Nvidia's lunchbox-sized DGX Spark puts a 128GB AI supercomputer on your desk.
The DGX Spark is a small desktop box built around Nvidia's GB10 Grace Blackwell superchip, pairing a Blackwell GPU with an Arm CPU on one die. It ships with 128GB of unified memory, enough to run and fine-tune mid-size open-source models like a 70-billion-parameter Llama variant without touching the cloud. Nvidia prices it at roughly $3,999 and sells it as a personal AI lab for developers who want to prototype before scaling up to real data-center hardware. The hands-on confirms the pitch: train, test, and fine-tune smaller models locally, with no cloud-compute bill and no shipping your proprietary data to someone else's server.
That matters because local fine-tuning has mostly meant either a $10,000-plus multi-GPU workstation or a recurring cloud bill that scales with every experiment. A machine the size of a lunchbox that handles both without a rack of fans changes who can afford to tinker with their own models in-house, particularly in regulated industries where sending data outside the building isn't an option. It also positions Nvidia to own the hobbyist and prosumer AI-dev tier the way it already owns the data center.
Whether $3,999 feels like a bargain or a vanity purchase depends on whether you were already paying a cloud provider that much to run the same experiments.