NVIDIA launches a 64GB DGX Spark from US$4,999 on 23 October

NVIDIA has announced a 64GB version of its DGX Spark desktop AI computer. In a 2 October 2026 blog post, the company said the machine will be sold only through manufacturer partners from Friday 23 October, starting at US$4,999. The post does not give a Malaysian price or name a Malaysian retailer.[1]

The pitch is local AI. Models, and the agent software built on them, run on a small machine at a desk instead of in a rented cloud. NVIDIA also says two units can be linked. The performance figure it gives comes from its own test, and we have seen no independent results.[1]

NVIDIA says the new configuration keeps the GB10 Grace Blackwell Superchip, DGX OS and the full NVIDIA AI software stack, the same as the 128GB model. The difference is memory: 64GB of unified memory shared by processor and graphics. NVIDIA says this keeps the platform at an accessible price point.[1]

The post says the 64GB machine supports models of up to 100 billion parameters and the agentic applications built on them, fully on the device. It does not say at what numerical precision or at what speed. It also does not give a price for the 128GB model, so the saving cannot be judged from the post alone.[1]

NVIDIA says two 64GB units can be joined with ConnectX-7 networking and a QSFP cable, with a new NVIDIA Sync Cluster Assistant handling setup. A pair gives 128GB of pooled memory and twice the memory bandwidth, according to the company. By our arithmetic, two units at the starting price would cost US$9,998.[1]

In NVIDIA's Qwen 3.8 27B test, two clustered systems delivered up to 1.7x the performance of one. The words up to matter, because a ceiling is not an average. The figure comes from NVIDIA and covers one model. Independent reviewers have not yet had the shipping hardware.[1]

NVIDIA lists DGX OS, its AI software stack, Agent Toolkit, CUDA-X AI libraries and Nemotron open models. It says tools such as Ollama, vLLM and PyTorch with CUDA are supported. It describes Sync Model Launcher as a way to run local AI by clicking a few buttons. Our reading is that setup effort will matter as much as hardware for small teams.[1]

NVIDIA also names Blender among the first major creator applications to support the platform. The post lists always-on agents, local inference for PC applications and workflow management as uses. Shipping begins only on 23 October, so nobody outside NVIDIA can yet confirm these claims in daily use.[1]

Bank Negara Malaysia's 2 October rates were RM4.0710 to buy and RM4.0960 to sell per US dollar. On those, US$4,999 is roughly RM20,350 to RM20,480. That is our arithmetic, before any distributor margin, tax or import cost, none of which NVIDIA has announced. The post names six global manufacturers and no Malaysian reseller.[1][2]

Local hardware matters because of data rules. Section 129 of the Personal Data Protection Act 2010 regulates sending personal data out of Malaysia, and the Commissioner's Guidelines No. 3/2025, issued on 29 April 2025, set out the conditions. Our reading is that a model running on a machine in the firm's own office does not by itself send customer data abroad, although cloud-connected tools can. Check each workflow with whoever handles compliance.[3]