Packaging is the chokepoint

The chip shortage did not end when factories caught up on making silicon wafers. It moved one step later, into “advanced packaging,” the stage where a finished compute chip and its memory are fused into one high-speed module that can actually go into an artificial intelligence server. (cnbc.com) Taiwan Semiconductor Manufacturing Company says its Chip on Wafer on Substrate process, the packaging method used for many top artificial intelligence chips, is growing at an 80% compound annual rate. Nvidia has already reserved most of that top-tier capacity, according to a rare CNBC interview with TSMC’s North America packaging head Paul Rousseau. (cnbc.com) That step matters because modern artificial intelligence chips are no longer one slab of silicon. TSMC says its Chip on Wafer on Substrate design uses a silicon interposer, which works like a dense wiring board, to connect the main processor to multiple stacks of high-bandwidth memory in one package. (tsmc.com) The United States is trying to build more of that capability in Arizona, where TSMC is expanding and Intel is pushing its own advanced packaging services. The catch is that new American capacity does not automatically mean broad access if one buyer is large enough to pre-book most of the slots. (cnbc.com) (tsmc.com) That is why Nvidia’s next product cycle now looks less like a pure design race and more like a logistics race. The Register, citing TrendForce, reported on April 8 that Rubin chips may ship later and in smaller volumes than expected, with Rubin now forecast at 22% of Nvidia’s high-end graphics processing unit mix in 2026, down from 29%. (theregister.com) One pressure point is high-bandwidth memory, which is the stacked memory sitting right next to the processor so data does not have to travel far. Rubin is expected to use fourth-generation high-bandwidth memory, and TrendForce said validation of that newer memory is one reason the rollout could slip. (theregister.com 1) (theregister.com 2) Another pressure point is the substrate, which is the base layer under the package that routes signals and power between the chip and the rest of the system. Reports around Rubin have highlighted Ajinomoto build-up film substrates, a specialized material used in those dense layers, as one of the parts that can cap output even when the processor die itself is ready. (theregister.com) The networking side is also part of the bottleneck. TrendForce pointed to the move to Nvidia’s faster ConnectX-9 network interface cards, plus higher power draw and tougher liquid-cooling needs, as reasons a finished Rubin server could be harder to validate than a finished Rubin chip. (theregister.com) That supply risk is pushing buyers back toward Blackwell, the current generation, because customers can only deploy what manufacturers can assemble in volume. When the next platform looks uncertain, the current platform gets pulled harder, and packaging lines stay full for longer. (theregister.com) (cnbc.com) Nvidia is also selling software as part of the hardware package now. Its Mission Control platform is designed for Blackwell rack-scale systems and handles scheduling, monitoring, and automated recovery, which means Nvidia is not just shipping chips but the operating layer for scarce clusters. (nvidia.com) (developer.nvidia.com) That changes what “capacity” means in 2026. A customer needs wafers, memory, substrates, interconnects, cooling gear, packaging slots, and now software that understands how a Blackwell rack is wired, or the server can sit half-used even after delivery. (developer.nvidia.com) (docs.nvidia.com) So the real choke point is no longer the clean room where the chip is etched. It is the crowded handoff where silicon, memory, networking, cooling, and control software all have to arrive at the same time, and right now Nvidia is big enough to absorb a large share of that queue. (cnbc.com) (theregister.com)