Mastodawn

undefined | AI's next bottleneck: Why even the best chips made in the U.S. take a round trip to Taiwan

Advanced packaging – the process that connects multiple dies into a single, larger chip – is rapidly becoming the next choke‑point for artificial‑intelligence hardware. Historically performed in Asia, the step is now in short supply as AI workloads demand ever‑higher density, performance and efficiency. TSMC’s most advanced method, Chip‑on‑Wafer‑on‑Substrate (CoWoS), is growing at an 80 % compound‑annual rate, and Nvidia has booked the bulk of the available capacity. Intel, which already does much of its final packaging in Vietnam, Malaysia and China, is expanding its U.S. footprint with facilities in New Mexico, Oregon and Arizona, and is courting customers such as Amazon, Cisco, SpaceX and Tesla for custom‑chip packaging.

Both TSMC and Intel are racing to supply the U.S. market. TSMC is constructing two new packaging plants in Arizona and two more in Taiwan, while still routing 100 % of its chips to Taiwan for final assembly. Intel’s embedded multi‑die interconnect bridge (EMIB) offers a cost‑effective alternative to CoWoS, and its recent tour of an Arizona packaging fab showed customers already using the service. The proximity of packaging to fab sites promises shorter turnaround times and lower logistics risk, a benefit that especially appeals to large AI investors like Nvidia and Elon Musk’s upcoming Terafab projects.

Looking beyond 2.5‑D, the industry is already developing true 3‑D solutions. TSMC’s System‑on‑Integrated‑Chip (SoIC) and Intel’s Foveros Direct aim to stack dies vertically, creating a single “chip‑like” entity with dramatically higher performance. Memory makers such as Samsung, SK Hynix and Micron are also advancing 3‑D packaging for high‑bandwidth memory, while hybrid bonding techniques replace traditional bumps with copper pads to further shrink interconnect distances. These innovations, together with the surge in advanced‑packaging capacity, will shape the next wave of AI hardware and determine whether the supply chain can keep up with the exploding demand for compute.