AI Computers
Agent Hub routes work to deskside NVIDIA machines built for agents: DGX Spark for private always-on staff, DGX Station when you need a data center on the desk.
Desktop agent computer
A 15 cm cube for always-on local agents. Private prototyping, air-gapped work, and small swarms.
Deskside data center
A data center on the desk. Hundreds of parallel agents, MIG slices, models approaching a trillion parameters.
Windows AI PC
Same Grace Blackwell family, aimed at creators who want agents inside a normal desktop OS.
Rough seats, not a quote. Light agents are chat and routing. Research is long context. Swarm is many specialists on one product.
~10 concurrent agents · ~128 GB coherent memory
Prototype registry. Production would SSH or call the node API. Air-gapped nodes stay local and never leave your desk.
Sparks run DGX OS with CUDA, vLLM or TensorRT-LLM, and an agent toolkit (NemoClaw / OpenShell class). Stations add MIG slices and Kubernetes so a swarm can be scheduled like a small cluster. Hub Agent can treat a registered node as a bridge: private inference first, SuperPOD later.