Rubin GPU delay could extend Blackwell era
Reports suggest Nvidia's next-generation 'Rubin' GPUs may be delayed, which would keep the industry dependent on current Blackwell systems for longer and extend optimisation cycles around today's hardware. The potential delay could push hyperscalers and cloud buyers to keep buying and tuning Blackwell-era racks rather than waiting for a platform handoff. (networkworld.com)
A graphics processing unit is the engine that does the math for modern artificial intelligence, and Nvidia’s current engine is called Blackwell. A new report on April 9 said some customers now expect the handoff to the next engine, Rubin, to slip, which would keep Blackwell in the driver’s seat longer. (networkworld.com) That matters because Nvidia has been selling whole racks now, not just chips. Its Blackwell NVL72 system links 72 Blackwell graphics processors in one cabinet, and Nvidia says the Rubin NVL72 is the next rack-scale step after that. (nvidia.com) Nvidia itself spent 2025 and early 2026 telling buyers that Rubin was close. At the January 5, 2026 Consumer Electronics Show announcement, Nvidia said Rubin was the “next generation of AI” and described a platform built around a Vera central processor and Rubin graphics processor. (nvidianews.nvidia.com) By March 16, 2026, Nvidia had gone further and said the Vera Rubin platform was “in full production.” The company said that platform included seven chips and five rack types aimed at what Jensen Huang calls “AI factories,” which is Nvidia’s term for data centers built mainly to produce artificial intelligence output. (nvidianews.nvidia.com) So the surprise is not that Rubin exists on a slide. The surprise is that, weeks after Nvidia said “full production,” Network World reported that delayed customer access could still slow the actual rollout of Rubin systems into enterprise and cloud deployments. (networkworld.com, nvidianews.nvidia.com) Blackwell is not a simple placeholder people can ignore while they wait. Nvidia’s January 2026 Rubin launch said Rubin would cut inference token cost by up to 10 times versus Blackwell and reduce the number of graphics processors needed to train some mixture-of-experts models by 4 times, which tells you how much money buyers thought might be on the line in the next upgrade. (investor.nvidia.com) If Rubin slips, the practical effect is not that data centers stop buying. The practical effect is that hyperscalers like cloud providers keep ordering and tuning Blackwell-era racks, because an artificial intelligence cluster is more like a power plant than a laptop and Nvidia has said these transitions take years of planning. (networkworld.com, datacenterdynamics.com) That can actually help Blackwell’s staying power. The longer Rubin hardware is scarce or staggered, the more software teams optimize models, networking, cooling, and power layouts around Blackwell racks they can buy today instead of a future platform they cannot deploy at scale yet. (networkworld.com, nvidia.com) Nvidia’s public roadmap still points to a long runway after Blackwell, not a cancellation. At GTC 2025, the company said Rubin was due in 2026, Rubin Ultra in 2027, and Feynman in 2028, so even a delay now would look more like a traffic jam in a planned convoy than a missing vehicle. (networkworld.com, blogs.nvidia.com) The near-term question is not whether Nvidia has a post-Blackwell story. The near-term question is whether customers can get enough Rubin systems, soon enough, to justify pausing Blackwell purchases when Nvidia is still selling Blackwell into a market it says has demand stretching through 2027. (datacenterdynamics.com, networkworld.com)
