Design flaws have slowed the deployment of Nvidia’s Blackwell GPUs in the data center last year, prompting a redesign of silicon and packaging. However, this did not affect Nvidia’s research on mid-cycle refresh blackwell (Blackwell Ultra) GPUs in AI and HPC and next-generation Vera Rubin GPUs. Nvidia can already share details about the Rubin GPU and its postrubin products.
“Blackwell Ultra (scheduled later),” Nvidia CEO Jensen Huang reaffirmed analysts and investors during the company’s revenue conference call. “Next Train (IS) Blackwell Ultra with new networking, new (12-HI HBM3E) memory and of course new processors. Above.”
Later this year, NVIDIA plans to release a Blackwell B300 series solution for AI and HPC (formerly known as Blackwell Ultra), which will offer up to 288GB of memory, offering higher computational performance and eight stacks of 12-HI HBM4E memory. Unofficial information shows that the performance uplift made possible by NVIDIA’s B300 series will be around 50% compared to comparable B200 series products, but the company has not yet confirmed this.
To further improve performance, the B300 will be provided on NVIDIA’s Mellanox Spectrum Ultra X800 Ethernet Switch. Nvidia is also expected to provide partners with additional system design freedom with its B300 series data center GPUs.
Nvidia’s next-generation GPU series is based on the company’s all-new codename, Rubin Architecture, which further improves AI calculation capabilities as industry leaders march towards achieving artificial general information (AGI). In 2026, the first iteration of the data center Rubin GPU will feature eight stacks of HBM4E memory (up to 288GB). The Rubin platform also includes a Vera CPU, a 3600 GB/s NVLink 6 switch, a CX9 network card that supports 1,600 GB/s, and an X1600 switch.
Jensen Huang will be talking about Rubin at the GPU Technology Conference (GTC) in March, but he is not sure what he will discuss. Surprisingly, Nvidia will also be talking about GTC’s postrubin products. With Jensen Huang’s announcement this week, it is unclear whether the company will reveal details about the Rubin Ultra GPU or the GPU architecture that will come after the Rubin family.
Speaking of Rubin Ultra, this can certainly be a very groundbreaking product. It is predicted that Nvidia will come with a 12 stack HBM4E in 2027 as it learns how to efficiently use the 5.5-reticle-sized Cowos interposer and 100mm x 100mm board.
“Come to GTC. I’ll tell you Blackwell Ultra,” Huang said. “There’s Rubin. Then we show what the clicks afterwards are. It’s a really, really exciting new product.”
