Nvidia open-sources its factory agent blueprint, with Siemens as first anchor

Nvidia's European industrial pitch now has a repository attached to it: the Factory Operations Blueprint, or FOX, is available in early access on GitHub, giving developers a path to build always-on factory manager agents plus specialist agents for visual inspection, root-cause analysis, report generation and fleet management.

The design is deliberately connective. FOX-built agents can reach frontier models such as Cosmos and Nemotron and also foundational knowledge in CAD, PLM, simulation software and IT/OT systems, so they can watch operations in real time, correlate what a camera saw with what a machine logged, and coordinate a response.

Cosmos in-context learning is the detail that makes this practical: agents adapt from examples and documentation without fine-tuning. Factory floors rarely have a clean, labeled dataset waiting, and they cannot pause production for a training run.

Siemens is the named anchor customer. It is integrating FOX-built agents into Intelligence Center X to identify root causes and coordinate factory operations, with an estimated 20% cost saving. Its Process Preparation X agents use Cosmos Data Factory tools to generate defect images for circuit-board inspection, an approach Siemens says cut calibration time for inspection processes by 40% at its Erlangen electronics plant.

Around the factory, Nvidia pushed the same accelerated-computing argument into engineering. cuDSS is now generally available for the dense math behind structural, fluid, electromagnetic and semiconductor simulation; Mercedes-Benz runs Siemens Simcenter STAR-CCM+ on Nvidia GPUs to compare aerodynamic designs and uses Cosmos 3 with the Metropolis VSS blueprint for vehicle inspection and windshield assembly verification.

The hardware numbers are concrete: imec benchmarks S-Litho at up to 40x speedup, Infineon uses Sentaurus Device with CUDA libraries for a 12x speedup in 3D breakdown simulation of power devices, and Leonardo works with Neural Concept to bring aircraft geometry and aerodynamic results together in Omniverse libraries.

Digital twins fill out the picture. Audi built a battery-production twin on Omniverse libraries to test production concepts before equipment exists and, with Dassault Systemes and Agile Robots, validates humanoid and logistics flows in DELMIA twins; Schaeffler connects humanoid actuator twins to Isaac, and Wandelbots validates assembly workflows with NOVA on Isaac Sim.

For manufacturing, the real threshold is not model capability but whether an agent can earn the trust of both IT and OT at once. Schneider Electric and AVEVA are applying the method to AI infrastructure itself, building an AI factory twin for Nscale so engineers can compare power, cooling and equipment layouts before anyone touches a rack; whether FOX becomes a default starting point or just another blueprint depends on how reproducible the Siemens result proves to be.