NVIDIA will let developers buy its DGX Spark desktop AI computer with 64GB of unified memory instead of 128GB, but only through six partner manufacturers.
Acer, ASUS, Dell, Gigabyte, HP and MSI start selling the configuration on Friday, Oct. 23, with prices from $4,999, NVIDIA says. NVIDIA describes that as an accessible price point and says the smaller version keeps the GB10 Grace Blackwell Superchip, DGX OS and the full NVIDIA AI software stack of the 128GB model. With the processor unchanged, memory capacity becomes the main variable a buyer is choosing.
NVIDIA puts the ceiling for a single 64GB unit at models of up to 100 billion parameters running entirely on the device, with no cloud dependency. Larger jobs point to a second machine. Per NVIDIA, two units connect with a QSFP cable, and the Cluster Assistant in the NVIDIA Sync app detects them, validates their configuration and configures the ConnectX-7 network. A pair pools 128GB and handles models of up to 200 billion parameters, the company says, and in its own Qwen3.8 27B test two clustered systems delivered up to 1.7x the performance of one.
Clustering is not new to the line. NVIDIA's June description of the assistant covered two to four units, with up to 512GB of unified memory across four. The 64GB announcement describes only the two-unit case and does not say whether the smaller machines can be chained in threes or fours.
Not everything is ready on day one. A Sync Model Launcher due at the end of the month is meant to download and start Qwen3.8 27B on one unit or a cluster and set up OpenCode, so coding can happen in a browser. Playbooks for vLLM, for running OpenClaw against a local model and for linking several Sparks are listed as coming soon for 64GB devices. NVIDIA also lists Blender among the first creative apps to support the platform, though that prebuilt installer is still to come.
A pair at the starting price costs at least $9,998 before any cabling, and the announcement gives no price for the single 128GB model. Whether two 64GB machines beat one larger machine bought outright therefore cannot be judged from NVIDIA's numbers.
