Nvidia Rubin GPUs hit delays

Nvidia’s next Rubin artificial-intelligence chips are running into supply delays, pushing some 2026 capacity plans back toward Blackwell systems instead. (trendforce.com) TrendForce said on April 8 that Rubin’s share of Nvidia’s high-end graphics processing unit shipments in 2026 is now expected to fall to 22%, down from an earlier 29% forecast. The same report raised Blackwell’s expected share to 71%, up from 61%. (trendforce.com) The bottlenecks are not limited to the chip itself. TrendForce cited slower validation of high-bandwidth memory 4, a shift from ConnectX-8 to ConnectX-9 networking, higher power draw, and tougher liquid-cooling requirements. (trendforce.com) Rubin is Nvidia’s next data-center platform after Blackwell, built as a rack-scale system that combines a Vera central processing unit, a Rubin graphics processing unit, ConnectX-9 networking, and new switches. Nvidia said in January that the platform was aimed at cutting inference token costs by as much as 10 times versus Blackwell at system scale. (nvidianews.nvidia.com) Nvidia has pitched Rubin as the hardware for the next wave of giant model training and inference, the step where a model generates answers after it is built. Nvidia said Microsoft’s Fairwater systems would scale to “hundreds of thousands” of Vera Rubin superchips, and CoreWeave was among the first cloud providers lined up to offer the platform. (nvidianews.nvidia.com) That makes timing important for cloud companies and large model developers that were planning 2026 upgrades. Network World reported on April 9 that delayed Rubin supply could leave enterprises relying longer on Hopper and Blackwell, while hyperscalers absorb the first shock by stretching existing deployments. (networkworld.com) Another pressure point is memory. Chosun Ilbo, citing TrendForce and KeyBanc, reported on April 9 that Nvidia appeared to have lowered its 2026 Rubin production target from 2 million units to 1.5 million units as suppliers worked through high-bandwidth memory 4 qualification; Nvidia had not publicly confirmed that figure. (chosun.com) Nvidia’s own public roadmap still points to Rubin deployments in the second half of 2026, not a cancellation. For now, the practical change is that Blackwell looks set to carry more of Nvidia’s 2026 artificial-intelligence server volume while Rubin ramps more slowly than buyers expected. (nvidianews.nvidia.com)