second unit adds up to 1.7x performance, not 2x, the pair costs more than a single 128GB unit, and it adds cabling and another node to manage. Where the real target is serving a model to a team rather than one developer's workstation, the comparison is no longer with a cloud subscription but with a rack server carrying one or two data centre or RTX PRO class GPUs, where memory, power and support sit inside an estate you already run.
We would not treat any of these prices as settled. The useful questions are which memory capacity the workload actually needs, whether a previous-generation GPU server already in the estate can carry it, and what a quote looks like when it is held for a fixed window. We can price GB10 systems, GPU servers and server memory side by side across vendors, and buy back the hardware being replaced.
Related infrastructure
Category: AI & GPU · Vendor: NVIDIA, Dell, ASUS, GIGABYTE · Technology: DGX Spark, GB10 Grace Blackwell, ConnectX-7, unified memory · Last verified: 04 Oct 2026 · ~2 min read
Sourcing this kind of infrastructure? Talk to ETON about availability, lead time and pricing.
