NVIDIA GB300 NVL72 / 72 GPUS PER RACK / LIQUID COOLED
GB300 NVL72 cluster rental price: per rack and per GPU
Only two providers publish an on-demand GB300 NVL72 price: $11.59 and $18.00 per GPU-hour. At those rates one 72-GPU rack costs about $600,000 to $933,000 for a 30-day month. Every other provider quotes GB300 clusters privately.
Updated October 17, 2026. By GPU Rentals, powered by Bridgepointe Technologies. Every figure comes from the provider's own pricing page. Sources are listed at the end.
Per GPU-hour, 2 providers publish it
Per 30-day month at those rates
Per GPU-hour, from one provider's published discount
Our GB300 NVL72 reserved cluster rates
We compare each term against the lowest AWS, Microsoft or Google rate. None of the three publishes a GB300 rate, so we compare against Azure's published rate for the older GB200 NVL72 rack. Rates are per GPU-hour, billed monthly, with no upfront payment.
| Term | GPU Rentals rate per GPU-hour | Lowest AWS, Microsoft or Google rate |
|---|---|---|
| 12 months | $10.83 | $17.31, Azure 1-year, GB200 NVL72 |
| 24 months | $9.14 | Not offered |
| 36 months | $7.44 | $11.90, Azure 3-year, GB200 NVL72 |
| 60 months | $6.70 | Not offered |
AWS, Microsoft and Google sell only 1-year and 3-year terms, so 24 and 60 months have no direct comparison. Every rate depends on confirmed capacity for your GPU count, start date and region.
GB300 NVL72 cluster price per GPU-hour
A GB300 NVL72 is a full rack of 72 GPUs that work as one unit. Some providers rent it by the GPU or by the 4-GPU tray. Others sell only the whole rack.
| Provider | Term | Per GPU-hour | How it is sold |
|---|---|---|---|
| Verda | On-demand | $11.59 | Per GPU, clusters up to 144 GPUs |
| Verda | 12 months | $10.66 | Our math: published 8% term discount |
| Verda | 24 months | $8.69 | Our math: published 25% term discount |
| Oracle Cloud | On-demand | $18.00 | BM.GPU.GB300.4, a 4-GPU bare metal node |
CoreWeave, Microsoft Azure, AWS, Google Cloud, Lambda, Nebius, Together AI, Hyperstack and Vultr offer GB300 but quote it privately. Vultr sells it only as a full 72-GPU rack. Spot rates are left out because the provider can take the GPUs back.
GB200 NVL72 cluster prices, for comparison
The GB200 NVL72 is the previous rack generation. More providers publish its price, which shows how far long terms bring the rate down.
| Provider | Term | Per GPU-hour | How it is sold |
|---|---|---|---|
| CoreWeave | On-demand | $10.50 | 4-GPU instance at $42.00 per hour |
| AWS | Capacity Block | $10.58 | 36 or 72 GPUs, prepaid, Dallas only |
| Oracle Cloud | On-demand | $16.00 | 4-GPU bare metal node |
| Microsoft Azure | Pay as you go | $27.04 | 4-GPU VM, West US 3 |
| Microsoft Azure | 1-year reserved | $17.31 | Same VM |
| Microsoft Azure | 3-year reserved | $11.90 | Same VM |
Azure's 3-year rate is 56% below its pay-as-you-go rate. Expect GB300 terms to move the same way once more providers publish them.
How GB300 NVL72 clusters are sized: 72, 144, 576 GPUs
The rack is the building block. Each one holds 72 GPUs and 36 Grace CPUs in 18 compute trays, joined by 9 NVLink switch trays so all 72 GPUs act as one. Clusters grow one rack at a time: 72, 144, 216, 288 and up. A request for about 500 GPUs becomes 7 racks (504 GPUs) or 8 racks (576 GPUs).
A rack needs up to 142 kW of power and is fully liquid cooled. That is why most buyers rent it in a provider's data center rather than install it themselves.
Pick GB300 NVL72 when the model is too large to split well across 8-GPU servers, such as the largest training runs and long-context inference. For most other work, B300 clusters or B200 clusters cost less per GPU.
GB300 runs on Arm-based Grace CPUs. Software built only for x86 servers may need work before it runs.
GET A GB300 CLUSTER QUOTE
GB300 NVL72 racks.
Checked for you.
Tell us how many GPUs or racks you need, for how long, and when. We check providers and reply by email with options and pricing. Nothing is booked or charged.
REQUEST SENT
Thanks. We have your GB300 request.
We will reply by email with provider options and pricing. Nothing has been booked or charged.
GB300 NVL72 QUESTIONS
Before you rent a GB300 cluster.
What are GPU Rentals' GB300 NVL72 cluster rates?
Our reserved GB300 NVL72 rates are $10.83 per GPU-hour for 12 months, $9.14 for 24 months, $7.44 for 36 months and $6.70 for 60 months, billed monthly. Rates are subject to confirmed capacity.
How much does it cost to rent a GB300 NVL72 rack?
At published on-demand rates of $11.59 to $18.00 per GPU-hour, one 72-GPU rack costs about $834 to $1,296 per hour. That is about $600,000 to $933,000 for a 30-day month.
How much does a 144-GPU GB300 cluster cost per month?
Two racks cost about $1.2 million to $1.87 million for a 30-day month at published on-demand rates. At the 24-month rate of $8.69 per GPU-hour, two racks are about $901,000 a month.
Is a reserved GB300 cluster cheaper?
Yes. Verda's published 24-month discount brings its rate to $8.69 per GPU-hour, 25% below on-demand. On the older GB200, Azure's 3-year rate is 56% below pay as you go. Most providers quote GB300 terms privately.
Can I rent part of a GB300 NVL72 rack?
Some providers rent single GPUs or 4-GPU trays. Others, like Vultr, sell only the full 72-GPU rack. The full benefit of NVL72 comes from all 72 GPUs working as one.
What is the difference between GB300 NVL72 and a B300 cluster?
A B300 cluster is built from 8-GPU servers joined over the network. A GB300 NVL72 joins 72 GPUs inside one liquid-cooled rack, with Arm-based Grace CPUs. It suits the largest models and costs more per GPU.
How much power does a GB300 NVL72 rack need?
Up to 142 kW per rack, with liquid cooling. Most data centers cannot host that without special preparation, which is why renting from a provider is common.
Sources, checked October 17, 2026
Verda: verda.com/pricing. Oracle: oracle.com/cloud/price-list. CoreWeave: coreweave.com/pricing. AWS Capacity Blocks: aws.amazon.com/ec2/capacityblocks/pricing. Microsoft Azure retail prices API, ND GB200 v6, West US 3, effective April 1, 2025: prices.azure.com. Vultr: vultr.com/pricing. NVIDIA GB300 NVL72: nvidia.com. NVIDIA NVL72 reference architecture, updated September 15, 2026: docs.nvidia.com.
Prices change often and are shown as published. Monthly figures use 720 hours. A published price is not a promise that capacity is available.