Siemens and NVIDIA have moved their industrial artificial-intelligence partnership into two practical factory workflows: preparing visual inspection before a physical product exists, and evaluating production schedules in a digital model before changing the live line. Siemens presented the work at its electronics factory in Erlangen, Germany, on October 19, 2026.
The demonstrations concern electronics manufacturing rather than a new CNC machine tool. Erlangen produces industrial drives and automation hardware, including printed circuit board (PCB) assemblies. Yet the engineering problem is familiar to precision manufacturers: launching new variants and revising capacity plans can involve extensive manual preparation, even when machining or assembly equipment is already automated.
Training inspection before the first physical board
In the October 19 announcement, Siemens said its engineers used digital representations of new PCB assemblies to generate photorealistic images for training automated visual-inspection systems. The aim is to remove a dependency that traditionally delayed quality-control setup: waiting for physical prototype assemblies before teaching vision equipment to recognize the intended design.
The workflow combines production-design information with NVIDIA Cosmos world-model technology to generate training material. That can bring some inspection preparation forward into the engineering phase, but it does not by itself establish whether synthetic images represent every real surface finish, lighting condition, solder defect or dimensional variation that an inspector must detect. Validation still requires physical production evidence.
Siemens reported that fully virtual production preparation can reduce inspection-related engineering effort by up to 85% and time to market by up to 40% for the use case described. These are the company's stated potential gains, not a publicly documented independent multi-factory comparison. The announcement does not disclose the complete baseline, test population or repeatability of those percentages across different product families.
Simulating the order queue before replanning the factory
A second application addresses orders that arrive after production plans are established. Siemens says an AI agent can prepare alternative schedules based on capacity, materials, due dates and line availability, then show planners how the choices affect defined performance indicators. Before execution, the proposed scenarios are assessed against a digital twin of the factory rather than directly applied to physical operations.
This separation between recommendation and execution matters. A scheduling agent may identify feasible alternatives more quickly, but its output depends on the accuracy of tool, labor, material and equipment data as well as the constraints encoded in the factory model. A plan accepted in simulation can still be undermined by unavailable components, maintenance interruptions or quality problems not represented in the inputs.
Siemens said the planning approach could enable cost savings of up to 20%. The release did not present a detailed independent cost model or specify which production configurations reach that figure. The claimed improvement therefore cannot be generalized to a CNC job shop with different changeover times, part mix or bottleneck equipment.
NVIDIA described its own participation in the October industrial-AI demonstration, including the use of its Factory Operations Blueprint and Omniverse simulation technologies. Siemens said its Erlangen operations already run more than 100 AI algorithms across manufacturing functions, but the existence of those deployments does not independently verify the new application's savings claims.
The factory use case illustrates a specific way to keep people and plant controls between an AI proposal and an operational change. For machining organizations exploring similar systems, the transfer question is whether production data, tooling constraints and quality feedback are reliable enough for a digital plan to be meaningful—not whether the Erlangen percentages can be copied directly.
