Blackwell still rules GPUs

Nvidia’s next artificial intelligence chip is already on the roadmap, but the chip most buyers are likely to get through 2026 is the one it is selling now: Blackwell. TrendForce said high-end Nvidia graphics processing unit shipments in 2026 should be led by Blackwell as demand stays strong and rack-scale systems use more chips per deployment. (communicationstoday.co.in) A graphics processing unit is the part of a server that does huge numbers of math operations at once, like a warehouse full of workers all moving boxes in parallel instead of one clerk at a desk. Artificial intelligence training uses thousands of these chips together because one chip alone would take too long to process the model’s calculations. (nvidia.com) Blackwell is not usually sold as a single loose chip for the biggest customers. Nvidia has been pushing full-rack systems such as GB200 NVL72 and GB300 NVL72, where dozens of Blackwell processors, networking gear, and memory are wired together as one machine. (communicationstoday.co.in) That rack design changes the bottleneck. The hard part is no longer just making the silicon die itself, but also advanced packaging, which is the step where chip pieces and high-bandwidth memory are bonded into one dense module with tiny connections. (trendforce.com) Rubin is the next Nvidia architecture after Blackwell, but TrendForce and other supply-chain reporting point to packaging limits and memory timing as reasons its ramp could be slower than the market once expected. TrendForce said Nvidia has revised Rubin production plans while Blackwell demand in the first half of 2026 keeps pulling in more high-bandwidth memory generation 3E supply. (trendforce.com) One sign of that pressure is design simplification. TrendForce reported on April 1 that Rubin Ultra appears to be sticking with a dual-die package instead of a rumored four-die package because a larger package would hurt yield and raise cost. (trendforce.com) That leaves Nvidia in an unusual position: the “old” product is still the main event because customers cannot wait for the “new” one. Communications Today, citing TrendForce on April 8, said Nvidia’s integrated GB and Vera Rubin rack strategy will still lift high-end graphics processing unit shipments in 2026 even as Blackwell remains the dominant mix. (communicationstoday.co.in) The second half of the story is software. Nvidia Mission Control is the company’s control layer for these Blackwell clusters, and Nvidia says it handles workload scheduling, orchestration, monitoring, and autonomous recovery from one control plane. (nvidia.com) Scheduling sounds abstract, but it solves a very physical problem: a 72-processor rack is only valuable if jobs land on the right processors at the right time without stranding memory and network links. Nvidia’s developer documentation this week said Mission Control maps jobs to the topology of GB200 NVL72 and GB300 NVL72 systems and works with Slurm and Run:ai to turn the rack into something operators can partition and share. (developer.nvidia.com) Nvidia said in March 2025 that Mission Control could boost graphics processing unit utilization by 5 times in some environments, which is why the software matters as much as the metal. If a company spends billions on Blackwell racks, the next check it writes is often for the software that keeps those racks full, stable, and recoverable when a job or node fails. (blogs.nvidia.com) So the 2026 market is shaping up less like a clean handoff from Blackwell to Rubin and more like a long extension of Blackwell’s peak. The chips are still scarce enough to matter, but the bigger sale now is the operating system for an artificial intelligence factory, not just the processors inside it. (nvidia.com)