Blackwell, memory and the supply squeeze

Nvidia’s next graphics processor was supposed to start taking more of the load in 2026, but analysts now think the current Blackwell line will do even more of the work instead. TrendForce cut Rubin’s expected share of Nvidia’s high-end graphics processor shipments for 2026 to 22% and raised Blackwell’s share to 71%. (trendforce.com) That shift is not just about one chip arriving late. Nvidia now sells whole rack systems, and its flagship GB200 NVL72 packs 72 Blackwell graphics processors into one liquid-cooled cabinet that Nvidia says behaves like a single giant machine for training and serving artificial intelligence models. (nvidia.com) Those racks need a special kind of memory called high bandwidth memory, which is built like a stack of tiny memory floors sitting right next to the processor instead of across the room on a motherboard. Samsung says its HBM3E version can move up to 1,180 gigabytes per second and hold up to 36 gigabytes per stack. (semiconductor.samsung.com) Blackwell uses that memory now, while Rubin is expected to move to the next step, high bandwidth memory 4, and that is where the supply chain gets tight. TrendForce said Rubin faces delays from high bandwidth memory 4 validation, new ConnectX-9 network interface cards, higher power draw, and tougher liquid-cooling requirements. (trendforce.com) So the bottleneck is no longer only the graphics processor die coming out of Taiwan Semiconductor Manufacturing Company’s fabs. It is also the memory stacks, the advanced packaging that bonds them to the processor, and the rack hardware around them. (trendforce.com) (nvidia.com) Memory makers have been signaling this squeeze for months. SK hynix said in its January 5, 2026 market outlook that demand for HBM3E and HBM4 will drive an “AI memory supercycle” as data-center buildouts expand. (news.skhynix.com) That creates an odd result for customers. If Rubin slips, cloud companies and model developers do not stop buying; they buy more Blackwell systems, which still need huge amounts of HBM3E memory right now. (trendforce.com) (micron.com) It also creates an odd result for suppliers. A Rubin delay can hurt near-term plans around high bandwidth memory 4, but it can keep factories for Blackwell-era high bandwidth memory 3E running flat out because Blackwell volumes rise instead of falling. (trendforce.com) (news.skhynix.com) Nvidia has been pushing customers toward bigger integrated systems because that is where performance jumps most sharply. The company says GB200 NVL72 delivers an exascale computer in one rack, and its Blackwell architecture page says the platform is now in full production. (nvidia.com 1) (nvidia.com 2) That means the 2026 fight is not simply Blackwell versus Rubin. It is Blackwell plus enough HBM3E stacks, enough packaging capacity, enough liquid cooling, and enough rack integration to turn thousands of chips into working data centers on schedule. (trendforce.com) (nvidia.com)