Nvidia's roadmap hit by bottlenecks

Nvidia’s next graphics processor problem is not just designing a faster chip. It is getting enough advanced memory, enough packaging capacity, and enough power-and-cooling hardware to turn that chip into a server customers can actually buy in volume. (theregister.com) TrendForce said on April 8 that Nvidia’s Blackwell family is now expected to make up 71 percent of the company’s high-end graphics processor shipments in 2026, up from an earlier 61 percent forecast. In the same forecast, Rubin fell to 22 percent from 29 percent, which is a big shift for a product that was supposed to be the next step up. (trendforce.com) A modern artificial intelligence chip is not one slab of silicon. It is more like a tiny city built from separate blocks, and one of the hardest jobs is connecting those blocks inside one package so data can move fast enough. (tsmc.com) That packaging method is called Chip-on-Wafer-on-Substrate by Taiwan Semiconductor Manufacturing Company, usually shortened to CoWoS. Taiwan Semiconductor Manufacturing Company says it combines chips on an interposer and substrate, which is the step Nvidia relies on for its biggest accelerator packages. (tsmc.com) Rubin adds another constraint because it uses High Bandwidth Memory 4, which is the next memory generation stacked vertically like a high-rise to sit close to the processor. Nvidia’s own Rubin materials say the platform is built around removing memory and communication bottlenecks, which only works if those memory stacks arrive in volume. (nvidia.com) The memory industry is only just moving into that new generation. Samsung says High Bandwidth Memory 4 can deliver up to 2.7 times higher throughput than its earlier generation, and Micron says it is in high-volume production of a 36-gigabyte High Bandwidth Memory 4 part for next-generation artificial intelligence platforms. (semiconductor.samsung.com) (micron.com) “Available” does not mean “easy to qualify.” The Register reported that TrendForce tied Rubin’s risk to the time needed to validate High Bandwidth Memory 4, the move to ConnectX-9 network cards, higher power draw, and more demanding liquid cooling. (theregister.com) That last part matters because Nvidia no longer sells just chips. Nvidia announced on March 16 that the Vera Rubin platform includes full rack products, central processors, storage gear, Ethernet switches, and inference accelerators, so one delayed part can slow an entire rack shipment. (nvidianews.nvidia.com) Blackwell benefits from being the part customers already know how to deploy. TrendForce said strong demand and Nvidia’s push toward integrated rack systems will still lift total high-end graphics processor shipments in 2026 even as the mix shifts away from Rubin. (trendforce.com) Geopolitics is part of the forecast too. TrendForce said the combined share of Hopper and Rubin will shrink because of geopolitical issues and supply-chain changes, which means the 2026 winner may be the chip with the fewest manufacturing surprises rather than the newest name on the roadmap. (trendforce.com) For cloud companies and enterprise buyers, that changes the shopping list. If Blackwell is the part that ships in volume while Rubin ramps more slowly, procurement teams may spend 2026 buying an older generation for longer, then redesigning power, cooling, and networking later when Rubin finally arrives at scale. (theregister.com)