Nvidia's 'Personal AI Supercomputer' (www.theverge.com)

Nvidia this week begins selling the DGX Spark — billed as a “personal AI supercomputer” — with orders opening October 15th via nvidia.com and select partners in the U.S. The compact desktop box, previously called Digits, ships with Nvidia’s GB10 Grace Blackwell Superchip, 128 GB of unified memory, up to 4 TB NVMe storage and delivers roughly a petaflop of AI throughput. Nvidia says Spark can run models up to about 200 billion parameters, and the listed price in today’s materials is $3,999 (up from an earlier $3,000 teaser). Multiple OEMs including Acer, Asus, Dell, Gigabyte, HP, Lenovo and MSI will offer customized Spark variants. For the AI/ML community, Spark’s significance is practical and symbolic: it brings datacenter-class model training and inference capability to individual researchers, labs and developers without rack space or specialized power/cooling. That can accelerate experimentation, improve privacy and lower friction for iterative model development by enabling large-model work locally. Technically, the unified memory and petaflop-level compute make it suited for fine-tuning large transformer models or running sizable inference workloads on-premises, while OEM availability may broaden access and foster a market of small-form-factor AI workstations.