Nvidia: Blackwell Wins, Rubin Delays
Analysts now expect Nvidia’s Blackwell architecture to carry most 2026 GPU shipments as Rubin faces supply‑chain and packaging risks. Reports suggest Blackwell could represent the bulk of shipments this year while Rubin and HBM4 memory bottlenecks constrain alternative supply, which means enterprise compute timelines will hinge more on packaging and memory availability than raw demand. For buyers and investors, that translates into continued pressure on cloud pricing and project sequencing if Blackwell remains the workhorse. (theregister.com) (communicationstoday.co.in)
Nvidia’s newest artificial intelligence chips are not being held back by a lack of buyers. They are being held back by the plumbing: advanced packaging, high-bandwidth memory, liquid cooling, and the network cards that tie whole racks together. (theregister.com) That is why analysts now think Blackwell, not Rubin, will do most of the work in 2026. TrendForce cut its estimate for Rubin’s share of Nvidia high-end graphics processor shipments in 2026 to 22 percent from 29 percent, which leaves Blackwell carrying the larger load. (theregister.com) A graphics processor is the chip that does the math for training and running large artificial intelligence models. In Nvidia’s data-center business, customers usually do not buy one chip at a time anymore; they buy whole racks packed with processors, memory, networking, and cooling as one system. (nvidia.com) Blackwell is the current generation of that system. Nvidia says its Grace Blackwell GB200 platform combines processors into rack-scale products like the GB200 NVL72, which is why shipment mix now depends on whether entire racks can be assembled, not just whether one chip comes off a wafer line. (nvidia.com) Rubin is the next platform after Blackwell, and Nvidia has already positioned it as a full supercomputer design rather than a single processor launch. Nvidia’s own Rubin materials describe six new chips in one platform, which makes every missing part a possible bottleneck. (nvidia.com) One bottleneck is high-bandwidth memory, which is the ultra-fast memory stacked like a vertical apartment tower beside the processor. Rubin is expected to use High Bandwidth Memory 4, and TrendForce says validation time for that newer memory is one reason Rubin volumes may slip. (theregister.com) Another bottleneck is packaging, which is the factory step that mounts several advanced parts together so they behave like one engine. The Register notes that Nvidia has already run into packaging constraints before, and Blackwell itself previously faced delays tied to the technology used to stitch its dies together. (theregister.com) The networking piece matters too. TrendForce said Rubin’s rollout is also being slowed by the move to Nvidia’s faster ConnectX-9 network interface cards, which are the adapters that let racks of chips talk to each other at full speed. (theregister.com) Power and cooling are now part of the chip story. TrendForce said Rubin systems need more power overall and more advanced liquid cooling, which means a cloud provider can have budget approval for processors and still wait on pumps, pipes, and data-center retrofits. (theregister.com) The demand side is still moving up. TrendForce expects high-end artificial intelligence server revenue to keep growing as Nvidia pushes integrated rack products, and one recent industry forecast said artificial intelligence server revenue could rise more than 30 percent in 2026 after a nearly 48 percent jump in 2025 driven largely by Blackwell systems. (communicationstoday.co.in 1) (communicationstoday.co.in 2) That leaves buyers in a familiar position. If Blackwell remains the workhorse while Rubin ramps slowly, cloud capacity will expand in steps instead of a smooth wave, and pricing will keep reflecting whichever piece of the rack is hardest to get that quarter. (theregister.com) (nvidia.com)
