Blackwell may stay longer
Possible delays to Nvidia’s next‑gen Rubin accelerator could extend commercial demand for existing Blackwell GPUs, which analysts expect to account for more than 70% of high‑end GPU shipments in 2026—keeping pricing pressure on buyers. Packaging capacity expansions at TSMC may ease supply over time, but the transition looks likely to be uneven. (parameter.io) (dqindia.com)
Nvidia’s next big artificial intelligence chip family was supposed to take over quickly, but new supply-chain reports say that handoff may slip and keep today’s Blackwell chips in the driver’s seat deeper into 2026. TrendForce said on April 8 that Blackwell’s share of high-end Nvidia graphics processor shipments could rise from 61% to 71% in 2026 as Rubin and older Hopper lose ground. (trendforce.com) That matters because data-center buyers do not buy a loose chip; they buy entire racks built around a chip generation, and changing generations is like swapping the engine design for a whole truck fleet. Nvidia has spent the last year pushing Blackwell-based rack systems such as GB200 and newer Blackwell Ultra systems, so a longer Blackwell cycle keeps those platforms commercially relevant for cloud companies already building around them. (nvidia.com 1) (nvidia.com 2) Rubin is not a small refresh. Nvidia said in January that the Rubin platform would bundle six major chips into one system design, and in March it said seven Rubin-platform chips were already in full production for what it calls “AI factories.” (nvidia.com 1) (nvidia.com 2) The snag is packaging. Modern artificial intelligence chips are built more like a layered wedding cake than a single slab of silicon, and Taiwan Semiconductor Manufacturing Company’s CoWoS packaging is the tray that holds the pieces together with memory and interconnects. (tsmc.com) Blackwell already eats a huge amount of that packaging capacity because its chips are physically larger and use far more high-bandwidth memory than the older Hopper line. TrendForce said in 2024 that Blackwell-class chips are about twice the die size of H100 and that Nvidia’s demand would take nearly half of Taiwan Semiconductor Manufacturing Company’s CoWoS capacity as Blackwell became mainstream. (trendforce.com) Now the same bottleneck is shaping Rubin. TrendForce reported on April 1 that Rubin Ultra is still being planned as a dual-die design for 2027, not a more aggressive four-die package rumored in the market, and that Nvidia has shifted more wafer allocation toward Blackwell instead of cutting orders. (trendforce.com) That shift helps explain why buyers may keep feeling price pressure even if total supply improves. If hyperscalers such as the biggest cloud companies keep absorbing Blackwell racks while Rubin ramps unevenly, the “old” generation does not get cheap; it stays scarce because it is still the preferred thing shipping in volume. (trendforce.com 1) (trendforce.com 2) Memory is part of the story too. Tom’s Hardware reported in January that suppliers had to rework high-bandwidth memory generation four products to meet Nvidia’s revised Rubin requirements, though Nvidia disputed that there was a mass-production delay, and the report said volume Rubin availability was then expected in the second half of 2026. (tomshardware.com) So the odd result is that a delay in the “next” chip can strengthen the market power of the “current” chip. If Blackwell holds around 71% of Nvidia’s high-end graphics processor mix in 2026 while Taiwan Semiconductor Manufacturing Company gradually adds packaging capacity, the transition to Rubin may look less like a clean launch and more like a traffic merge that clears one lane at a time. (trendforce.com) (tsmc.com)
