Foxconn and NVIDIA Boost AI Training with DGX SuperPOD

Highlights

Foxconn (鴻海, 2317-TW) and NVIDIA (輝達, NVDA-US) are working together to advance AI training and deployment. Using NVIDIA DGX SuperPOD and DGX GB200 systems, Foxconn trained its internally developed FoxBrain large language model. A model with 1200 billion parameters trained 2.48 times faster than with the company’s previous H100-based solution, processing 1 trillion Tokens in 12 days. Foxconn is also applying AI to factory scheduling, production processes, quality analysis, and digital-twin simulations. The collaboration links accelerated model training with practical AI applications across manufacturing operations.

Sentiment Analysis

• The article has a predominantly positive tone, emphasizing faster AI training, broader computing capacity, and operational improvements. It reports a 2.48-fold speed increase for the 1200-billion-parameter FoxBrain model and processing of 1 trillion Tokens in 12 days.
• Its positive framing also rests on reported factory outcomes: automated production scheduling reduced scheduling work hours by 25%, while AI-assisted injection-molding process changes cut related engineering workload by 50%. The article further states that more than 20 major SMT defect root causes were addressed and the abnormality improvement rate exceeded 90%.
• The overall sentiment is nevertheless mostly factual and company-focused. The source describes capabilities and reported results but provides no independent validation, comparison methodology, or discussion of implementation costs and limitations. Its claims should therefore be understood as reported company outcomes rather than an external assessment.

Article Text

Foxconn (鴻海, 2317-TW) and NVIDIA (輝達, NVDA-US) are collaborating on AI infrastructure and applications. Foxconn has introduced NVIDIA DGX SuperPOD and DGX GB200 systems to accelerate training of FoxBrain, its internally developed large language model. According to the company’s reported results, training a model with 1200 billion parameters was 2.48 times faster than under its previous solution based on H100 systems. Foxconn also processed 1 trillion Tokens in 12 days, supporting continued development of the model and its capabilities.

The DGX SuperPOD combines NVIDIA Blackwell architecture with NVLink switches to bring together computing resources from multiple GPUs. This configuration is intended to accelerate the training of large AI models. Foxconn uses the infrastructure to train FoxBrain and is also working to combine industrial data with multimodal reasoning. The company’s stated direction includes developing vision-language models that can be applied to manufacturing, electric vehicles, and smart cities.

The reported training performance provides a central measure of the collaboration. Foxconn says the 1200-billion-parameter model trained 2.48 times faster with DGX SuperPOD than with its earlier H100-based solution. The system processed 1 trillion Tokens over 12 days, providing computing capacity for ongoing FoxBrain training and development of multimodal functions. These figures describe Foxconn’s reported results; the source does not provide an independent benchmark methodology or additional comparison details.

Foxconn is also applying AI to production management. Its automated production scheduling system, known as APS, is reported to have reduced factory scheduling work hours by 25%, lowering the labor burden associated with production planning. In injection molding, AI is used to optimize process-setting parameters. The company says this reduced engineering workload associated with process changes by 50%, with the aim of improving production-flow efficiency.

In production-line quality management, Foxconn uses AI to analyze issues on SMT lines. The company reports that it has addressed root causes associated with more than 20 major defects, with an abnormality improvement rate exceeding 90%. AI model analysis and real-time decision-making are intended to help identify the causes of production-line irregularities and improve the speed and effectiveness of follow-up action. These examples illustrate the company’s effort to extend AI from computing infrastructure into factory operations.

For managing AI resources, Foxconn uses NVIDIA Mission Control to coordinate AI workloads across more than 200 factory sites globally. The company also combines NVIDIA Run and NIM microservices to support flexibility in deploying AI models. Separately, NVIDIA RTX PRO servers are used to support factory digital-twin simulations and AI-agent workflows. Through NVIDIA Omniverse and Isaac, Foxconn simulates robots and production environments, supporting line validation and replication as it works to introduce AI technology at manufacturing locations around the world.

Foxconn says it will continue working with NVIDIA to develop applications in smart manufacturing, smart electric vehicles, and smart cities. Its plans also include an AI Token ecosystem and an open agent framework intended to help enterprises develop customized AI agents. Together, the reported training results, factory-process applications, workload management, and simulation tools describe a broad approach: building AI capacity while adapting models and systems for use in industrial settings. The source outlines the company’s objectives and reported outcomes but does not specify timelines for future deployments or provide independent assessments of their impact.

Key Insights Table