NVIDIA's latest chip roadmap: launching Rubin GPU in 2026, Rubin Ultra in 2027, with the new generation of GPUs named after Feynman
Although the Blackwell B200 has just been fully put into production and the B300 will be launched in the second half of 2025, its successor products Vera Rubin (2026) and Rubin Ultra (2027) have already been planned, with significant performance improvements. The reasoning speed of Vera Rubin can reach 50 petaflops, with 288 GB of memory, making its performance 2.5 times that of Blackwell. A new architecture named after Feynman will also be introduced later
Blackwell has not yet been delivered on a large scale, and NVIDIA has already laid out two generations of successor products.
On March 18, local time, NVIDIA CEO Jensen Huang delivered a keynote speech at the GTC 25 conference, announcing the data center GPU roadmap for 2026-2027, with Rubin and Rubin Ultra set to succeed Blackwell, followed by a new architecture named after theoretical physicist Feynman.
Although Blackwell B200 has just begun full production and B300 is set to launch in the second half of 2025, NVIDIA has already started preparing for future technology transitions. Jensen Huang unveiled the product roadmap for the next two years, with a new generation of GPUs named after astronomer Vera Rubin set to launch in 2026. The performance of Vera Rubin NVLink 144 will be 3.3 times that of GB300 NVL 72.
Rubin can achieve a speed of 50 petaflops during inference, more than double the current Blackwell chip's speed of 20 petaflops per second. Rubin can also support up to 288 GB of fast memory.
After Vera Rubin, NVIDIA expects the next generation Rubin Ultra NVL576 to launch in the second half of 2027, with performance 14 times that of GB 300 NVL72.
Additionally, NVIDIA revealed the architecture codename following Rubin—Feynman, named after theoretical physicist Richard Feynman, further continuing the company's tradition of naming GPU architectures after scientists.
Notably, Jensen Huang stated that the name Blackwell is incorrect, as Blackwell B200 actually consists of two chips, which changes the topology of NVLink. Although Blackwell B200 is currently referred to as NVL72, Huang believes it would be more appropriate to call it NV144L.
This rethinking of the naming convention may indicate that the Rubin architecture will have further innovations in NVLink