Nvidia's GB300 NVL72 is the Blackwell Ultra generation's flagship rack, and it is already showing up in production: a Microsoft Azure cluster built for OpenAI, and — in a less obvious venue — a U.S. military university. For anyone planning data-center power or cooling capacity, the GB300 is the number to build around for the next planning cycle.

We pulled the rack's published specs, its real-world deployments, and its price tag together to show what a single one of these racks demands from a facility, and what it costs to own.

Key takeaways

  • A single GB300 NVL72 rack draws 132–142 kW at nominal load and peaks near 155 kW — roughly 15x the ~9 kW an average enterprise rack pulls, per the Uptime Institute's 2025 survey.
  • The rack packs 72 Blackwell Ultra GPUs and 36 Grace CPUs, about 20 TB of pooled HBM3E memory, and roughly 130 TB/s of NVLink aggregate bandwidth.
  • Liquid cooling removes about 90% of the rack's heat; integrators are shipping in-rack CDUs rated around 250 kW to keep up.
  • A GB300 NVL72 rack runs $3.7–$4.0 million; Microsoft's Azure cluster for OpenAI already uses more than 4,600 of them.
  • Nvidia donated a DGX GB300 system to the Naval Postgraduate School, activated July 22, 2026, giving over 1,500 resident students and about 600 faculty access to it.

What's inside a GB300 NVL72

Each rack integrates 72 Blackwell Ultra GPUs alongside 36 Grace CPUs. Every GPU carries 288 GB of HBM3E memory in a 12-Hi stack, good for 8 TB/s of bandwidth per chip — which is also the number Nvidia and integrators cite for the standalone B300 GPU. Pooled across the rack, that adds up to roughly 20 TB of shared HBM3E, tied together by NVLink 5 at an aggregate bandwidth of about 130 TB/s all-to-all. Each GPU is rated around 1,400 W TDP. For scale-out networking beyond the rack, Nvidia pairs the system with ConnectX-8 SuperNICs delivering 800 Gb/s per GPU.

Spec Value
GPUs per rack 72 Blackwell Ultra
CPUs per rack 36 Grace
GPU memory 288 GB HBM3E per GPU (~20 TB pooled)
Memory bandwidth 8 TB/s per GPU; ~130 TB/s NVLink aggregate
GPU TDP ~1,400 W
Rack power draw 132–142 kW nominal, ~155 kW peak
Rack weight ~1,580 kg
Scale-out networking ConnectX-8 SuperNICs, 800 Gb/s per GPU

The power and cooling math

The headline number for facility planners is the draw: 132–142 kW at nominal load, spiking to around 155 kW. The Uptime Institute's 2025 survey puts the average enterprise rack at about 9 kW — so a single GB300 NVL72 is doing the work (and pulling the power) of more than a dozen ordinary racks in the same footprint.

That density is only survivable with liquid cooling. Roughly 90% of the heat the rack generates is rejected into liquid rather than air, and integrators are now shipping in-rack coolant distribution units (CDUs) rated for around 250 kW of cooling capacity to match. For operators still running air-cooled halls, a GB300 deployment is effectively a forcing function to bring liquid cooling in-row or in-rack.

What it costs, and who's buying at scale

Loop Capital's analyst estimate puts a GB300 NVL72 rack at $3.7–$4.0 million. That's the unit economics; the deployment economics are what make the number matter. Microsoft's Azure buildout for OpenAI is already running more than 4,600 of these racks — among the largest known GB300 deployments disclosed to date. Nvidia has claimed the platform delivers up to 50x the AI-factory output of its Hopper-generation systems, though that figure is Nvidia's own claim rather than an independently verified benchmark.

Beyond hyperscalers: a GB300 at a military university

Not every GB300 deployment is inside a hyperscaler's cloud. Nvidia activated a DGX GB300 supercomputer at the Naval Postgraduate School on July 22, 2026, during the school's Converge @ NPS event, donating the system to the university's foundation. DDN, VAST Data, and Vertiv contributed storage, data management, cooling, and integration for the deployment. Once online, it will be available to more than 1,500 resident students and about 600 faculty — a sign that GB300-class compute is starting to reach research institutions, not just the handful of companies that can absorb a multi-thousand-rack order.

What this means for planning

The throughline across every GB300 deployment we found — hyperscaler or university — is the same: power and cooling, not chip supply, are the constraints operators are now designing around. A facility that can't deliver 140+ kW per rack and reject 90% of that as liquid heat isn't in the running for this generation of hardware, regardless of budget. That's the planning question the GB300 ramp is forcing across the industry over the next year.

Sources

This article was researched and fact-checked against the following sources: