Nvidia Rubin GPU delay

Nvidia’s next chip after Blackwell may arrive later than expected, and the holdup is not the main processor but the memory bolted beside it. Multiple reports this week say Rubin systems are running into supply and validation friction around High Bandwidth Memory 4, the next generation of stacked memory used in artificial intelligence servers. (networkworld.com) (trendforce.com) That memory is the part that keeps a graphics processor fed with data. If the processor is the engine, High Bandwidth Memory is the fuel line, and Nvidia’s Rubin platform is designed to use High Bandwidth Memory 4 rather than the older High Bandwidth Memory 3E used in Blackwell systems. (nvidia.com 1) (nvidia.com 2) High Bandwidth Memory is hard to make because manufacturers stack many dynamic random-access memory dies vertically and connect them with microscopic vertical wires. TrendForce said in March that Nvidia was asking suppliers for data rates above 10 gigabits per second, above the 8 gigabits per second High Bandwidth Memory 4 standard, which forced design changes and new qualification work. (trendforce.com) That extra speed target matters because Rubin is supposed to be a big step, not a small refresh. Nvidia said on January 5, 2026 that Rubin could cut inference token cost by up to 10 times and reduce the number of graphics processors needed to train mixture-of-experts models by 4 times versus Blackwell. (investor.nvidia.com) The awkward part is that Nvidia also told the world on March 16, 2026 that Vera Rubin was “in full production.” What seems to be happening is that the platform launch is real, but the volume ramp for the most advanced memory and the broad shipment mix inside 2026 may still be sliding around underneath that announcement. (investor.nvidia.com) (trendforce.com) TrendForce now expects Blackwell to account for 71% of Nvidia’s high-end graphics processor shipments in 2026, up from 61% in its earlier view. The same note says Hopper and Rubin together lose share because of geopolitical issues and supply-chain changes, which is a polite way of saying the next handoff is not going cleanly. (trendforce.com) Tom’s Hardware reported earlier that High Bandwidth Memory 4 mass production had slipped to the end of the first quarter of 2026, while Rubin volume availability moved to the second half of 2026. Nvidia pushed back on the idea of a mass-production delay, but even that report described a cadence shift tied to stricter memory specifications. (tomshardware.com) For cloud companies and large data-center buyers, a Rubin delay does not mean spending stops. It usually means more racks of Blackwell get ordered instead, because Nvidia’s current GB200 NVL72 rack already links 72 Blackwell graphics processors and ships with 130 terabytes per second of NVLink bandwidth inside one rack. (nvidia.com) That is why this kind of delay can keep demand smooth instead of creating a cliff. If the next platform is late, customers still buy networking gear, power equipment, cooling, software licenses, and Blackwell servers, just on a longer runway than they expected in January. (networkworld.com) (nvidia.com) The companies with the most immediate problem may be the memory suppliers, not Nvidia. TrendForce and Korean reporting both point to Samsung and SK hynix still fighting through qualification, speed, heat, and power targets for High Bandwidth Memory 4, which means the bottleneck sits in one of the hardest parts of the bill of materials. (trendforce.com) (chosun.com) So the near-term picture looks less like a missed product cycle and more like a traffic jam at the on-ramp. Rubin is still Nvidia’s next destination, but Blackwell now looks set to carry more of 2026 than buyers, suppliers, and maybe Nvidia itself expected three months ago. (networkworld.com) (trendforce.com)