Nvidia: Blackwell mix and Mission Control

Nvidia’s newest problem is no longer just making enough artificial intelligence chips. Analysts now think its Blackwell line will take 71% of Nvidia’s high-end graphics processing unit shipments in 2026, up from an earlier 61% forecast, because the newer Rubin line is running into supply-chain friction. (trendforce.com) That sounds backward until you look at what these machines are. A Blackwell rack is not one chip in one server; Nvidia’s GB300 NVL72 packs 72 Blackwell Ultra graphics processing units and 36 Grace central processing units into one liquid-cooled rack. (nvidia.com) When a rack holds 72 chips, the hard part is not only silicon. It is making all 72 chips, the memory, the network links, the power delivery, and the cooling system behave like one machine instead of 72 expensive parts fighting each other. (docs.nvidia.com, nvidia.com) That is where Mission Control comes in. Nvidia introduced it as software that provisions Blackwell systems, checks that the rack is healthy, and then moves jobs across the cluster as companies switch between training and inference. (blogs.nvidia.com) Training is the stage where a model learns from giant piles of data. Inference is the stage where the finished model answers questions for users, and Nvidia says Mission Control can reallocate the same Blackwell infrastructure between those two jobs as demand changes. (blogs.nvidia.com) Inside these racks, placement matters. Nvidia’s April 7, 2026 technical post says Mission Control maps the physical layout of GB200 and GB300 NVL72 systems so schedulers such as Slurm and Run:ai can place work on the right group of chips instead of treating the rack like a random pile of processors. (developer.nvidia.com) Nvidia’s own hardware guide shows why that mapping is necessary. Each NVL72 rack contains 18 compute trays, and each tray holds 4 Blackwell graphics processing units, so one bad placement decision can strand memory bandwidth or split a job across slower paths. (docs.nvidia.com) TrendForce’s new shipment forecast says Blackwell is gaining share partly because Rubin is harder to land on time. The firm points to high-bandwidth memory 4 validation, a network transition from ConnectX-8 to ConnectX-9, higher power draw, and more advanced liquid-cooling demands as the main trouble spots. (trendforce.com) So Nvidia is selling two things at once. It is selling Blackwell hardware that is mature enough to ship in volume, and it is selling Mission Control software that tries to keep those packed racks busy enough to justify their cost. (trendforce.com, blogs.nvidia.com) Nvidia says Mission Control can boost infrastructure utilization by up to 5 times and cut job recovery time by up to 10 times with checkpointing and automated restarts. Those numbers are the clearest sign of where the market is moving: once a company has a room full of Blackwell racks, the next bottleneck is no longer getting a chip, but getting every chip in the room pointed at the right job. (blogs.nvidia.com)