NVIDIA's $4,999 DGX Spark 64GB puts 100B-parameter local inference on a desk

A cheaper tier of NVIDIA's desktop AI box

NVIDIA has begun selling a 64GB configuration of its DGX Spark desktop through Acer, ASUS, Dell, Gigabyte, HP and MSI, available from October 23 with a starting price of $4,999. According to a post on dev.to by Netics Labs, which cites NVIDIA's announcement, the new variant keeps everything else from the existing 128GB model: the GB10 Grace Blackwell superchip, DGX OS and the same AI software stack. The unified memory is the main difference, cut in half to 64GB.

NVIDIA positions the machine, per the post, at developers, researchers and enthusiasts who want capable AI agents running privately on the device rather than depending on a cloud provider. The company claims a single unit supports models of up to 100 billion parameters entirely on device.

What ships in the box

The software bundle is what decides whether the hardware gets used, the post argues. The platform ships with the NVIDIA Agent Toolkit, CUDA-X AI libraries and NVIDIA's Nemotron open models, and supports runtimes teams already run, including Ollama, vLLM and PyTorch. Blender is named as an early supporter on the creative side, with an installer on the way. Taken together, the list describes a machine for development and continuous inference, sized to the models that fit in 64GB of unified memory, rather than a replacement for rack hardware.

Clustering doubles the ceiling

The more consequential half of the announcement is the two-box story. Every unit carries a ConnectX-7 network interface, and two units can be linked directly with a QSFP cable, with no switch in between. That link forms a 200 GbE fabric, combines the two systems' memory into 128GB, doubles memory bandwidth and extends supported model size to 200 billion parameters. NVIDIA's own figures, relayed in the post, show two clustered 64GB systems delivering up to 1.7 times the performance of one on a 27-billion-parameter Qwen 3.8 workload. A tool called the NVIDIA Sync Cluster Assistant handles detection, configuration validation and network setup so both nodes run an identical stack.

Netics reads the cluster option as a purchasing decision more than a performance claim: a team can begin with one unit and add a second when its workload exceeds the memory ceiling, using capacity as the trigger to spend. The post also flags the limitation that two units in one office are still one physical location, which matters for anyone tempted to treat this as distributed infrastructure. A related tool, the NVIDIA Sync Model Launcher, is due at the end of the month and will download and start the same Qwen model on a single unit or a cluster, configuring OpenCode to use it.

The costs that do not appear on the invoice

The post's central argument is that local inference is a narrower deal than it is usually sold as. Fixed hardware beats rented compute when a workload runs steadily enough to keep the box busy, when the data benefits from never leaving the building, and when the model fits within the memory purchased. Renting wins when usage is spiky, when the team needs whatever the largest model happens to be next quarter, or when nobody is responsible for the machine's upkeep.

That upkeep is the hidden line item: a model registry, a schedule for driver, firmware and runtime updates, and monitoring that notices when a node starts answering slowly. Before ordering, the post recommends confirming that the largest model actually needed fits the memory ceiling, that the workload runs at least a few hours a day, that data residency requirements are understood, that someone owns updates, and that the second unit and its connecting cable are priced into the plan.

Why it matters

A desk-side box that runs 100-billion-parameter models for $4,999 moves a whole class of inference work from recurring cloud bills into a capital purchase a small team can approve. The two-unit path to 200 billion parameters adds headroom without changing the software environment, and the vendor's own 1.7x scaling figure, if it holds in practice, makes the second box a rational upgrade rather than a leap. But as Netics makes clear, the economics only close for workloads that are steady, sensitive and sized to the memory ceiling. For bursty teams or those chasing the newest models, rented GPUs remain the cheaper answer, and the operational burden of owning a local stack should be weighed as carefully as the price tag.