Foxconn Reports 2.48x Faster AI Training in NVIDIA-Based Self-Learning Factories

Foxconn reports 2.48 times faster training performance for 120-billion-parameter models on its new NVIDIA DGX SuperPOD infrastructure compared with its previous H100-based systems. The electronics manufacturer disclosed the figure as it detailed an AI factory programme that now spans more than 200 manufacturing facilities, with AI agents deployed on NVIDIA software supporting real-time operational decisions.

The company says the new infrastructure processed one trillion tokens in 12 days. All performance claims come from Foxconn; the company did not publish independent verification or specify the test configuration behind the training comparison.

What sits at the core of the AI factory?

FoxBrain, Foxconn's proprietary large language model, anchors the programme. The company adapted open foundation models using its own industrial data. FoxBrain supplies the intelligence for specialist platforms including Genesis, aimed at manufacturing applications, and CityGPT, which serves urban services.

The stated objective is a self-learning manufacturing ecosystem: AI systems analyse operational data, identify problems and support decisions across increasingly complex production environments. Foxconn develops the ecosystem with NVIDIA, extending it beyond manufacturing into electric vehicles and smart city applications.

Where has AI delivered measurable gains?

Foxconn reports three concrete results from live production use:

• Production scheduling shows a 25 percent reduction in labour hours.
• Engineering workload for injection-moulding process changeovers has fallen by 50 percent.
• On surface-mount technology (SMT) lines, AI systems have addressed more than 20 major defect root causes, with a correction rate exceeding 90 percent.

The SMT result matters most for test and quality engineers. The systems go beyond conventional automation and data analysis: they identify recurring problems, run root-cause investigation and trigger corrective action. For high-volume electronics production, continuous analysis of production data shortens the time needed to resolve quality issues and improves consistency.

These figures are vendor-reported and come from Foxconn's own operations, not from a standardised benchmark.

How do digital twins fit the deployment?

Foxconn uses NVIDIA RTX PRO servers, Omniverse and Isaac technologies to simulate factories and robotic operations before physical deployment. Engineering teams model production lines, equipment and robotic systems virtually, then assess configurations, optimise workflows and validate automation concepts before committing hardware.

The approach targets a specific constraint in Foxconn's business: replicating production processes across global facilities demands consistent engineering standards and reliable deployment procedures. Combined simulation and AI-driven analysis aims to reduce implementation risk and speed the transfer of manufacturing capabilities between sites.

Does the AI factory become a product?

Yes, through subsidiary Visionbay.ai, an NVIDIA Cloud Partner. Foxconn uses NVIDIA-Certified Systems to sell AI-as-a-Service to enterprises and government organisations across Asia. The move turns internal infrastructure into external revenue rather than confining deployment to Foxconn's own network.

What does this mean for measurement and metrology?

The programme points to a structural shift in how quality data gets used. AI-driven defect analysis and root-cause identification can improve manufacturing quality, but their effectiveness depends on reliable measurement data, consistent inspection processes and the ability to link detected deviations to corrective manufacturing actions.

The longer-term direction Foxconn describes would make inspection and measurement data an integral part of a continuously learning production environment, feeding decisions across the whole manufacturing lifecycle. Test engineers should read the 90 percent correction rate as evidence that this integration has started, not as a certified capability.

Foxconn's programme shows large-scale AI infrastructure moving from experiments into industrial deployment, with the potential to reshape how factories are designed, operated and optimised.

The open question for adopters: will AI root-cause analysis hold up under audit, or will verification labs need new measurement traceability standards to trust it?

By Grace Kim