NVIDIA announced on Oct. 2 a new DGX Spark configuration with 64GB of unified memory. It retains the company’s Grace Blackwell-based GB10 superchip, DGX OS and NVIDIA software stack, while offering less memory than the existing 128GB model. The manufacturer says the product will go on sale Friday, Oct. 23 through Acer, ASUS, Dell, Gigabyte, HP and MSI, at a starting price of $4,999. This is the announced starting retail price, not a business procurement rate or total cost of ownership.

NVIDIA says one 64GB system can run models with up to 100 billion parameters on device. Connecting two systems directly brings combined unified memory to 128GB; the company says the configuration can support models up to 200 billion parameters. In NVIDIA’s Qwen 3.8 27B test, two clustered 64GB systems delivered up to 1.7 times the performance of one system. That figure is the company’s result for a particular model and setup; it does not establish the same acceleration across models or an independent comparison.

The configuration is aimed at developers, researchers and users experimenting with local AI. NVIDIA highlights inference, fine-tuning, data science and agent workflows on the device without renting a cloud instance for every task. Clustering two units may simplify network and software configuration, but doubling memory alone does not determine application latency, model quality or cost. Results depend on the model, context length, data transfers and workload concurrency.

Acer, ASUS, Dell, Gigabyte, HP and MSI were named as sales partners; NVIDIA says the 64GB SKU will be offered through those manufacturers and will not replace the 128GB version. This channel distinction separates product availability from NVIDIA’s own hardware sales strategy. The $4,999 starting price applies to the configuration scheduled for Oct. 23; the announcement does not specify taxes, regional prices or inventory.

For organizations assessing AI investment, the economics of local inference depend on utilization. A device may keep sensitive data local and reduce variable cloud consumption, while purchase, electricity, cooling, operations and updates remain user costs. The announcement gives no sales forecast, customer count or total savings calculation versus cloud services. It should therefore be read as a priced product configuration and a vendor performance claim, not realized market demand or verified return on investment.