On the near horizon — H2 2025 — Nvidia will release Blackwell Ultra, a GPU that’ll come in several configurations. A single Ultra chip will offer the same 20 petaflops of AI performance as Blackwell, but with 288GB of memory — up from 192GB in vanilla Blackwell.

Nvidia will also offer a single rack called the GB300 NVL72 that offers 1.1 exaflops of FP4, 20TB of HBM memory, 40TB of “fast memory,” and 130TB/sec of NVLink bandwidth and 14.4 TB/sec networking.

The Vera Rubin die will have two GPUs sandwiched together, delivering over 50 petaflops of FP4 performance per chip. Nvidia claims that a full rack of these will deliver over 3.6 exaflops of performance—that's more than three times what's possible on existing Blackwell hardware.

Chief Jensen Huang unveils Vera Rubin semiconductor and says appetite for ever-faster ones and software will continue to grow

Jensen Huang headlined the Nvidia GTC 2025 keynote, giving updates on Blackwell, Vera Rubin, and a new Disney and Google Deepmind partnership.

Jensen Huang, Nvidia Corp.’s chief executive officer, used his annual developer conference to address concerns that the cost of artificial intelligence computing is spiraling out of control.