Rent NVIDIA Vera Rubin NVL72: reserve rack-scale capacity

A Vera Rubin NVL72 reservation lets you claim a place for NVIDIA's next rack-scale system before it ships. Reservations are open now: you register interest, and we confirm allocation and terms in writing before any payment. If you want to rent Vera Rubin capacity for training or inference, this page covers what is known, how reserving works, and what to run in the meantime.

What we can say about Vera Rubin NVL72 today

• GPUs 72 Rubin GPUs per rack, with 288 GB of HBM4 each, about 20.7 TB in total
• CPUs 36 Vera CPUs
• Interconnect NVLink 6, the generation after the NVLink 5 used in Blackwell systems
• Form Liquid-cooled, rack-scale system
• Availability Limited, expected in late 2026

That list is deliberately short. The compute and bandwidth figures in the specifications above are preliminary numbers published by NVIDIA and may change before systems ship; we confirm the exact configuration of the systems we rent in writing. What you can plan with today is memory: 288 GB per GPU, the same capacity as the B300, now in HBM4, a newer memory generation than the HBM3e of Blackwell systems.

How a Vera Rubin NVL72 reservation works

1. Register interest. Create a free account and send a reservation request from the deployment page, with these details in the notes: what you plan to run, how many GPUs or racks you need, the term you have in mind, and your earliest and latest acceptable start dates.
2. We confirm in writing. When we can allocate capacity, we send the allocation, price and term in writing. Nothing is due before this confirmation.
3. You decide. If the terms work for you, you accept them; if not, you walk away without having paid anything.
4. Payment. You pay from your account balance, which you top up in crypto.

For an order of this size, we may ask for verification of your company and intended use before confirming, as described in our verification policy. Sharing how flexible your dates are helps us match you with the capacity we can allocate.

Should you reserve Vera Rubin or rent what ships now?

A reservation makes sense when your plans already point past what is available today:

• You have a training roadmap that runs into 2027 and calls for a 72-GPU NVLink domain on the newest generation.
• Your models, or their training state, already outgrow an 8-GPU server, so you will need rack-scale capacity in any case.
• You want an allocation and a price in writing before you commit budget.

It makes less sense if you need GPUs this quarter, if your jobs fit in eight GPUs, or if you are still deciding what to train; monthly servers keep your options open in those cases. If you do reserve, a useful request states the model size and precision you plan to run, the number of GPUs, the term, your start window and any constraints on location, so that our written answer can be specific.

Planning a Vera Rubin workload now

Because memory per GPU matches the B300, you can size Vera Rubin deployments today with the same arithmetic. Weights take roughly parameters × bytes per parameter (2 in BF16, 1 in FP8, 0.5 in FP4), plus KV cache and 10–30% headroom, and full training with Adam needs roughly 16–20 bytes per parameter before activations.

• Inference of a one-trillion-parameter model in FP8: about 1 TB of weights, so at least four GPUs for the weights alone; a full rack leaves roughly 19 TB for KV cache, activations and other models.
• Full training of a 405B-class model: about 6.5–8.1 TB of training state, or 23 to 29 GPUs' worth of memory before activations, well within one 72-GPU rack.
• Full training of a one-trillion-parameter model: about 16–20 TB, close to the rack's entire 20.7 TB of HBM4, so plan for activation memory on top, either across racks or with offloading.

Our VRAM calculator runs these numbers for other sizes. On the software side, a new GPU architecture usually means new driver, CUDA and framework versions, so keep your stack containerized and pinned, and plan time to validate it on the new systems before you move production work.

Need capacity before Vera Rubin ships?

Two options are available today, with up to 288 GB of memory per GPU:

• HGX B300 servers: eight Blackwell Ultra GPUs per server at $39,900 per month, with a one-month minimum term. Because plans run month to month, you can keep them until your Vera Rubin capacity is delivered and then cancel with no fee. Stock is limited.
• GB300 NVL72: 72 Blackwell Ultra GPUs in one NVLink domain, from $7,990 per GPU per month on 12 to 36 month terms. Choose it if you need a rack-scale NVLink domain now, and set the term length with your Vera Rubin timeline in mind, since the two commitments would otherwise overlap.

One practical approach is to start training on HGX B300 servers, keep checkpoints and data pipelines portable, and move the next run to Vera Rubin once it is delivered. HGX B300 servers are ordered directly on the deployment page and delivered in under 10 minutes after payment; your Vera Rubin reservation request is sent from the same page.

Vera Rubin NVL72 pricing

There is no list price for Vera Rubin NVL72 yet. Each reservation is confirmed in writing with its configuration, price and term, and nothing is payable until you accept it. Prices are in US dollars and exclude VAT and other taxes where applicable. When the time comes, you pay from your account balance, topped up in crypto: Bitcoin (BTC), Ethereum (ETH), Monero (XMR), USDT on TRON (TRC-20) or Solana (SOL), at an exchange rate locked for a limited time for each top-up. Card payments are coming soon, and every payment comes with an invoice in your dashboard. For an overview of everything we offer at this scale, see rack-scale systems.