TSMC Flies Arizona Chips to Taiwan for Packaging, Exposing U.S. Supply Chain Gaps

The United States’ efforts toward semiconductor self-sufficiency continue to face critical challenges. According to a new report from Taiwan Economic Daily, TSMC's Arizona-fabricated wafers are being flown back to Taiwan for packaging, underscoring the persistent reliance on Taiwan in the advanced chip supply chain—especially for AI server applications.

The reason? TSMC lacks sufficient advanced packaging capacity in the U.S., particularly for high-end processes like CoWoS (Chip-on-Wafer-on-Substrate), which are essential for AI accelerators used by companies such as NVIDIA, AMD, and Apple. To meet soaring demand, especially from the AI sector, TSMC is air-shipping processed wafers to Taiwan, where they undergo final packaging and assembly before being delivered to customers.

With AI demand surging globally and 3nm supply tight, TSMC’s Taiwanese facilities are working at full capacity. To ease production bottlenecks, the company has shifted some front-end fabrication to its Arizona fab, but must still rely on Taiwan for back-end processes like packaging and testing.

Despite TSMC’s pledged $165 billion U.S. investment, including future plans to build advanced packaging capabilities in America, meaningful progress remains slow. Industry insiders note that air cargo transport, while costly, is currently more feasible and faster than waiting for new packaging infrastructure to come online.

Meanwhile, U.S. policy is making strides. Government initiatives aim to fulfill over 50% of domestic chip demand by 2032, with support from incentives and tariffs introduced under President Trump’s “chip-first” policies. TSMC has also announced plans to introduce 1.6nm (A16) chip production in the U.S., suggesting that domestic capabilities will eventually catch up.

Still, the current situation underscores a key reality: the U.S. semiconductor supply chain continues to depend on Taiwan, especially for cutting-edge chip packaging—a critical bottleneck in the AI race.