How much does it cost to rent a B300 cluster?
A B300 is sold inside an 8-GPU server, so clusters grow in steps of 8. The table prices three common sizes for a 30-day month (720 hours) at the lowest and highest published on-demand rates, and at the lowest published 36-month reserved rate.
| Cluster | On demand, $6.60/GPU-hr | On demand, $15.00/GPU-hr | 36-month reserved, $4.70/GPU-hr |
|---|---|---|---|
| 8 GPUs (1 server) | $38,016 | $86,400 | $27,072 |
| 64 GPUs (8 servers) | $304,128 | $691,200 | $216,576 |
| 512 GPUs (64 servers) | $2.43 million | $5.53 million | $1.73 million |
512 GPUs is not an arbitrary size. NVIDIA's DGX B300 SuperPOD reference design is built from units of 64 servers, which is 512 GPUs. Clusters that large are negotiated one contract at a time, so a quote can land below the published floor.
Published B300 on-demand rates
On demand means you pay by the hour and can stop at any time. It is the most expensive way to run a B300 for months. Spot and preemptible rates are left out, because the provider can take the GPUs back.
| Provider | Per GPU-hour | How it is sold |
|---|---|---|
| Massed Compute | $6.60 | 8-GPU instance at $52.80 an hour |
| Hyperstack | $7.40 | Per GPU |
| Civo | $7.49 | 8-GPU node |
| Latitude.sh | $8.00 | 8-GPU bare metal server at $64 an hour |
| Runpod | $8.99 | Secure Cloud |
| Nebius | $9.50 | Per GPU |
| Together AI | $9.99 | GPU cluster |
| Verda | $10.08 | Self-service instance |
| Oracle Cloud | $15.00 | BM.GPU.B300.8 bare metal |
AWS sells B300 as short-term Capacity Blocks at $16.15 per GPU-hour, prepaid for 1 to 182 days.
B300 reserved cluster rates by term
A reserved term commits you to a set number of GPUs for months or years, and the hourly rate drops in return. Only a few providers post their reserved rates, and almost none post a 60-month rate. Most quote them privately.
| Provider | 12 mo | 24 mo | 36 mo | 60 mo |
|---|---|---|---|---|
| Our rates (GPU Rentals) | $8.77 | $6.79 | $4.81 | $4.33 |
| Civo | $6.49 | $5.99 | $5.49 | Private |
| TheAI Cloud | None | $5.00 | $4.70 | Private |
| DigitalOcean | $7.94 | Private | Private | Private |
Our rates are billed monthly with no upfront payment. Civo rents a whole 8-GPU node. TheAI Cloud rents whole 8-GPU nodes, does not offer 12 months, and takes 30% at signing. DigitalOcean sells its 12-month term through sales.
The first row is ours. GPU Rentals is our GPU rental service, also powered by Bridgepointe Technologies, and the only one here that posts a 60-month rate. Its terms and a reserved-cluster quote form are on gpurentals.ai. Flexicloud posts only "from $4.50" for its longest term, without saying how long. Nebius says its reserved B300 runs up to 35% below on demand, and CoreWeave says committed use runs up to 60% below, but neither posts a dollar rate. Lambda offers B300 only in clusters of 4,000 or more GPUs on terms of three years or more.
What it costs to own a B300 cluster and colocate it
The alternative to renting is buying the servers and placing them in a colocation facility. Three numbers decide it: the server price, the power each server draws, and what colocation costs per kW.
- Server price. October 2026 reseller listings and appraisals put a configured 8-GPU HGX B300 server at roughly $430,000 to $550,000. Volume orders come in lower.
- Power. NVIDIA rates a DGX B300 at about 14 kW. A 64-GPU cluster is eight servers, so about 112 kW of GPU load before network switches and storage.
- Colocation. Our recent quotes for up to 500 kW ran $180 to $400 per kW per month, depending on region. At 14 kW, that is $2,520 to $5,600 a month per server.
| Cost line | Per GPU-hour | How it is worked out |
|---|---|---|
| Server, paid off over 36 months | $2.07 to $2.65 | $430,000 to $550,000 spread over 36 months of 5,760 GPU-hours |
| Colocation space and power | $0.44 to $0.97 | 14 kW at $180 to $400 per kW per month |
| Own and colocate, total | $2.51 to $3.62 | Before networking, storage, financing and staff |
| Rent, lowest 36-month rate | $4.70 | Published reserved rate, checked October 17, 2026 |
Owning looks cheaper on paper, and for a team that will keep the GPUs busy for three years it often is. The gap closes once you add what the table leaves out:
- The network between servers. A multi-node cluster needs a high-speed fabric, which is a large purchase of its own.
- Financing. The table spreads the price evenly with no interest. Paying cash up front, or financing it, costs more.
- Electricity. Ask whether a colocation rate includes it. Contracts differ.
- People. Someone has to install, run and repair the cluster. A rental includes that.
- Resale value. In the other direction, the servers are still worth something after 36 months, which the table also ignores.
Where a B300 cluster can physically go
An 8-GPU B300 server draws about 14 kW. Two in one rack is about 28 kW, several times the 3 to 5 kW a standard cabinet is built for. That rules out most ordinary colocation space and points to high-density halls with engineered cooling. Air-cooled B300 servers exist, but many are liquid-cooled, so confirm the cooling your hardware needs before you shortlist a site.
The GB300 NVL72 is a different product: a liquid-cooled rack of 72 GPUs that work as one unit, drawing up to about 140 kW. Few colocation sites can host it today. See GPU and high-density colocation for what to confirm about power, cooling and delivery dates.
Rent or own: how to decide
- Rent when the work is short, uncertain or bursty, or you need GPUs within weeks. Idle hours on owned hardware are the costliest hours you will pay for.
- Reserve when you know you will keep a cluster busy for a year or more but do not want to run hardware.
- Own and colocate when you will keep the GPUs busy for three years or more, have people to run them, and can get high-density space on your start date.
A published rate is not a promise that capacity is free on your start date. The same is true of colocation space. We check both against your GPU count, term and date before anything reaches a shortlist.
Price a 12, 24, 36 or 60-month B300 term.
Tell us how many GPUs, which term and from when. We come back with reserved rental options and colocation space that can take the power, side by side.
Sources
Rental rates, checked October 17, 2026, each from the provider's own pricing page: Massed Compute (vm.massedcompute.com/pricing), Hyperstack (hyperstack.cloud/gpu-pricing), Civo (civo.com/pricing), Latitude.sh (latitude.sh/pricing), Runpod (runpod.io/pricing), Nebius (nebius.com/prices), Together AI (together.ai/pricing), Verda (verda.com/pricing), Oracle (oracle.com/cloud/price-list), AWS Capacity Blocks (aws.amazon.com/ec2/capacityblocks/pricing), TheAI Cloud (cloud.theai.com), DigitalOcean (digitalocean.com/pricing/gpu-droplets), CoreWeave (coreweave.com/pricing), Lambda (lambda.ai). Server prices: American Compute B300 appraisals (amcompute.com/appraisals/b300), GPU Smith HGX B300 datasheet (gpusmith.com) and a Servermall DGX B300 listing (servermall.com). Power and cluster design: NVIDIA DGX B300 (nvidia.com/en-us/data-center/dgx-b300) and the DGX B300 SuperPOD reference architecture (docs.nvidia.com). Colocation rates: our own quotes, updated October 7, 2026, listed by region on colocation pricing.
Prices change often and are shown as published. A published price is not a promise that capacity is available.