Blackwell dominates, Rubin delayed
Nvidia’s Blackwell chips are now expected to dominate 2026 AI GPU shipments while next‑generation Rubin faces supply‑chain and packaging pressures that could delay deliveries. Analysts and industry reports say Blackwell will make up the bulk of this year’s GPU flow even as HBM4 and CoWoS packaging constraints threaten real delivery volumes. For buyers, predictability is slipping: demand is clear but getting chips on acceptable timelines and terms is the real bottleneck. (theregister.com, communicationstoday.co.in)
Nvidia’s newest artificial intelligence chips are selling so fast that the older “next big thing” is now the safer bet for 2026 deliveries. TrendForce says Blackwell is now expected to make up 71% of Nvidia’s high-end graphics processing unit shipments in 2026, up from an earlier 61% forecast. (trendforce.com) The surprise is that Rubin, the architecture Nvidia has been pitching as Blackwell’s successor, is losing share before it fully ramps. TrendForce cut Rubin’s expected 2026 share to 22% from 29%, while Hopper fell to 7% from 10%. (trendforce.com) This is not a demand problem. TrendForce says total high-end graphics processing unit shipments should still rise in 2026 because cloud companies are buying full Nvidia rack systems that pack in more chips per order. (trendforce.com, communicationstoday.co.in) The bottleneck is the part after the chip is made. Taiwan Semiconductor Manufacturing Company’s Chip on Wafer on Substrate packaging is the step that bolts the processor and high-bandwidth memory together on one advanced package, and it is built for exactly these giant artificial intelligence accelerators. (tsmc.com) Rubin leans harder on the newest memory generation. Nvidia’s Vera Rubin platform is built around high-bandwidth memory 4, and TrendForce says that memory, along with networking, power, and liquid-cooling constraints, is one reason Rubin’s 2026 flow is slipping. (nvidianews.nvidia.com, trendforce.com) Nvidia is still publicly talking like Rubin is ready. On March 16, 2026, the company said the Vera Rubin platform had seven new chips “in full production,” and its product page says a Vera Rubin NVL72 rack combines 72 Rubin graphics processors with 36 Vera central processors. (nvidianews.nvidia.com, nvidia.com) That creates a split between announcement and delivery. A chip can be in production on paper, but customers still need enough memory stacks, enough packaging slots, enough networking gear, and enough cooled racks to turn wafers into working server clusters. (tsmc.com, trendforce.com) Blackwell benefits because it is newer than Hopper but less exposed than Rubin to the hardest parts of the 2026 supply chain. TrendForce’s revised mix says buyers who want Nvidia capacity this year are more likely to get Blackwell than wait for a clean Rubin ramp. (trendforce.com) The practical effect is that the artificial intelligence chip market is no longer constrained mainly by who can design the fastest processor. In 2026, it is constrained by who can secure memory, packaging, power, and cooling at the same time. (trendforce.com, tsmc.com) So Nvidia can keep winning and customers can still lose time. The company may ship more high-end chips overall in 2026, but the hardest question for cloud providers is shifting from “Which chip is best?” to “Which chip can actually arrive, in a rack, on schedule?” (trendforce.com, theregister.com)
