Foxconn Reports 2.48 Times Faster Model Training on DGX SuperPOD

Foxconn says its DGX SuperPOD trained 120-billion-parameter models 2.48 times faster than its H100 baseline, while supporting AI work across the electronics maker’s factory operations.

Foxconn’s benchmark used DGX GB200 systems built on NVIDIA’s Blackwell architecture. The company compared training performance with an H100-based baseline and counted one trillion tokens processed over 12 days using training-job telemetry.

“By utilizing the NVIDIA Blackwell-based DGX SuperPOD, we achieved a 2.48x training speedup, enabling us to process one trillion tokens in just 12 days to power our global AI factory,” said Yung-Hui Li, founding director of AI Research Center at Hon Hai Research Institute.

Foxconn says its factory work has reduced plant scheduling labor hours by 25% and cut engineering work for injection-molding process changeovers by 50%. On a surface-mount technology line, the company says its work addressed more than 20 major defect root causes, with a correction rate above 90%.

“The DGX SuperPOD serves as our AI ‘Super Brain,’ driving agentic workflows that have delivered significant efficiency improvement and a 25% reduction in production scheduling hours,” said Kathy Yang, Foxconn’s head of AI software business.

FoxBrain, Foxconn’s Traditional Chinese language model, was announced in March 2025 and initially designed for the company’s internal systems. Foxconn’s wider production environment also includes GB300 NVL72 systems and RTX PRO servers, distinct from the DGX GB200 setup used for the cited training results.

In May 2025, Foxconn announced a partnership with NVIDIA to build an AI supercomputing center in Taiwan, initially describing Blackwell Ultra systems that included GB300 NVL72 racks. The 2.48-times training comparison concerns the DGX GB200 configuration.

Foxconn has also launched commercial computing capacity through its Visionbay.ai unit. In July 2026, the unit began commercial operation of a separate 5-megawatt cluster using NVIDIA HGX B300 systems. In its July announcement, Foxconn said utilization was above 90%.