GB300 NVL72 rental from $7,990 per GPU-month

A GB300 NVL72 rental gives you 72 NVIDIA Blackwell Ultra GPUs and 36 Grace CPUs in one liquid-cooled rack, all in a single NVLink domain. We rent it on 12 to 36 month terms, from $7,990 per GPU per month. This page explains what the system is, who actually needs it, and when several HGX B300 servers are the better buy.

What a GB300 NVL72 is

The GB300 NVL72 is NVIDIA's rack-scale Blackwell Ultra system. Instead of eight GPUs per server, it connects 72 GPUs through NVLink switches into one domain, with 130 TB/s of total NVLink bandwidth per NVIDIA's figures. NVIDIA lists 20 TB of HBM3e per rack, about 277 GB per GPU (up to 288 GB on the B300 module), at 8 TB/s per GPU. Each of the 36 Grace CPUs is paired with two GPUs, and together they add 17 TB of LPDDR5X.

Per GPU, NVIDIA lists 15 PFLOPS of dense FP4, 5 PFLOPS of dense FP8 and 2.5 PFLOPS of dense BF16 for the GB300 NVL72, slightly above its figures for the B300 in an HGX server (13.5, 4.5 and 2.25 PFLOPS). The rack is liquid-cooled and hosted by us, so you get the capacity without building a facility around it.

Who needs a rack instead of HGX B300 servers

72 GPUs could also be nine HGX B300 servers. The GPUs belong to the same Blackwell Ultra family and memory per GPU is similar; the difference is how they talk to each other.

The bandwidth gap

Inside an HGX B300 server, each GPU has 1.8 TB/s of NVLink bandwidth, but only to the other seven. Between servers, traffic goes through the network: eight 800G adapters per server, or 800 Gb/s (100 GB/s) per GPU. In a GB300 NVL72, every GPU keeps its 1.8 TB/s of NVLink bandwidth to all 71 others, about 18 times what crosses the network per GPU on HGX servers.

Workloads that need it

• Wide expert parallelism: large mixture-of-experts models whose experts are spread across dozens of GPUs exchange tokens all-to-all at every MoE layer. In a 72-GPU NVLink domain, that traffic never touches the network.
• Frontier-scale training: full training with Adam needs roughly 16–20 bytes per parameter before activations, so a one-trillion-parameter model needs about 16–20 TB, about the rack's entire HBM3e. On HGX B300 servers, the same state would span seven to nine servers linked by the network.
• Tensor or pipeline parallelism beyond eight GPUs, where each step synchronizes more GPUs than one HGX server holds.

When HGX B300 servers are the better choice

If each job fits in eight GPUs, the rack's NVLink domain goes unused. That covers most fine-tuning and the inference of models up to several hundred billion parameters: a 671B model in FP8 fits on one HGX B300 server. Data-parallel training, where each replica fits in one server and only gradients cross the network, is far less sensitive to the gap. In those cases, HGX B300 servers cost less per GPU ($4,987.50 per GPU per month at $39,900 per server), start with a one-month term and can be added one at a time, subject to stock.

Memory math at rack scale

For inference, capacity alone is rarely the reason to choose a rack. A one-trillion-parameter model in FP8 needs about 1 TB of weights, which already fits on one HGX B300 server with 2.3 TB. What the rack adds is room to serve such models with very large KV caches and high concurrency, and the bandwidth to spread them across many GPUs. With about 20 TB of HBM3e, a 1T model in FP8 leaves roughly 19 TB for KV cache and activations, or you can run several large models side by side in one domain. The 17 TB of LPDDR5X attached to the Grace CPUs adds space for data you offload from the GPUs.

Software and operations

The Grace CPUs are Arm-based, so your container images and compiled dependencies need arm64 builds. NVIDIA ships arm64 builds of its core software, but check your own images and Python wheels early. Multi-GPU communication uses NCCL, as on HGX servers, with the difference that collectives across all 72 GPUs stay on NVLink. The software baseline, access method and network setup for your rack are part of the written configuration we confirm before you pay. Our multi-GPU NCCL documentation covers the communication side.

GB300 NVL72 rental pricing and terms

• Price From $7,990 per GPU per month, about $575,000 per month for 72 GPUs
• Term 12 to 36 months
• Per GPU-hour equivalent From about $10.95, the monthly price per GPU divided by 730 hours
• Availability By reservation request from the deployment page, with the price confirmed before you pay

Compared with HGX B300 servers at $4,987.50 per GPU per month, the rack costs about 60% more per GPU. The premium buys the 72-GPU NVLink domain, the Grace CPUs and their memory, and a liquid-cooled rack. Nine HGX B300 servers would come to $359,100 per month before any volume pricing, so the rack is worth it only when your workload uses the domain. At this scale, the comparison is usually with other reserved contracts rather than hourly billing: in 2026, many providers prefer one-year contracts for large capacity, and on-demand stock is often sold out. Your reservation confirmation states the exact price for your term, configuration and start date.

How to rent GB300 NVL72 capacity

1. Send a reservation request from the deployment page, creating a free account or logging in on the same form, with these details in the notes: what you plan to run, how many GPUs you need, the term and your target start date.
2. We confirm the configuration, price, term and location in your dashboard and by email. Nothing is charged before you accept.
3. Verification may be requested for an order of this size; our verification guide explains what we may ask for and why.
4. You accept the confirmed price and pay from your account balance, topped up in crypto (card payments are coming soon), and access details appear in your dashboard.

If your timeline allows it, the Vera Rubin NVL72 is open for reservations; for capacity today, the GB300 NVL72 and HGX B300 are the options. More on our rack-scale systems.