NVIDIA's Feynman platform is expected to go into mass production in 2028, with TSMC's A16 and SoIC expanding production simultaneously

According to DIGITIMES, NVIDIA's next-generation AI platform Feynman aims for mass production in the second half of 2028, utilizing TSMC's A16 process and incorporating SoIC 3D stacking and CPO technology. The Rubin generation focuses on integrating multiple small chips with HBM, while Feynman further develops towards 3D small chips, SoIC, and CPO. The report indicates that TSMC is accelerating the construction of related production capacity, with the original plan for SoIC monthly capacity to reach 20,000 pieces by the end of 2026, now updated to a target of 50,000 pieces by the end of 2027.

The construction progress of TSMC's Chiayi AP7 and Tainan AP8 is also accelerating, with AP7 planned in 8 phases. P1 and P2 have entered installation, P3 and P4 have obtained construction permits, and P5 to P8 are still in planning. In addition to mass-producing Apple-specific WMCM, the factory will configure SoIC, CoPoS, and CPO production lines according to customer demand. NVIDIA has rapidly shifted its R&D and supply chain resources to the Feynman platform, and the spillover effect of TSMC's advanced packaging orders continues to expand, with Siliconware becoming the main testing and packaging factory for NVIDIA's CoWoS and CPO orders. The NVLink interconnect bandwidth of NVIDIA's Feynman platform is expected to exceed 1 PB/s, further improving from Rubin Ultra's 520 TB/s. TSMC's COUPE technology integrates EIC and PIC in 3D through SoIC to form a light engine, advancing optoelectronic conversion from traditional circuit boards to the internal packaging structure. Since 2026, NVIDIA has invested over $40 billion in building the AI ecosystem, covering areas such as models, wafer manufacturing, data centers, and optical communications.