NVIDIA GB300 NVL72 by HPE

One integrated rack: 18 liquid-cooled compute trays (36 NVIDIA Grace CPUs and 72 Blackwell Ultra GPUs across 2 GB300 boards per tray), liquid-cooled NVLink switch trays for the 72-GPU NVLink domain, 8 power shelves on a DC bus bar, and your choice of NVIDIA Quantum-X800 InfiniBand or Spectrum-X Ethernet networking with ConnectX-8 SuperNICs. HPE integrates it as a single delivered unit rather than a parts list you assemble on-site.

As an authorized HPE partner, we scope and quote the GB300 NVL72 rack and the surrounding AI Factory configuration for you, including networking, services, and support. We can source TAA-compliant configurations and support purchases through GSA MAS (application in progress), SAP/FAR channels, and GPC direct, plus GPC for smaller line items, alongside HPE GreenLake consumption if you prefer an as-a-service model over a capital purchase.

Plan for roughly 132 kW nominal thermal design power per rack, with electrical design power peaking around 155 kW and HPE recommending bus way provisioning up to 192 kW EDPp. Cooling is direct liquid cooling at the rack, capturing about 90% of heat to liquid and 10% to air. This is not a retrofit into a standard air-cooled row: your facility needs liquid cooling distribution (CDUs) and power delivery sized for this density before the rack lands.

GB300 swaps in Blackwell Ultra GPUs with 288 GB of HBM3e memory per GPU (up from 192 GB), delivering roughly 1.5x the dense FP4 compute and materially higher per-GPU token throughput on large reasoning models. Each compute tray also shifts to 4 GPUs plus 2 Grace CPUs, versus 2 GPUs per tray on GB200. If you already run GB200 NVL72, GB300 is a memory and inference-throughput upgrade on the same NVLink rack-scale architecture, not a different platform to relearn.

Sizing starts with your workloads: model sizes, training versus inference mix, context length, dataset volumes, and concurrent user counts. From there we map GPU count, whether a single 72-GPU NVLink domain covers it or you need multiple racks networked together, and which fabric (InfiniBand or Ethernet) fits your operations team. We can arrange access to HPE and NVIDIA technical resources to validate the design before you commit budget.

Yes. HPE positions the GB300 NVL72 within its AI Factory portfolio for service providers, sovereign entities, and large enterprises, with configurations aimed at regulated industries and national AI programs that need in-country data control. We help align the specific configuration, contract vehicle, and support terms to your regulatory and security requirements as an authorized HPE partner.

GB300 NVL72 is shipping today and running in production at hyperscale (Vultr's global buildout is one public example), while HPE has already announced its successor, the Vera Rubin NVL72 platform, for December 2026 availability. If you need trillion-parameter-class capacity in 2026, GB300 is the proven, supportable choice now; if your timeline allows waiting a year and you want the newest silicon, it's worth discussing both paths with us before you commit.

Submit your workload details and target timeline through our quote request, and we'll come back with a scoped configuration covering the rack, networking fabric, facility prerequisites, and support tier, along with applicable contract-vehicle pricing for federal, SLED, healthcare, or enterprise procurement.