Blackwell dominates; packaging is the choke-point
Analysts now expect Nvidia’s Blackwell GPUs to make up the lion’s share of high‑end AI shipments this year, shifting demand while Rubin delays loosened expectations for next‑gen parts. At the same time Nvidia has reserved much of TSMC’s advanced packaging capacity, and industry observers warn that packaging—how dies and memory are integrated—may become the critical bottleneck for accelerator supply. That means chip roadmaps are being constrained by assembly capacity as much as by the GPU designs themselves. (theregister.com (cnbc.com))
The bottleneck in artificial intelligence chips is no longer just making the silicon. Nvidia has locked up most of Taiwan Semiconductor Manufacturing Company’s top-end packaging capacity, so the slow step is increasingly the factory work that bolts the chip pieces together. (cnbc.com) Packaging is the stage where a bare chip stops being a fragile slice of silicon and becomes a usable part. For an artificial intelligence accelerator, that means mounting the processor next to high-bandwidth memory, wiring them together at very short distances, and putting the whole assembly on a substrate that can plug into a server board. (cnbc.com) Taiwan Semiconductor Manufacturing Company’s main method for this is called Chip on Wafer on Substrate. The short version is that several chip pieces and memory stacks are attached in one tightly packed module, which is why this step matters more for artificial intelligence hardware than for an ordinary laptop processor. (cnbc.com) Nvidia needs a lot of that capacity because Blackwell is now expected to dominate its premium shipments this year. TrendForce said on April 8 that Blackwell should account for 77 percent of Nvidia’s high-end graphics processing unit shipments in 2026, while Rubin was cut to 22 percent from an earlier 29 percent forecast. (theregister.com) That Rubin cut was not about demand disappearing. TrendForce said the next system is running into practical delays around validating new high-bandwidth memory generation 4, moving to Nvidia’s ConnectX-9 network cards, handling higher power draw, and meeting tougher liquid-cooling requirements. (theregister.com) So customers that wanted to wait for Rubin are being pulled back toward Blackwell systems that can ship sooner. That shift pushes even more orders into the same packaging lines that Blackwell already depends on. (theregister.com (cnbc.com)) Taiwan Semiconductor Manufacturing Company says demand for this assembly step is exploding. Paul Rousseau, the company’s North America packaging head, told CNBC that Chip on Wafer on Substrate capacity is growing at an 80 percent compound annual rate. (cnbc.com) Even that growth is not enough if one customer reserves most of the best lines. CNBC reported that Nvidia has booked the majority of Taiwan Semiconductor Manufacturing Company’s most advanced packaging capacity, which leaves less room for rival accelerator makers and custom chip projects. (cnbc.com) This is why the supply story has changed shape. A company can have leading chip designs, access to the latest wafer process, and enough memory on paper, but still miss shipments if it cannot get enough advanced packaging slots at the end of the line. (cnbc.com (theregister.com)) It also helps explain why Intel is suddenly talking up packaging as a competitive weapon. CNBC reported that Intel is technologically on par with Taiwan Semiconductor Manufacturing Company in advanced packaging, even though Taiwan Semiconductor Manufacturing Company remains the volume leader. (cnbc.com) Nvidia’s own roadmap now looks less like a clean yearly handoff and more like traffic merging into one crowded lane. Blackwell is taking the bulk of 2026 volume, Rubin expectations have been trimmed, and the real limit is no longer just chip design but how many finished modules the packaging plants can actually assemble. (theregister.com (cnbc.com))
