Summary

• Siemens and NVIDIA have developed digital twin and AI applications for quality inspection preparation and production scheduling at the Erlangen electronics factory.
• Siemens reports potential engineering-effort reductions of up to 85%, time-to-market improvements of up to 40%, and planning cost savings of up to 20% in distinct workflows.
• Testing is taking place in a functioning factory, but wider deployment and fully autonomous production have not been demonstrated.

Siemens is testing artificial intelligence inside its electronics factory in Erlangen, Germany, using digital twins to prepare manufacturing processes before physical prototypes are available and AI agents to examine production schedules before changes reach the factory floor.

The applications have been developed with NVIDIA and are intended to reduce some of the engineering effort required to introduce new products, while helping manufacturers respond to changes in orders, equipment availability, and material supplies.

Siemens reports that virtual production preparation can reduce engineering work by up to 85% and shorten time to market by up to 40% for specified printed circuit board assembly processes. Separately, its agent-based production planning approach offers potential cost savings of up to 20%.

The figures describe different applications and are company-reported results or estimates, not independently audited improvements across the entire factory. Siemens and NVIDIA are developing and testing the technology in live production, but have not established that the Erlangen facility is operating without human supervision.

Inspection preparation begins before physical assembly

The Erlangen plant manufactures power electronics and automation equipment, including components used in industrial drives and frequency converters. Introducing a new printed circuit board assembly can require extensive preparation before production equipment is ready to manufacture and inspect it consistently.

Engineers must establish appropriate manufacturing conditions and configure inspection processes capable of identifying defects. Some of that work has traditionally depended on having physical components or prototypes available for examination.

Siemens and NVIDIA are using digital representations of new assemblies to begin inspection preparation earlier. NVIDIA Cosmos world models generate synthetic training data intended to help vision-based inspection systems recognise relevant product characteristics.

That allows engineers to develop and test parts of the inspection process before the first physical examples of a new assembly have been produced. The approach could reduce delays between design completion and manufacturing readiness, particularly where products are frequently modified or introduced in several variants.

Siemens attributes the potential 85% reduction in engineering effort and 40% improvement in time to market to this virtual production preparation process. Neither figure should be interpreted as a plant-wide reduction in staffing requirements or evidence that every product introduced at Erlangen will achieve comparable savings.

Synthetic training information also introduces an engineering requirement of its own. Digitally generated images and simulations must correspond sufficiently closely to the conditions encountered in a real factory, including variations in components, lighting, assembly tolerances, and defects.

A system performing well against simulated examples may still require adjustment when physical production begins. Validation against real components remains necessary to determine whether the inspection process detects relevant faults without generating excessive false alarms.

The advantage of the virtual approach lies in reducing the amount of preparation that must wait for physical prototypes, rather than removing the need for production engineering and quality assurance.

AI agents examine production alternatives

A second application concerns the planning decisions required to keep manufacturing operations running when orders, material supplies, or production requirements change.

Factories must balance available machines, staff, components, delivery commitments, and other constraints when deciding what should be produced and in which sequence. A change affecting one part of that schedule can alter the availability of equipment or materials elsewhere.

Siemens and NVIDIA have developed AI agents capable of generating alternative production plans using information about the factory’s actual operating conditions. Those plans can then be evaluated through a digital twin before being introduced into the physical production environment.

NVIDIA’s Factory Operations Blueprint supports the agent-based planning functions, while Omniverse libraries provide simulation capabilities for testing proposed arrangements against the factory model.

The arrangement allows production planners to consider different responses to changing circumstances without immediately committing equipment and employees to an untested schedule.

If a component becomes unavailable, for example, planning software could generate alternatives that account for other work, available materials, and the implications for delivery commitments. The resulting options would still need to be checked against the actual manufacturing constraints and business priorities.

Siemens associates potential planning cost savings of up to 20% with the approach. That estimate must remain distinct from the engineering and time-to-market figures associated with virtual inspection preparation.

The company has not published a comprehensive independent assessment showing that the planning technology produces equivalent savings across different production environments.

Factory data determines the quality of decisions

Industrial AI depends on more than the capabilities of an individual model. Production systems must provide sufficiently accurate information about machinery, inventory, work orders, product specifications, and operating conditions for automated recommendations to be useful.

An agent may generate a schedule that appears efficient in simulation but becomes impractical if it relies on outdated material records or fails to account for maintenance requirements.

Digital twins likewise depend on the quality of the information used to represent the physical environment. As equipment changes and manufacturing processes evolve, the virtual model must remain aligned with actual factory conditions.

Connecting those systems can be demanding in facilities containing equipment and software introduced over many years. Some machines expose detailed operating information through modern interfaces, while others require additional engineering work before their condition and availability can be used reliably in automated planning.

Siemens’ position as a supplier of industrial software, automation equipment, and manufacturing systems gives it access to several components of that integration process. NVIDIA contributes accelerated computing, AI models, and simulation technology, but the complete application still needs to work within established production processes.

Factories considering comparable systems would have to assess how much integration is required and whether the expected savings justify the investment. The economics will vary with the complexity of the products manufactured, the frequency of design changes, and the extent to which existing production information is already available.

A live factory provides a demanding test

Established in 1971, the Erlangen electronics factory employs approximately 1,000 people and produces equipment used in industrial and infrastructure applications. Siemens has used the facility to develop and test digital manufacturing technologies for more than a decade.

The partners announced a broader ambition in January 2026 to develop increasingly autonomous and adaptive manufacturing sites. The Erlangen applications represent practical work towards that objective rather than confirmation that fully autonomous factories have been delivered.

Testing in a functioning manufacturing environment provides evidence that the technologies can be developed around actual operating requirements. It also exposes them to the variability that controlled demonstrations may not reproduce, including production changes, equipment constraints, and conflicting priorities.

Even where software can generate alternative decisions, employees remain responsible for establishing whether those decisions are appropriate and whether a proposed change can be introduced safely.

European manufacturers have strong incentives to reduce delays in production preparation and make better use of existing facilities, particularly where capital expenditure, skilled labour, and energy costs place pressure on operating margins.

However, the commercial value of the Erlangen work will depend on repeatable results rather than the promise of autonomous production. That means assessing the engineering effort saved, reliability of inspection systems, quality of production plans, and additional costs incurred through integration and maintenance.

Siemens and NVIDIA have established specific manufacturing applications for further testing in Erlangen. Their wider significance will depend on whether the reported improvements can be reproduced in other factories with different equipment, products, and operating requirements.