Nvidia’s Blackwell cycle widens
Nvidia’s Blackwell architecture is poised to dominate high-end AI GPU shipments this year as demand for rack-scale AI systems climbs and Rubin, the next-gen family, faces supply-chain delays. The result is concentration around Blackwell for 2026 shipments, which helps standardisation but risks keeping pricing and memory bottlenecks tight across the industry. Analysts flag that integrated software and rack offerings — not just chips — are becoming the commercial product being sold to enterprise customers. ( - )
Nvidia is not just selling more artificial intelligence chips in 2026. It is selling more of one generation of chips, because the next one is slipping and customers still want whole racks right now. (theregister.com) (trendforce.com) TrendForce said on April 8 that Blackwell’s share of Nvidia’s high-end artificial intelligence graphics processor shipments in 2026 is now expected to rise to 71%, up from 61% in its earlier forecast. Rubin’s share was cut to 22% from 29%. (trendforce.com) The reason is not weak demand. TrendForce still expects total high-end artificial intelligence graphics processor shipments to grow by about 26% in 2026, driven by customers buying bigger systems for training and inference. (trendforce.com) (communicationstoday.co.in) Rubin is running into the ugly part of hardware launches: memory validation, network changes, power delivery, and cooling. TrendForce and The Register both pointed to delays around High Bandwidth Memory 4, the move to ConnectX-9 networking, higher power draw, and more complex liquid cooling. (theregister.com) (trendforce.com) That matters because Nvidia’s newest systems are no longer single cards you slide into a server. The flagship Blackwell rack, called GB200 NVL72, links 72 Blackwell graphics processors with 36 Grace central processors in one liquid-cooled rack. (nvidia.com) (docs.nvidia.com) Nvidia says that rack behaves like one giant machine through a 72-graphics-processor NVLink domain. In plain English, that means buyers are purchasing a prewired block of compute, networking, and cooling instead of shopping for chips one by one. (nvidia.com) (docs.nvidia.com) That is why Blackwell’s wider cycle helps Nvidia even if Rubin arrives later. A slower handoff gives cloud providers and enterprises more time to standardize on Blackwell racks, software, and operating procedures. (trendforce.com) (nvidia.com) It also keeps the bottlenecks in place. Blackwell systems rely on High Bandwidth Memory 3E today, while Rubin is tied to High Bandwidth Memory 4, so delays in the next memory generation keep pressure on the current supply chain and on pricing for the systems customers can actually get. (theregister.com) (trendforce.com) (nvidia.com) The commercial product is also getting taller in the stack. Nvidia’s own enterprise pitch now bundles Blackwell hardware with networking, orchestration, and Nvidia AI Enterprise software in what it calls an Enterprise AI Factory validated design. (nvidia.com 1) (nvidia.com 2) Even Nvidia’s cloud offer follows the same pattern. DGX Cloud Lepton is marketed as a unified platform that lets developers build, train, and deploy across a network of providers without managing the infrastructure underneath. (nvidia.com) (nvidianews.nvidia.com) So the 2026 story is not just “Blackwell beats Rubin.” It is that Nvidia’s sales unit is becoming the full artificial intelligence factory — rack, network, memory, cooling, and software together — and Blackwell now has a longer window to own that package. (communicationstoday.co.in) (nvidia.com) (theregister.com)
