Compute is the choke point

The shortage is not for ideas. It is for the metal, memory, and cooling gear needed to turn artificial intelligence models into running machines, and TrendForce now expects Nvidia’s Blackwell family to take more than 70 percent of its high-end artificial intelligence chip shipments in 2026 while Rubin slips to about 22 percent. (trendforce.com) (cloudnews.tech) A high-end artificial intelligence server is not one chip in a box. Nvidia’s GB200 NVL72 system packs 72 Blackwell graphics processors into one liquid-cooled rack and links them with 130 terabytes per second of internal chip-to-chip bandwidth, which is why each delayed component can stall a whole system. (nvidia.com) The part creating the bottleneck is called high-bandwidth memory, which is memory stacked like a tower directly beside the processor so data can move much faster than it can from ordinary server memory. Rubin is built for the next version, called high-bandwidth memory 4, and Nvidia’s own platform description says Rubin pairs that memory with its new Vera central processor and ConnectX-9 network card. (developer.nvidia.com) TrendForce said on January 8 that Nvidia changed Rubin’s memory target in the third quarter of 2025 and pushed required speed above 11 gigabits per second per pin. That forced Samsung, SK hynix, and Micron to revise designs and resubmit samples, which moved broad high-bandwidth memory 4 production to late first quarter or early second quarter of 2026. (trendforce.com) By February 13, TrendForce was saying certification for those three suppliers was still expected in the second quarter of 2026. It also said no single supplier could satisfy Rubin alone, which turns memory approval into a gating item for the whole launch instead of a side detail. (trendforce.com) (techpowerup.com) Memory is only one choke point. Supply-chain reporting cited by The Register says Rubin is also dealing with the move to ConnectX-9 networking, higher total power draw, and stricter liquid-cooling requirements, which is why analysts cut its 2026 shipment mix from an earlier 29 percent estimate to 22 percent. (theregister.com) When the next generation slips, buyers do not stop buying. TrendForce said Nvidia raised shipment targets for B300 and GB300 Blackwell systems and increased orders for older high-bandwidth memory 3E, which pushes more cloud spending onto the current generation instead of the next one. (trendforce.com) That shift spreads the squeeze outward. The same racks need more advanced liquid cooling, the same contract manufacturers need to assemble more dense systems, and the same memory makers now have to choose between profitable artificial intelligence memory and conventional dynamic random-access memory, whose prices TrendForce said jumped sharply from the fourth quarter of 2025. (trendforce.com) (nvidia.com) So the story is not that Nvidia lacks a roadmap. Nvidia launched the Rubin platform at the start of 2026 with six tightly linked chips, but the market is still being pulled back toward Blackwell because the real limit is no longer model design on a whiteboard; it is certified memory stacks, packaged silicon, and enough power-and-cooling hardware to keep a rack of 72 chips alive. (nvidianews.nvidia.com) (developer.nvidia.com)