288GB of HBM3e and 1.5x training throughput vs B200, shipping to hyperscalers in Q3.
Blackwell Ultra packs 288GB of HBM3e and a redesigned NVLink fabric, delivering 1.5x training throughput over B200. Microsoft, Meta, and Oracle are first in line, with general availability targeted for Q3. Nvidia also previewed Rubin, the next-gen architecture due in 2027.