NVIDIA is adding a 64GB unified-memory option to its DGX Spark AI workstation, with sales through partner manufacturers due to begin on 23 October. The useful twist is that two systems can be linked to pool 128GB of memory, giving local AI developers room to try larger workloads without moving them to the cloud.
NVIDIA Watch analysis
What happened
The 64GB configuration retains the GB10 Grace Blackwell Superchip, according to NVIDIA, and is due through Acer, ASUS, Dell, Gigabyte, HP and MSI. NVIDIA says one system can run models with up to 100 billion parameters entirely on-device, provided they fit within its memory.
The company also says its Sync Cluster Assistant can configure a connection between two DGX Spark systems over 200 GbE, combining their memory and distributing workloads without manual network setup. NVIDIA reports that two systems achieved up to 1.7 times the performance of one when running its Qwen 3.8 27B model. That is a company-reported result, not a general guarantee for other models or workloads. The details were published by Quantum Zeitgeist.
Why it matters
A smaller-memory configuration could make NVIDIA’s local AI workstation more accessible to developers who do not need the 128GB option. The ability to connect two machines also offers a practical path to larger experiments, though the report gives no price for the new configuration, so “more accessible” remains a question for the checkout page.
The larger point is that local AI hardware is being sold not just as a box with a chip, but as a managed setup for running and scaling workloads. That may save configuration work; it does not make every model fit, or promise that two machines will perform like one larger system.
Our read
The 64GB option and automatic clustering are tangible additions, not just another chip-count boast. But the case for buying one will turn on the missing number: its price. If you are weighing local AI hardware, wait for partner listings and check the memory, model and workload you actually need before treating NVIDIA’s benchmark as your own forecast.
What to watch
- Partner pricing and whether the 64GB model is materially cheaper than the 128GB configuration.
- The models and workloads supported by the two-system clustering setup.
- Independent performance tests beyond NVIDIA’s reported Qwen benchmark.
Discussion spark: Would you rather buy one higher-memory AI workstation or link two smaller systems, if the price and performance were comparable?
Sources and evidence
- NVIDIA DGX Spark Gets 64GB Memory Option For Local AI Work – Quantum Zeitgeist (4 October 2026, 21:25 UTC)
Independent WittyWires tracker for public updates about NVIDIA. Not affiliated with or endorsed by NVIDIA; this is not an official account.