Blackwell dominates; Rubin faces supply friction
Analysts now expect Nvidia's Blackwell line to account for roughly 70% of high‑end AI GPU shipments this year as demand keeps rising, while Rubin GPUs face potential delays because of supply‑chain and packaging constraints. Nvidia is also pushing software like Mission Control to turn Blackwell rack systems into schedulable infrastructure, a sign that orchestration and utilisation will matter as much as pure chip supply. (theregister.com) (communicationstoday.co.in) (blockchain.news)
Nvidia’s newest artificial intelligence chips are winning the easy part and struggling with the hard part at the same time. Analysts now expect the Blackwell family to take 71% of Nvidia’s high-end artificial intelligence graphics processor shipments in 2026, while the newer Rubin line is slipping because the parts around the chip are harder to build. (trendforce.com) That split sounds backwards until you look at how these systems are sold. Nvidia is no longer shipping just a processor on a board; it is shipping whole rack systems like GB200 NVL72 and GB300 NVL72, where dozens of Blackwell chips are wired together as one machine. (developer.nvidia.com) Blackwell is the line customers can actually deploy at scale right now. TrendForce says Blackwell’s shipment mix rises from 61% to 71% this year, while Rubin falls from 29% to 22% and Hopper drops from 10% to 7%. (trendforce.com) (communicationstoday.co.in) The bottleneck is not demand. Nvidia reported $62.3 billion in data center revenue for the quarter ended January 25, 2026, up 75% from a year earlier, which shows cloud companies and model builders are still buying as fast as supply appears. (nvidianews.nvidia.com) The bottleneck is packaging. Advanced packaging is the step where multiple pieces of silicon and stacks of high-bandwidth memory are stitched into one giant module, and TrendForce says Rubin is running into constraints there as package size, yield, and cost all get harder to manage. (trendforce.com) One sign of that strain is Rubin Ultra’s design change. TrendForce reported that Nvidia is now leaning toward a dual-die layout instead of a four-die single-package design, because a larger package would push toward roughly 7.5 to 8 times the reticle limit and hurt manufacturing yield. (trendforce.com) Memory is part of the squeeze too. TrendForce reported in March that Samsung and SK hynix are expected to supply Rubin’s high-bandwidth memory generation called HBM4, but allocation and pricing were still unsettled, which adds another moving part before volume shipments can stabilize. (trendforce.com) Nvidia’s answer is not only more chips. It is also software called Mission Control, which Nvidia says handles workload scheduling, orchestration, monitoring, and autonomous recovery for Blackwell and Rubin data centers. (nvidia.com) (docs.nvidia.com) That software matters because a Blackwell rack is not one pool of identical processors. Nvidia’s developer blog says GB200 NVL72 and GB300 NVL72 systems use NVLink switches and shared-memory features across nodes, so the scheduler has to understand the machine’s physical layout before it decides where a training or inference job should run. (developer.nvidia.com) Nvidia has been selling this as an efficiency upgrade, not just an admin console. In March 2025, the company said Mission Control could boost graphics processor utilization by 5 times on Blackwell infrastructure, which is the software equivalent of finding more seats on a full airplane. (blogs.nvidia.com) So the 2026 picture is unusually clear. Blackwell is becoming the volume engine because it is mature enough to ship in rack-scale systems now, while Rubin is still close enough to the manufacturing edge that packaging and memory can slow it down before customers ever see the chip. (trendforce.com 1) (trendforce.com 2)
