Nvidia sells tiny new computer that puts big AI on your desktop (arstechnica.com)

Nvidia unveiled the DGX Spark, a $4,000 desktop AI workstation that packs one petaflop of compute and 128 GB of unified memory into a pocket‑sized 2.65‑pound box. Orders open Oct. 15 via Nvidia, partners, and select U.S. retailers. Built on the GB10 Grace Blackwell Superchip, the Spark includes ConnectX‑7 200 Gb/s networking and NVLink‑C2C interconnect (claimed to offer five times PCIe Gen5 bandwidth), runs on 240 W, and is designed to let developers run and fine‑tune much larger models locally than typical consumer GPUs allow. The Spark’s headline feature—large unified memory shared between CPU and GPU—means users can run models reportedly up to ~200B parameters or fine‑tune models as large as ~70B without cloud infrastructure, enabling use cases like local open‑weights LLMs, image synthesis (Flux.1), vision search and summarization with Cosmos Reason, or chatbots using Qwen3 optimized for the platform. That makes it a compelling option for low‑latency, privacy‑sensitive, or offline workflows, though Nvidia and market watchers acknowledge tradeoffs versus cloud pay‑as‑you‑go economics. Technically, NVLink‑C2C and high‑speed networking are key enablers of the Spark’s memory and throughput, positioning it as Nvidia’s attempt to create a new desktop category for serious AI development outside the data center.