The Rack-Scale AI Factory Arrives
Nvidia has officially initiated full-scale mass production and global shipments of its flagship Blackwell Ultra computing platforms, spearheaded by the GB200 NVL72 liquid-cooled rack architecture.
Each GB200 NVL72 enclosure functions as a single unified supercomputer, consolidating 72 Blackwell GPUs and 36 Grace CPUs interconnected via fifth-generation NVLink copper spine cables delivering 1.8 terabytes per second of bidirectional bandwidth per GPU.
“Blackwell is not merely a microchip; it is a full-stack datacenter architecture engineered to build the world’s generative intelligence engines,” declared Nvidia CEO Jensen Huang during an investor briefing.
Zero-Downtime Closed-Loop Liquid Cooling
With aggregate power draws approaching 120 kilowatts per rack, the NVL72 relies exclusively on liquid-to-liquid heat exchangers. Hyperscale cloud providers including Microsoft Azure, AWS, and Google Cloud are re-architecting their datacenter mechanical designs to support high-density coolant distribution manifolds.
- Inference Speedup: 30x over Hopper H100 generation on FP4 models
- FP4 Compute Density: 1.44 Exaflops per single rack
- Total NVLink Spine Bandwidth: 130 TB/s aggregate