Santa Clara, California, April 9, 2026, 04:13 PDT.
Nvidia’s next chip rollout may hit a snag. TrendForce, citing HBM4 qualification issues, higher cooling needs, and a shift in networking, said on April 8 that Rubin AI processor shipments might slip, pushing a larger chunk of this year’s output onto the Blackwell lineup. Blackwell could now represent over 70% of Nvidia’s high-end GPU shipments in 2026, up from earlier expectations, while Rubin’s projected share drops to 22%, down from 29%.
Right now, customer appetite shows no signs of slowing. Amazon reported on Thursday that AWS’s AI revenue run rate hit $15 billion in the first quarter and is “ascending rapidly.” CoreWeave and Meta have now ramped up their cloud-capacity agreement to approximately $21 billion, with some of the earliest deployments set to run on Nvidia’s Vera Rubin platform for inference—the part that generates answers from trained AI models. “Their most demanding workloads” are landing on CoreWeave’s cloud, CEO Michael Intrator said. Reuters
Nvidia’s messaging to investors has been that Rubin is on the near horizon. Back in March, Jensen Huang told investors the company had resumed making an AI chip tailored for China and noted, “Our supply chain is getting fired up.” Reuters later reported Rubin is already in full-scale production, and the chip is baked into Huang’s projection of topping $1 trillion in combined Blackwell and Rubin revenue by 2027. Reuters
Memory is the constraint. Samsung expects its first-quarter operating profit to top its entire 2025 total, driven by heavy AI infrastructure demand that’s tightening supply and sending prices up. In February, according to Reuters, the South Korean giant began shipping HBM4 — high-bandwidth, stacked memory built for AI workloads — to Nvidia. “Actual contract prices came in higher,” Meritz Securities analyst Kim Sunwoo said. Reuters
This week, Barron’s pointed to a KeyBanc client note where analyst John Vinh flagged a holdup for Nvidia’s Rubin GPU. According to Vinh, “The ramp of Nvidia’s Rubin GPU has been delayed” as HBM4 qualification over at SK Hynix—plus, though less of a factor, Micron—is dragging out longer than anticipated. Barron’s
Rivals are stepping up efforts to challenge Nvidia’s dominance just as delay risks emerge. Broadcom disclosed Monday it has secured a long-term agreement to help develop Google’s next-generation tensor processing units, or TPUs, through 2031—a move that underscores how major cloud players continue to search for less expensive or more customized options to Nvidia’s GPUs.
China’s proving a tough battleground. According to IDC figures cited by Reuters, Chinese AI chip makers grabbed almost 41% of the country’s AI accelerator server market last year. Nvidia’s slice slipped to 55%, while AMD managed roughly 4%—export curbs nudged customers to homegrown options, most notably Huawei.
Nvidia’s efforts to shore up its defenses are ramping up. The company just put $2 billion into Marvell, aiming to draw more semi-custom silicon and optical interconnect tech onto its own platform. Jacob Bourne at eMarketer pointed out Nvidia’s play targets areas where “bandwidth and power efficiency are key bottlenecks.” Reuters
Nvidia’s CFO Colette Kress, speaking as the company released quarterly earnings in February, made it clear they’re pushing hard on AI investment—especially with hyperscalers like Meta gearing up for a projected $630 billion outlay in 2026, much of it earmarked for data centers and processors. That’s the backdrop for why a hiccup with Rubin isn’t trivial: Blackwell’s momentum can only do so much. Any slip buys competitors time to experiment with custom silicon and alternative hardware.
