Rack-scale GPU systems: GB300 NVL72 and Vera Rubin NVL72

Rack-scale GPU systems turn a whole rack into one NVLink domain. In an NVIDIA GB300 NVL72, 72 Blackwell Ultra GPUs and 36 Grace CPUs are connected so that tensor, expert and pipeline parallel traffic can stay on NVLink across all 72 GPUs instead of crossing a network. We rent GB300 NVL72 on 12 to 36 month terms, and we are taking reservations for Vera Rubin NVL72.

How a 72-GPU NVLink domain works

In an 8-GPU HGX server, NVLink connects eight GPUs. Anything beyond those eight crosses the cluster network, which offers far less bandwidth per GPU than NVLink. Large training and serving jobs are usually laid out around that boundary: tensor parallelism inside a server, data or pipeline parallelism between servers.

An NVL72 moves the boundary to 72 GPUs. NVIDIA publishes 130 TB/s of total NVLink bandwidth for the rack. Each GPU keeps 288 GB of HBM3e at about 8 TB/s, for about 20 TB of HBM3e per rack, and the 36 Grace CPUs add 17 TB of LPDDR5X. NVIDIA's vendor-published figures also put each GPU in the NVL72 at 5 PFLOPS of dense FP8, against 4.5 PFLOPS for a B300 in an HGX server.

The rack is liquid-cooled. Cooling at this density needs facility-level support, which is part of what you rent rather than build. The full vendor figures are on our GB300 NVL72 page.

GB300 NVL72 rental: terms and pricing

GB300 NVL72 starts at $7,990 per GPU per month. For a full rack of 72 GPUs, that starting rate works out to $575,280 a month. Terms run from 12 to 36 months, unlike our servers, which you can rent month to month. Our team confirms each system in your dashboard with its configuration, price and location before you pay.

A 12-month minimum makes sense for a sustained program of work, such as a pre-training run followed by post-training and evaluation. If you are not sure your workload needs a 72-GPU domain, read the comparison below first, or start with an HGX B300 server on monthly terms.

Vera Rubin NVL72 reservations

Vera Rubin NVL72 is the next NVIDIA rack-scale generation: 72 Rubin GPUs with 288 GB of HBM4 each, 36 Vera CPUs and NVLink 6, in a liquid-cooled rack. Availability is limited in late 2026, and reservations are open. There is no list price yet: the price comes with the confirmation of your reservation. In the notes of a reservation request, give the scale you need, the date you need it and the term you have in mind, and we come back with what we can deliver. Nothing is charged when you send it.

Rack-scale GPU systems or several HGX servers?

Seventy-two B300 GPUs can also be nine HGX B300 servers. The right choice depends on how far your model has to be split.

Choose an NVL72 when tensor or expert parallelism has to span more than eight GPUs: very large mixture-of-experts models whose all-to-all traffic would otherwise cross the network, or serving setups whose weights and KV cache outgrow an 8-GPU server. Choose HGX servers when your parallelism fits inside eight GPUs and data parallelism handles the rest. You keep monthly flexibility and pay less per GPU, and we can link the servers with private networking.

How a rack-scale order works

1. Send a reservation request from the deployment page, creating a free account or logging in on the same form, and describe the workload, scale, start date and term in the notes. Nothing is charged when you send it.
2. Our engineers scope the system from your request: configuration and location.
3. Rack-scale orders may require verification under our verification policy. We tell you exactly what is needed and why.
4. We confirm the configuration, price and location in writing, in your dashboard and by email, before you pay.

Scoping goes faster when you bring:

• the model size, the precision and the parallelism layout you plan, such as tensor, expert and pipeline sizes;
• how much data you need to move in, and where checkpoints should be copied;
• your software stack. Grace is an Arm CPU, so containers and Python dependencies need arm64 builds. Check custom packages before you commit to a term.

Support from engineers is available 24/7 for the whole term.