Nvidia supply squeeze deepens

Nvidia’s newest artificial intelligence chips are selling so fast that the next chip after them is slipping, and that is making the current chip even harder to get. TrendForce said Blackwell is now expected to take 71% of 2026 high-end artificial intelligence graphics processor shipments, up from 61%, while Hopper and Rubin lose share. (trendforce.com) A graphics processor is the engine that trains and runs large artificial intelligence models, and companies buy them by the rack, not by the card. Nvidia’s Blackwell systems are being sold as giant rack-scale machines, which means each customer order now pulls in more chips, more networking, and more power gear at once. (nvidia.com) That rack-scale design is unusually dense. Nvidia says one GB200 NVL72 system links 72 Blackwell graphics processors with 36 Grace central processors inside one liquid-cooled rack and treats them like one huge computing pool. (nvidia.com) The next pressure point is memory. High Bandwidth Memory is the stacked memory glued close to the chip so data can move fast enough for artificial intelligence work, and TrendForce said High Bandwidth Memory 4 mass production was pushed to no earlier than the end of the first quarter of 2026 after specification changes. (trendforce.com) Rubin is the platform meant to follow Blackwell, and it depends on that newer memory. TrendForce said strong demand for Blackwell led Nvidia to adjust Rubin’s mass-production timeline, which pushed more demand back onto Blackwell instead of relieving it. (trendforce.com) Memory suppliers are not exactly sitting on spare inventory either. The Register reported in February that Micron had pre-sold every High Bandwidth Memory 4 chip it could make in 2026, while Samsung had also started shipping and SK hynix was still the last major supplier yet to announce production. (theregister.com) When the scarce part is not just the chip but the whole 72-processor rack, software starts to matter more. Nvidia’s Mission Control is the control layer that slices those giant racks into smaller chunks, watches failures, and places jobs where the fastest chip-to-chip links are available. (developer.nvidia.com) Nvidia said Mission Control works with Slurm and Run:ai, two common job schedulers, and uses identifiers like cluster UUID and clique ID to map which graphics processors share the same NVLink fabric. That lets one expensive rack run multiple teams without breaking the high-speed paths that make the rack valuable in the first place. (developer.nvidia.com) Nvidia is also selling that software as part of the product, not as an afterthought. Its Mission Control page says the stack now powers Blackwell and Rubin data centers and handles scheduling, orchestration, monitoring, and autonomous recovery across what Nvidia calls artificial intelligence factories. (nvidia.com) So the squeeze is no longer just “not enough chips.” It is a stack problem where delayed Rubin timing, tight High Bandwidth Memory 4 supply, and bigger Blackwell rack orders all point buyers to the same systems at the same time, which is why Blackwell’s share is rising even before its successor fully arrives. (trendforce.com)