Blackwell stays the AI‑chip workhorse
Analysts expect Nvidia’s Blackwell generation to dominate high‑end AI‑chip shipments through 2026, keeping customers anchored on a single architecture while Rubin faces delay and HBM4 supply constraints. Nvidia is also moving beyond silicon with 'Mission Control' software to manage rack‑scale AI jobs, signalling that efficient cluster utilisation and workload placement are becoming as important as raw GPU supply. For cloud and enterprise buyers, that means capacity planning, software integrations and topology‑aware scheduling will drive procurement choices as much as chip specs. (news.futunn.com) (blockchain.news)
The surprise in Nvidia’s 2026 chip roadmap is that the “current” generation is now expected to get even bigger. TrendForce said on April 8 that Blackwell’s share of Nvidia’s high-end artificial intelligence chip shipments is now projected to rise from 61% to 71% in 2026, while Hopper and Rubin both lose share. (trendforce.com) That sounds odd until you remember how these systems are sold. Nvidia is no longer just shipping single chips; it is selling entire rack-scale systems like the GB200 NVL72, which links 72 Blackwell graphics processing units and 36 Grace central processing units into one liquid-cooled rack. (nvidia.com) A rack-scale system is basically a whole aisle of a data center acting like one giant machine. Nvidia says the GB200 NVL72 creates a 72-graphics-processing-unit NVLink domain that behaves like a single massive graphics processor for very large language-model inference. (nvidia.com) That design changes the buying decision. A cloud provider choosing Blackwell is not just picking a faster chip; it is picking a cabinet layout, a cooling system, a network fabric, and a software stack that all have to fit together. (nvidia.com) The reason Rubin is slipping is not a simple product delay. TrendForce said the next-generation platform is running into validation and supply-chain friction around High Bandwidth Memory 4, which is the ultra-fast stacked memory that sits next to the graphics processor and feeds it data like a very wide fuel line. (trendforce.com) TrendForce had already warned in January that High Bandwidth Memory 4 mass production had moved to the end of the first quarter of 2026 because of specification upgrades and Nvidia product-strategy changes. That matters because Rubin depends on that memory generation, so a memory bottleneck can hold up an entire server launch. (trendforce.com) Rubin also brings other moving parts at the same time. TrendForce’s April note pointed to interconnect transitions, higher power draw, and more demanding liquid-cooling requirements, which means customers are not just waiting for one chip but for a whole new platform to be qualified. (trendforce.com) Nvidia is responding by making the software layer more central. Its Mission Control product is pitched as the operating system for an “artificial intelligence factory,” handling workload scheduling, monitoring, and automatic recovery across Blackwell and Rubin data centers. (nvidia.com) Scheduling sounds abstract, but it decides whether a training job lands on graphics processors that are tightly linked or scattered across the room. Nvidia’s developer documentation this week said Mission Control maps things like NVLink domains and partitions into schedulers such as Slurm and Run:ai so jobs can be placed on the right part of the machine. (developer.nvidia.com) That is why Blackwell staying dominant is bigger than one generation winning another. If customers spend 2026 standardizing on Blackwell racks and Mission Control-style orchestration, the lock-in sits in the cluster design and scheduling software as much as in the graphics processor itself. (nvidia.com)
