NVIDIA GB300: Rack-Scale Grace Blackwell for the Largest Clusters

NVL72's entire premise is treating a full rack as one coherent accelerator — every GPU sees every other GPU's memory over NVLink at rack scale, rather than being bottlenecked at the 8-GPU node boundary. That matters most for frontier-scale training and for inference workloads serving the largest models with the lowest latency.

It's a liquid-cooled system, which means it needs a site built or retrofitted for direct-liquid-cooling — one of the specific things we evaluate when matching a GB300 deployment to a powered site.

Same two paths as our other platforms, sized to rack-scale deployments:

Not every powered site can take a GB300 deployment without retrofit — direct-liquid-cooling plant, CDUs, and piping are real infrastructure asks. This is exactly the kind of site-viability question we run before matching a deployment, not something discovered mid-buildout.

If your training runs are bottlenecked by rack-scale memory and bandwidth rather than raw GPU count — the largest frontier-scale training jobs and the highest-throughput, lowest-latency inference for the biggest models — NVL72's coherent 72-GPU domain is built specifically for that problem. It's also the platform to reach for if you're deploying into a facility already built or being built for direct liquid cooling.

If your workloads fit within an 8-GPU NVLink domain and your site runs conventional air-cooled infrastructure, B300 gets you the same Blackwell Ultra generation without the cooling-plant requirement — see the B300 page for the fuller comparison.

Rack-scale deployments carry more site-side variables than an 8-GPU node — cooling infrastructure chief among them — so timelines depend more heavily on whether the target site already has direct-liquid-cooling capability or needs it built. GPUs-as-a-Service capacity on infrastructure we've already built for liquid cooling is the faster path when your timeline is tight; a dedicated buildout on a matched site follows our standard qualify-assess-match-execute process.

Target scale (racks or GPU count), deployment type, timeline, and region preference. If you already know your cooling and power requirements, tell us — it speeds up matching considerably.