Nvidia Ships the 64GB DGX Spark: Two Units, 128GB, 200B Models

The new SKU, and the price context

The 64GB DGX Spark went on sale on 23 October, starting at 4,999 dollars and sold exclusively through Acer, ASUS, Dell, Gigabyte, HP and MSI. It uses the same GB10 Grace Blackwell Superchip as the original model, with 64GB of unified memory and a ConnectX-7 network interface built into every unit. The timing is not accidental. The original Spark launched in October 2025 at 3,999 dollars, the 128GB model was raised to 4,699 dollars in February, and Tom's Hardware reports it has since jumped again to 6,950 dollars amid a memory price surge that the outlet bluntly calls the RAM-pocalypse. A 64GB tier at 4,999 dollars is Nvidia re-opening the door it was pricing people out of, one year after launch.

Two units, one cable

The interesting configuration is the pair. Each unit's ConnectX-7 ports connect directly over a QSFP cable on a 200GbE link with no switch in between, pooling the two boxes into 128GB of unified memory. A single 64GB unit supports models up to 100 billion parameters; a pair extends that to 200 billion parameters. On Nvidia's internal test with Qwen 3.8 27B, two clustered units delivered up to 1.7 times the performance of a single system.

The economics are worth stating plainly, because two 64GB units cost about 10,000 dollars while a single 128GB model, before the latest increase, was 4,699 dollars. Nobody pairs two Sparks to save money. The pairing exists so that people who bought one unit today have a documented growth path that does not involve selling it, and so the memory shortage does not entirely close off the 200B-parameter tier for desk-side use.

The software story is the quiet half

Nvidia is shipping this as a complete local-AI kit rather than a bare box. DGX OS, the CUDA-X libraries, the Agent Toolkit, Nemotron models, Ollama, vLLM and PyTorch with CUDA are supported out of the box, and llama.cpp and LM Studio are supported for getting started. The Sync app detects a connected second unit, validates the configuration and sets up the ConnectX-7 network automatically, and a Sync Model Launcher due by the end of the month will download and launch models on one or both units with a few clicks, including an option to set up OpenCode for browser-based coding.

What it means for local AI

For readers deciding where to spend a local-AI budget, the Spark 64GB lands in a specific slot: it is not competing with a gaming PC running a quantized 70B, it is competing with renting GPUs by the hour for models in the 70B to 200B class, with the data never leaving the desk. The verified performance claim to take seriously is the 1.7x scaling on Qwen 3.8 27B, which says the interconnect is real for at least one open model; independent numbers on larger models will be the real test.

The price history is the other lesson. The same hardware class cost 3,999 dollars a year ago and the big-memory tier has risen 74 percent since February, all tracking the memory market rather than any change in the product. Anyone who has been waiting for local hardware to get cheaper has just watched it get more expensive for twelve straight months. The 64GB Spark is the counter-programming: less memory per dollar, but a cluster path and a price that still exists.