Packaging is the new chip choke

Nvidia has grabbed most of Taiwan Semiconductor Manufacturing Company’s top chip-packaging slots, so the hard part of making an artificial intelligence chip is no longer just printing the silicon wafer. The new wait is the assembly step that bolts the processor to its memory and wiring so the whole thing can actually run inside a server. (cnbc.com) That assembly step is called advanced packaging, and it matters because today’s artificial intelligence chips are too big and too power-hungry to live as one neat piece of silicon. Companies now split the job across several pieces and then reconnect them inside one package at extremely short distances. (tsmc.com) Taiwan Semiconductor Manufacturing Company’s best-known method is called Chip-on-Wafer-on-Substrate, which is a stack-up where logic chips and memory chips sit together on a dense base before the whole module is mounted onto a substrate. Taiwan Semiconductor Manufacturing Company says this method is built for artificial intelligence and supercomputing parts that need huge bandwidth between processor and memory. (tsmc.com) That is why packaging has become a choke point for Nvidia’s H100, H200, and Blackwell-class systems: the package is no longer a plastic shell, but part of the computer itself. If the package cannot be built, the finished graphics processing unit cannot ship even if the wafer is already done. (cnbc.com) (trendforce.com) The immediate problem is concentration. CNBC reported on April 8 that almost all of this advanced packaging capacity still sits in Asia, while Nvidia has reserved the majority of Taiwan Semiconductor Manufacturing Company’s most advanced capacity. (cnbc.com) Taiwan Semiconductor Manufacturing Company is expanding, but demand is rising almost as fast as new lines can be built. Paul Rousseau, the company’s North America packaging head, told CNBC that Chip-on-Wafer-on-Substrate demand is growing at about an 80% compound annual growth rate. (cnbc.com) That scramble is pushing customers toward Intel, which has spent years building its own advanced packaging tools instead of relying only on wafer production. Intel’s main options are Embedded Multi-die Interconnect Bridge and Foveros, two ways of linking multiple chip pieces inside one product with shorter paths than a traditional circuit board can provide. (arstechnica.com) (cnbc.com) The reason Intel is suddenly getting calls is not that rivals prefer its design tools; it is that spare packaging capacity has become strategic inventory. Reports this week said Google and Amazon were in talks with Intel for packaging services as companies look for any route around Taiwan Semiconductor Manufacturing Company’s fully booked lines. (trendforce.com) (cnbc.com) The bottleneck does not stop at the packaging machine. DigiTimes reported on April 8 that integrated circuit substrates are tightening too, which means even companies that find packaging capacity can still get stuck waiting for the thin layered base that carries signals and power under the chip package. (digitimes.com) Those substrates depend on their own upstream materials, including copper-clad laminates, prepregs, and high-end fiberglass cloth, so one shortage can spill into the next. Industry reporting over the past year has shown the artificial intelligence build-out pulling those materials tight as Chip-on-Wafer-on-Substrate volumes rise. (globalsmtasia.com) (globalsmt.net) That is why new wafer fabs in Arizona or elsewhere do not solve the whole problem by themselves. Several reports say wafers made in the United States still often need to go back to Taiwan for advanced assembly and test, which leaves the supply chain exposed to the same packaging pinch even after billions of dollars of fab spending. (cnbc.com) (trendforce.com) The result is a semiconductor shortage that has moved one step downstream. In 2021 the scarce thing was wafer capacity; in 2026 the scarce thing is the ability to turn finished dies, memory stacks, and substrates into one working artificial intelligence module at scale. (cnbc.com) (tsmc.com)