NeuralStackNvidia announced the Blackwell Ultra GPU on July 19 2026. The chip contains 288 GB of HBM4 memory and delivers 4.2 petaflops of FP8 performance. It will enter volume production in Q1 2027.
The new architecture adds native support for 8-bit floating point training. Nvidia claims a 2.8 times improvement over the original Blackwell on large language model training. Power consumption is rated at 1400 watts.
Blackwell Ultra builds on the B200 design released in 2025. The company has already received orders from Microsoft and Meta for 2027 deployments.
TSMC will manufacture the chips on its 2nm process. Nvidia also introduced a new NVLink 6 interconnect with 1.6 TB/s bandwidth.
Why this matters
The release keeps Nvidia ahead in the race for memory-dense accelerators needed for trillion-parameter models. Cloud providers will redesign racks around the higher power and bandwidth requirements.
AMD and Intel face increased pressure to match memory capacity in their next generations. Nvidia expects 35 billion dollars in Blackwell Ultra revenue in 2027.