Nvidia will start taking orders Wednesday for the DGX Spark, a compact desktop AI computer priced at $3,999. The system delivers one petaflop of computing performance and includes 128GB of unified memory in a box small enough to sit on a desk. According to Ars Technica, the DGX Spark aims to run larger AI models locally without relying on cloud services or data centers.
Large Memory Enables Bigger Models
The DGX Spark can run AI models with up to 200 billion parameters and fine-tune models containing up to 70 billion parameters. This capability comes from its 128GB of unified memory shared between system and GPU tasks. The system runs on Nvidia’s GB10 Grace Blackwell Superchip and uses NVLink-C2C technology that provides five times the bandwidth of PCIe Gen 5.
Consumer GPUs like the RTX 5070 offer only 12GB of video memory, while the RTX 5090 typically ships with 24GB. The DGX Spark’s 128GB memory allows developers to run much larger models locally. For example, running OpenAI’s 120 billion-parameter gpt-oss language model requires about 80GB of memory, far more than consumer GPUs provide.
Technical Specifications
The device weighs 2.65 pounds and measures 5.91 x 5.91 x 1.99 inches. It uses 240 watts of power and includes ConnectX-7 200Gb/s networking. The ARM-based system runs Nvidia’s DGX OS, an Ubuntu Linux-based operating system built for GPU processing. It comes with Nvidia’s AI software stack preinstalled, including CUDA libraries and NIM microservices.
Market Position and Availability
Nvidia will sell the DGX Spark through its website starting October 15. Manufacturing partners and select US retail stores will also carry the system. The Register reports that the GB10 chip’s GPU computing performance roughly matches an RTX 5070, but the larger memory pool allows it to run bigger models at slower speeds.
The $3,999 price point sits well below high-end GPUs like the RTX Pro 6000, which costs about $9,000, and AI server GPUs like the H100, which starts at $25,000. Nvidia CEO Jensen Huang delivered one of the first units to Elon Musk at SpaceX’s Starbase facility in Texas, repeating a similar delivery Huang made to Musk at OpenAI in 2016 with the original DGX-1 system.
