TSMC says CoWoS demand is exploding — Nvidia has snapped up most of the capacity, it has been reported

The headline: packaging is suddenly front and center

Taiwan Semiconductor Manufacturing Co. told CNBC that its most advanced packaging technology, Chip on Wafer on Substrate (CoWoS), is growing at an eye-popping 80% compound annual growth rate, and the company is rapidly ramping capacity to keep pace. Paul Rousseau, head of packaging solutions for TSMC North America, told the network the numbers "are growing very substantially." It has been reported that Nvidia has reserved a majority of the most advanced CoWoS capacity — leaving the rest of the industry watching nervously.

Why this matters: it's not just glue and tape anymore

Advanced packaging is no longer an afterthought. As AI chips stack logic and high-bandwidth memory in tighter, three-dimensional layouts, packaging becomes the architecture that actually lets those pieces talk to each other at speed and efficiency. Rousseau called it "the natural extension of Moore's Law into the third dimension." Short on packaging capacity, and fabs can churn out silicon that simply can't be turned into usable AI hardware fast enough. Bottleneck? You bet.

The scramble: who’s building what, and where

TSMC is expanding — new packaging sites in Taiwan and two facilities planned in Arizona — and it has allegedly outsourced some simpler steps to specialists such as ASE and Amkor to keep throughput up. ASE says advanced-packaging sales will double in 2026 and is building big new capacity in Taiwan; rival SPIL recently opened a large site that Nvidia’s Jensen Huang attended. Intel, meanwhile, is pitched as technologically competitive and already packages for customers like Amazon and Cisco; it was also reportedly tapped by Elon Musk to help package chips for SpaceX, xAI and Tesla at his planned Terafab. For now, TSMC still ships 100% of wafers back to Taiwan for packaging — even chips made at its Phoenix fab — so the question is simple: can U.S. and partner capacity scale fast enough to keep the AI hardware race from tripping over its own shoelaces?

Sources: cnbc.com