Blackwell chips fill a Rubin-shaped gap
Delays to Nvidia’s next-generation Rubin GPUs are redirecting demand toward Blackwell, making Blackwell the dominant workhorse for high-end AI this year. Multiple supply-chain reports expect Blackwell to account for roughly 70%+ of high-end AI GPU shipments as customers and data centers shift orders, and Nvidia’s packaging and scheduling tools are becoming strategic for customers locking in scarce capacity. The practical consequence is that access to accelerated compute will stay constrained and procurement terms will matter as much as architecture choices. (coincentral.com) (theregister.com)
Nvidia’s newest chips were supposed to take over this year, but a supply-chain jam is pushing buyers back to the older line instead. TrendForce now says Blackwell will make up 71% of Nvidia’s high-end artificial intelligence graphics processor shipments in 2026, up from an earlier 61% forecast. (trendforce.com) The chip getting squeezed is Rubin, Nvidia’s next platform after Blackwell. TrendForce cut Rubin’s expected share of 2026 high-end shipments to 22% from 29% because the system is taking longer to get ready at full scale. (trendforce.com) That delay is not one broken part. TrendForce and The Register point to four separate bottlenecks: validating new high-bandwidth memory called High Bandwidth Memory 4, moving to Nvidia’s faster ConnectX-9 networking chips, handling higher power draw, and tuning more advanced liquid cooling. (theregister.com) Blackwell is not a consolation prize. Nvidia said in January that Rubin could cut inference token costs by as much as 10 times versus Blackwell, which tells you Blackwell is already the baseline machine customers are measuring everything against. (nvidia.com) The memory piece matters because these chips are not sold naked. Rubin depends on High Bandwidth Memory 4, which is the stacked memory sitting right next to the graphics processor so it can feed data fast enough for giant artificial intelligence models. (theregister.com) When that memory slips, the whole rack slips with it. The Register says the extra time needed to validate High Bandwidth Memory 4 is one reason Rubin may ship later and in smaller volumes than expected. (theregister.com) Customers are also buying more complete systems, not just loose chips. TrendForce says Nvidia’s push for integrated GB and VR rack solutions will raise total high-end graphics processor shipments in 2026 even as the mix shifts toward Blackwell. (trendforce.com) That changes what buyers compete for. If you need a full rack with networking, cooling, and memory all lined up, the scarce thing is no longer just the chip die coming out of a factory; it is the delivery slot for the whole package. (theregister.com) Nvidia has been telling customers that Rubin is coming fast. On March 16, 2026, the company said the Vera Rubin platform was “in full production” with seven new chips for artificial intelligence factories, which makes the supply-chain warnings look less like a cancellation and more like a difficult ramp. (nvidia.com) So the 2026 fight is not really Blackwell versus Rubin in a clean handoff. It is buyers trying to lock in Blackwell capacity while Rubin works through memory, networking, power, and cooling problems that have turned procurement timing into part of the product. (trendforce.com)
