TSMC Arizona Produces First GB300 AI Chips

According to market sources, NVIDIA’s first batch of GB300 AI GPUs based on the Blackwell architecture has recently completed wafer production at TSMC’s Fab 21 semiconductor facility in Arizona, United States. Manufactured using TSMC’s 4nm process technology, these chips represent the first GB300 AI processors to be produced on U.S. soil, marking a major milestone in the country’s efforts to establish advanced AI semiconductor manufacturing capabilities.

The production milestone also highlights TSMC’s growing manufacturing capabilities in the U.S. However, industry observers noted that the Arizona facility has currently completed only the wafer fabrication stage. Several critical steps remain before the chips can become finished AI products ready for shipment.

Since the U.S. has not yet established large-scale advanced CoWoS packaging capacity, the completed wafers still need to be transported back to Taiwan for TSMC’s advanced packaging process. After CoWoS packaging, the chips will be integrated by system manufacturers into AI modules and servers. As a result, GB300 production currently follows a model of “wafer manufacturing in the U.S. and advanced packaging in Taiwan,” meaning the AI semiconductor supply chain has not yet achieved full localization.

Industry analysts believe that advanced packaging remains the final key component needed for the U.S. to build a complete AI chip supply chain. Once CoWoS packaging capabilities are established domestically, the entire production process — from wafer fabrication and advanced packaging to AI server assembly — could potentially be completed within the United States.

To address this gap, TSMC is accelerating its advanced packaging expansion in the U.S. The company plans to invest approximately $265 billion to build two advanced packaging facilities in Arizona, creating an integrated semiconductor manufacturing ecosystem that combines advanced wafer production with next-generation packaging technologies.

Market expectations suggest that once CoWoS and SoIC advanced packaging lines begin operations in the U.S., future NVIDIA Blackwell platforms and next-generation Rubin AI chips could achieve localized production covering wafer manufacturing, advanced packaging, and AI server assembly. This would significantly strengthen supply chain resilience and align with U.S. efforts to bring semiconductor manufacturing back domestically.

Beyond packaging investments, TSMC is also advancing its U.S. process technology roadmap. The company plans to introduce next-generation nodes, including 2nm and 1.6nm process technologies, to Arizona production lines by 2028, allowing its U.S. facilities to manufacture some of the world’s most advanced chips alongside its Taiwan operations.

Industry analysts said that as demand for AI infrastructure continues to accelerate, the U.S. government’s push for semiconductor supply chain localization, combined with TSMC’s expanding manufacturing and packaging footprint, will help strengthen domestic AI chip production capabilities. In the future, TSMC’s U.S. facilities could become a critical manufacturing hub for AI chip designers such as NVIDIA while reducing supply chain and geopolitical risks.