Summary
NVIDIA says DGX Spark can now be configured in clusters of up to eight systems, with software to help developers run local AI models. The company also announced a Windows version of DGX Station and open-source robotics workflows.
NVIDIA announced new local AI hardware and software at GTC Berlin on Oct. 21, 2026, including support for manually configured clusters of up to eight DGX Spark systems. The company also outlined software for setting up local models and a Windows version of its larger DGX Station system, aimed at enterprise and engineering teams.
DGX Spark scales to eight systems
New DGX Spark 64GB systems are scheduled to begin rolling out from partners on Friday, Oct. 23. Developers can connect up to eight units using built-in NVIDIA ConnectX-7 networking and manage the cluster with NVIDIA Sync Cluster Assistant.
NVIDIA says an eight-system configuration provides 1TB of unified memory. The company describes the added capacity as supporting larger models, longer contexts, concurrent AI agents and distributed development. It is also intended to let small teams use a cluster as a shared backend, with developers running separate sessions against the same model.
NVIDIA Sync Model Launcher, due later this month, is designed to simplify setup on a single DGX Spark or a cluster. NVIDIA says it can download a model and container suited to the system configuration, connect the model to OpenCode and load DGX Spark Skills for tasks such as troubleshooting and device maintenance. An October software update adds a Wi-Fi Assistant for changing networks in appliance mode.
Local inference runs a model on hardware controlled by its user or organisation, rather than sending each request to a cloud-hosted model endpoint. A cluster can therefore give a development team a shared, locally managed resource, though actual performance depends on the model and workload being run.
DGX Station targets enterprise and robotics work
For larger workloads, NVIDIA lists DGX Station with 748GB of coherent memory and up to 20 petaflops of AI performance. The company says its systems are shipping from manufacturer partners. New integrations with enterprise IT management tools are intended to fit Station into organisations’ existing setup and management processes.
DGX Station for Windows is expected from system-manufacturer partners in December. NVIDIA says it will support manufacturing applications for design, simulation, validation and engineering automation, alongside locally run AI agents. The aim is to let engineering teams use AI models and tools within Windows-based workflows on the same system.
NVIDIA also released three open-source reference workflows for robotics on DGX Station. They use NVIDIA Isaac Lab, Isaac Lab-Arena and Cosmos to refine GR00T 1.7 models with reinforcement learning, train dexterous manipulation skills and post-train Cosmos world models for synthetic-data generation. These workflows provide starting points for development and evaluation on local compute; they are not, by themselves, reports of a finished robot deployment.