Nvidia's pint-sized desktop AI supercomputer is getting a price shake-up that captures just how expensive memory has become. As first detailed by ServeTheHome, Nvidia is introducing a DGX Spark 64GB model priced from $4,999 — even as the original 128GB configuration climbs to a roughly $6,950 tier across OEM partners, a steep rise from the $3,999 launch price of the 128GB/4TB model about a year ago.
According to the report, the new 64GB DGX Spark will be available from Acer, ASUS, Dell, Gigabyte, HP and MSI starting Friday, October 23. MIXED Reality News also reported the launch, highlighting that two of the 64GB units can pool their memory into a single 128GB cluster. For more context on this story, see our ongoing AI trends.
What You Get for $4,999
Despite the halved memory, Nvidia is quick to point out that the 64GB version keeps the feature that made the DGX Spark interesting for clustering: NVIDIA ConnectX-7 200GbE networking, carried over from the 128GB model. The company also says two of the new 64GB machines can be easily clustered for a combined 128GB of memory at around $8,000, with software support for the configuration built in.
That clustering story is a big part of the pitch. The DGX Spark line is built on Nvidia's GB10 chip, a compact system-on-chip designed to bring Blackwell-generation AI performance to a desktop box. For developers and researchers running mid-sized models locally — or stacking several units into small clusters — the machines slot into a niche that laptops and full data-center GPUs leave open.
Why Prices Are Climbing
The sticker shock has a straightforward cause: memory prices are still increasing, driven by the enormous demand for high-bandwidth memory in AI data centers. ServeTheHome's editors were blunt about the economics. "We are paying $2,000 for 64GB ECC RDIMMs, so pricing logically must increase," the site's chief analyst said, describing the market reality behind the new tiers.
The DGX Spark is not the only Nvidia product feeling the squeeze. Reports from Ars Technica and gHacks this week noted that Nvidia has raised the price of the aging Shield TV Pro by $100 to $299.99, citing memory costs — a sign that the memory supercycle is feeding through to consumer hardware, not just AI accelerators.
A Crowding Mid-Range Market
The repricing lands in an increasingly competitive corner of the market. ServeTheHome notes that the new $6,950-ish 128GB tier puts the DGX Spark in the same price range as AMD's upcoming "Gorgon Halo" systems, which offer around 192GB of memory capacity at similar memory bandwidth when they arrive in October. Apple, meanwhile, is preparing to ship a $10,000 256GB Mac Studio M5 Ultra configuration that would undercut a two-unit Nvidia cluster on both price and unified memory capacity.
For buyers, the calculus now involves trade-offs that barely existed a year ago: Nvidia's superior networking versus AMD's higher memory capacity, or Apple's memory density at the top end. ServeTheHome's own guidance is that most buyers who can afford it should still prefer a single 128GB node over two clustered 64GB units, based on the site's experience running GB10 clusters.
What It Means for Local AI Builders
The DGX Spark launched at $3,999 and was widely seen as an accessible entry point for running capable open-weight models on a desk. At $4,999 for 64GB — or roughly $8,000 for a clustered 128GB pair — that accessibility has eroded, and the price of entry to local AI hardware is now firmly in workstation territory.
ServeTheHome frames the shift with a wry comparison: $6,950 is nearly three years of a $200-per-month frontier AI subscription. The counterargument, the site notes, is privacy and control — cloud subscriptions mean sending data to a third party, and with each frontier release the gap between closed models and the smaller open models that fit on a desktop has, by ServeTheHome's assessment, been widening.
Either way, the message from this week's pricing is clear: the memory feeding the AI boom has become the binding constraint, and it is repricing everything from $100,000 DGX Station workgroup systems down to a seven-year-old streaming box. Builders hoping memory would get cheaper before committing to local hardware just got their answer.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
