First Batch of NVIDIA's GB300 AI Chips Produced in the U.S. – TSMC's Arizona Fab Enters Mass Production Era

Market reports indicate that NVIDIA's (NVDA-US) first batch of GB300 AI GPUs, based on the Blackwell architecture, has recently completed production at TSMC's (TSM-US) Fab 21 in Arizona, USA, manufactured using the 4-nanometer process. This marks the first batch of GB300 AI chips to be mass-produced on U.S. soil, signifying that the United States now possesses the capability to produce advanced AI GPUs and further validating TSMC's Arizona facility as capable of high-end chip mass production.

However, industry sources note that the current operations in Arizona are limited to the wafer fabrication stage, with critical steps still required before the chips are ready for shipment.

Since the U.S. currently lacks CoWoS (Chip-on-Wafer-on-Substrate) advanced packaging capacity, these wafers must be air-freighted back to Taiwan, where TSMC will complete the CoWoS packaging process before handing them over to system manufacturers for AI module and server assembly. As a result, GB300 currently follows a 'manufactured in the U.S., packaged in Taiwan' production model, meaning the U.S. AI chip supply chain has not yet achieved full end-to-end localization.

Industry analysts suggest that advanced packaging is the final missing piece for the U.S. to establish a complete AI chip supply chain. Once CoWoS packaging capabilities are established in the U.S., the entire process—from wafer fabrication and advanced packaging to AI server assembly—could potentially be completed domestically.

To complete this crucial capability, TSMC is actively advancing its advanced packaging footprint in the U.S. According to plans, the company will invest $265 billion to build two advanced packaging plants in Arizona, creating an integrated production cluster encompassing both wafer manufacturing and advanced packaging, thereby further strengthening the U.S. semiconductor supply chain.

Market expectations suggest that once the CoWoS and SoIC (System on Integrated Chips) advanced packaging lines are fully operational, future NVIDIA Blackwell platforms and next-generation Rubin architecture AI chips could be produced entirely within the U.S., from wafer fabrication and advanced packaging to AI server assembly. This would significantly enhance supply chain autonomy and align with the U.S. government's policy of bringing semiconductor manufacturing back onshore.

In addition to packaging expansion, TSMC is also continuously enhancing its process technology in the U.S. According to company plans, it aims to introduce next-generation advanced processes such as N2 (2nm) and A16 (1.6nm) into mass production in Arizona by 2028, enabling its U.S. facilities to produce the most advanced chips at the same level as its Taiwan operations.

Market observers point out that as AI demand continues to grow rapidly, the U.S. government is actively promoting domestic semiconductor supply chains. With TSMC expanding its manufacturing and packaging footprint in the U.S., the country’s ability to independently produce AI chips will improve. The Arizona facility is also expected to become a key manufacturing hub for AI chip design companies like NVIDIA, further reducing geopolitical and supply chain risks.