NVIDIA Blackwell-2 Ultra: Liquid-First

NVIDIA has officially unveiled the **Blackwell-2 Ultra (B2U)** architecture, marking a pivotal shift in data center engineering. For the first time, liquid cooling is not an option—it is the **native design requirement**, enabling a 2.5x increase in thermal density over the original Blackwell series.

HBM4e & 12TB/s Bandwidth

The core of the B2U is the integration of **HBM4e memory**, delivering a record-breaking **12TB/s of memory bandwidth**. This is specifically engineered to handle the massive weights of trillion-parameter agentic models that require near-instantaneous state transitions.

The Cooling Mandate

By mandating direct-to-chip liquid cooling, NVIDIA has eliminated the physical constraints of massive heat sinks and high-velocity fans. This allows for tighter rack density, where a single **GB300 rack** can now deliver over **1.5 Exaflops** of AI compute power.

Data center providers like **Equinix** and **Digital Realty** have already begun retrofitting facilities to accommodate the coolant manifold requirements of the B2U architecture.