Nvidia Rubin delay boosts Blackwell demand
Reports say Nvidia’s next‑generation Rubin GPUs may be delayed, pushing hyperscalers to absorb more Blackwell shipments and slowing the upgrade cycle for enterprise AI infrastructure. The practical effect is that cheap, frontier inference remains unlikely in the near term and organisations will keep optimising around current hardware. That timeline matters for builders deciding whether to invest in migration to new accelerators or keep tuning for existing chip generations. (networkworld.com)
Nvidia sells artificial intelligence chips in waves, not one at a time. The current wave is Blackwell, and the next wave is Rubin, which Nvidia said at its March 2025 conference would follow on an annual rhythm after Blackwell and Blackwell Ultra. (blogs.nvidia.com) That rhythm matters because the biggest buyers do not order a few servers. Nvidia’s GB200 NVL72 system packs 72 Blackwell graphics processors and 36 Grace central processors into one liquid-cooled rack, so a single product launch can reshape entire data-center buildouts. (nvidia.com) Now that schedule is wobbling. Network World reported on April 9 that Nvidia’s Rubin graphics processors may be delayed, and TrendForce said on April 8 that Rubin shipment delays are one reason Blackwell is now expected to take more than 70% of Nvidia’s high-end graphics processor shipments in 2026. (networkworld.com) (trendforce.com) The bottleneck is memory. Rubin is tied to High Bandwidth Memory 4, which is a stack of very fast memory chips sitting next to the processor like a pantry built into the stove, and TrendForce said validation of that memory is taking longer than expected. (trendforce.com) TrendForce listed three other pressure points on the Rubin launch. Nvidia and its suppliers still have to move networking from ConnectX-8 to ConnectX-9, handle higher power draw, and tune more advanced liquid-cooling systems before Rubin ships at scale. (trendforce.com) When the next wave slips, the current wave gets pulled harder. TrendForce now sees Rubin’s 2026 share of Nvidia’s high-end graphics processor shipments falling from 29% to 22%, while Blackwell rises from 61% to 71%. (trendforce.com) That shift lands first with hyperscalers, which are giant cloud companies like Amazon Web Services, Microsoft Azure, and Google Cloud that buy entire fleets of servers at once. Nvidia said in 2025 that Blackwell Ultra systems would ship in the second half of 2025, giving those buyers a newer Blackwell option to absorb while Rubin slips. (blogs.nvidia.com) It also changes the math for everyone below the hyperscalers. Nvidia’s DGX B200 server uses eight Blackwell graphics processors and delivers up to 15 times the inference performance of the previous generation, so enterprises deciding what to deploy now have a strong reason to keep tuning software around hardware they can actually get. (nvidia.com) Inference is the part where a trained model answers your question, writes the paragraph, or ranks the search results. Nvidia’s GB200 rack is marketed for 30 times faster real-time inference on trillion-parameter language models, which is why a longer Blackwell cycle keeps today’s fast-but-expensive economics in place instead of unlocking a sudden cheap leap on Rubin. (nvidia.com) TrendForce also trimmed its 2026 growth forecast for high-end artificial intelligence graphics processor shipments from 26.8% to about 26% because of Rubin delays. That is not a collapse in demand; it is a sign that buyers still want more compute, but more of it will arrive on mature Blackwell systems instead of the next platform. (trendforce.com) So the near-term playbook is getting simpler. If you are building around Nvidia in 2026, the safer bet is not to wait for Rubin magic but to assume Blackwell, especially GB300 and B300 systems, will stay the main workhorse longer than planned. (trendforce.com)
