Nvidia: demand strong, supply messy
Nvidia still anchors AI infrastructure demand, but the rollout is being reshaped by memory and packaging bottlenecks rather than lack of customer interest. (theregister.com) TrendForce and others now expect Blackwell GPUs to dominate 2026 shipments while Rubin chips face delay risk—a dynamic that boosts near‑term Blackwell monetisation but extends dependence on constrained HBM and advanced packaging capacity. (cloudnews.tech)
Nvidia’s problem right now is not finding buyers for artificial intelligence chips. It is getting enough memory stacks and advanced packaging slots to turn orders into working servers on time. (theregister.com) That is why 2026 is looking more like a Blackwell year than a Rubin year. TrendForce now expects Blackwell to take 71% of Nvidia’s high-end graphics processing unit shipments in 2026, up from 61%, while Rubin falls to 22% from 29%. (cloudnews.tech) Blackwell is Nvidia’s current chip family for giant artificial intelligence systems. Nvidia says Blackwell is already in full production, and its GB200 NVL72 rack links 72 Blackwell graphics processing units inside one cabinet with 130 terabytes per second of chip-to-chip bandwidth. (nvidia.com 1) (nvidia.com 2) Rubin is the next family after Blackwell, and Nvidia has already unveiled it for future systems. Nvidia says Rubin uses High Bandwidth Memory 4 and ConnectX-9 networking, which means it depends on a newer memory generation and a newer network stack than Blackwell does. (nvidia.com 1) (nvidia.com 2) High Bandwidth Memory is the short, fat fuel tank bolted right next to the chip so the chip does not have to wait for data. Rubin’s shift to High Bandwidth Memory 4 adds speed, but TrendForce says validation of that memory is one reason Rubin faces delay risk. (cloudnews.tech) (nvidia.com) Advanced packaging is the factory step that glues the processor and those memory stacks into one dense module. Taiwan Semiconductor Manufacturing Company’s CoWoS packaging has been the narrow door for these systems, and TrendForce said in late 2024 that demand was still above supply even after capacity doubled in 2024 and 2025. (trendforce.com) Blackwell still needs that same packaging door, but it uses High Bandwidth Memory 3E that is already further along in production. Nvidia’s Blackwell graphics processing units carry up to 192 gigabytes of High Bandwidth Memory 3E, and Blackwell Ultra pushes that to 288 gigabytes. (nvidia.com) (developer.nvidia.com) The result is a strange kind of good news for Nvidia. If Rubin slips, customers do not stop spending; they keep buying Blackwell racks, which lets Nvidia monetize the current generation for longer while the next one waits for memory, cooling, and networking pieces to line up. (theregister.com) (cloudnews.tech) TrendForce also says total high-end graphics processing unit shipments should still grow by about 26% in 2026. So the bottleneck is not a collapse in demand; it is a traffic jam in the most specialized parts of the chip supply chain. (cloudnews.tech) That traffic jam now stretches beyond Nvidia to the companies around it. Taiwan Semiconductor Manufacturing Company has to expand CoWoS lines, and memory suppliers such as SK hynix, Micron, and Samsung have to keep feeding the High Bandwidth Memory pipeline that every new Nvidia rack consumes. (trendforce.com) (theregister.com) So the story is not that artificial intelligence spending cooled off in April 2026. The story is that Nvidia’s roadmap is being paced by the hardest parts of semiconductor manufacturing, and right now those hard parts favor shipping more Blackwell before Rubin can fully take over. (theregister.com) (cloudnews.tech)
