During Nvidia GTC, Jensen Huang reveals the future GPUs up to 2028. First up is Blackwell Ultra, followed by Vera Rubin in 2026.
Blackwell is barely rolling off the production line, but Nvidia has a lot more in store for the coming years. During his keynote at GTC, Nvidia’s high mass, Jensen Huang reveals the plans for the coming years. In the second half of the year, Blackwell Ultra will appear: nowadays, CPUs and GPUs must also be equipped with a Pro, Max, or Ultra version.
The Ultra variant of Blackwell doesn’t bring more computing power: the 20 petaflops that the ‘base version’ can handle is already more than adequate. Nvidia makes the difference in memory with Blackwell Ultra. The on-chip memory is expanded to 288 GB compared to the current 192 GB.
It’s now exactly one year since Nvidia dropped the Blackwell bomb on the AI world. A year later, Nvidia is having a lot of problems shipping the chip in large volumes. A bug in the production process causes the chips to overheat when placed in a server rack. Nvidia is struggling to solve this problem.
Vera Rubin
In 2026, Nvidia will make the next generational leap with Vera Rubin. This generation will consist of a CPU Vera and a GPU Rubin. The power of Vera Rubin lies in the sum of its parts. In combination, the GPU and CPU deliver 50 petaflops of computing power for inference: more than double that of Blackwell. In the second half of 2027, Vera Rubin Ultra will double the number of petaflops once again.
The generation after Vera Rubin will bear the name of David Feynman. As this GPU generation is not expected until 2028 at the earliest, Nvidia hasn’t shared too many details about it yet.