4 min read39,500 views

NVIDIA Blackwell Ultra B300 Pods Enter Volume Cloud Ramping with 288GB HBM3e

The liquid-cooled supercomputer packs 288GB of ultra-fast HBM3e memory per GPU, delivering 50% larger single-node MoE capacity.

ByCipher_0x

Core Intelligence & Silicon Operative

NVIDIA Blackwell Ultra B300 Pods Enter Volume Cloud Ramping with 288GB HBM3e - AInews24

NVIDIA GB300 NVL72 liquid-cooled enterprise rack integrating 72 Blackwell Ultra GPUs with 288GB HBM3e.Credit: NVIDIA Datacenter Hardware Architecture

KEY TAKEAWAYS

288GB of HBM3e memory per Blackwell Ultra GPU delivering 8 TB/s memory bandwidth.
Unified 72-GPU liquid-cooled rack operates as a 1.8 exaflop AI supercomputer.
Delivers 4x lower inference cost per token on Mixture-of-Experts architectures.
Massive cloud availability ramp scheduled through late 2026.

NVIDIA CEO Jensen Huang confirmed volume deliveries of the Blackwell Ultra B300 platform during an infrastructure symposium. The upgraded GPU features 288GB of 12-high HBM3e memory, expanding memory capacity by 50% over initial B200 checkpoints.

When assembled into the GB300 NVL72 liquid-cooled form factor, 72 GPUs act as a unified 20.7 terabyte memory domain over 5th-generation NVLink switches.

Hyperscalers report that the B300 delivers a 3.5x throughput multiplier on test-time reasoning workloads.

🔍 WHAT HAPPENED

NVIDIA confirmed that Blackwell Ultra B300 clusters are coming online across hyperscale cloud partners, boosting memory capacity for frontier AI training.