Nvidia used its Berlin conference to argue that the next wave of agents will not all live in a hyperscaler. The pitch rests on a smaller object than a data centre: a tower that sits under a desk.
DGX Spark systems with 64GB of unified memory began shipping from partners on October 23. The company says they will run models such as Nemotron 3.5 Bolt 74B, Qwen3.8 27B and Gemma 4 26B-A4B without anything leaving the building.
The more consequential detail is clustering. ConnectX-7 networking and a Sync Cluster Assistant let up to eight units work together in manual configuration, which Nvidia reckons at 1TB of unified memory. That is enough headroom for frontier-class open weights like GLM 5.3, a model that would otherwise need cloud capacity.
Nvidia also shipped software to make the box less of a project. A Model Launcher due later in October downloads a model and its container, wires it to OpenCode and loads troubleshooting guidance. A Wi-Fi Assistant handles network switching when the machine runs as an appliance.
For bigger teams there is DGX Station, with 748GB of coherent memory and up to 20 petaflops, plus a Windows variant arriving in December. Early access to NVIDIA Cloud Functions would let a team register local endpoints behind one compatible entry point.
The bet is straightforward. Regulated buyers and agent developers want inference where the data already is, and Nvidia wants the software stack to follow it there.
