NVIDIA B300 GPUs are available from a growing list of GPU clouds and from AWS, but mostly not the way a pricing page suggests. A handful of providers list B300 on demand, at roughly $7 to $18 per GPU per hour in early November 2026. Most B300 capacity, and almost all rack-scale GB300 NVL72 capacity, is sold as reserved clusters and committed contracts, often with minimum terms, minimum cluster sizes, and a lead time between signing and delivery.
Treat any list of B300 providers, including the one in this post, as a starting point that changes week by week. New regions come online, prices move as supply grows, and listed capacity sells out. This post explains what B300 is, how its capacity is sold, and what to check in an offer, so the advice holds even when the snapshot does not. The current provider list lives on the GPU Kubernetes page, which we keep up to date.
What the B300 is
The B300 is NVIDIA’s Blackwell Ultra GPU, the successor to the B200 in the Blackwell generation. NVIDIA ships it as the 8-GPU HGX B300 board and as part of the GB300 NVL72 rack. Compared with HGX B200, the HGX B300 board brings more memory, more low-precision compute, and twice the network bandwidth:
| HGX board, 8 GPUs | B300 | B200 |
|---|---|---|
| Total GPU memory | 2.1 TB | 1.4 TB |
| FP4 Tensor Core, dense | 108 PFLOPS | 72 PFLOPS |
| FP8 Tensor Core, dense | 36 PFLOPS | 36 PFLOPS |
| NVLink bandwidth per GPU | 1.8 TB/s | 1.8 TB/s |
| Networking bandwidth | 1.6 TB/s | 0.8 TB/s |
| FP64 | 10 TFLOPS | 296 TFLOPS |
Per GPU, providers list B300 memory at about 270 to 288 GB, against 180 GB for the B200. That extra memory is the main reason to pick a B300: larger models, longer contexts, and bigger batches fit on fewer GPUs, which matters most for inference of large reasoning models and for training with FP4 and FP8.
The table also shows the trade-off. The B300 gives up most of its FP64 and INT8 throughput, so scientific computing that depends on double precision runs better on the B200 or H200.
HGX B300 or GB300 NVL72
The two B300 systems are built for different scales. An HGX B300 node has 8 GPUs connected by NVLink inside the node, and multi-node jobs connect nodes over InfiniBand or Ethernet, at 800 Gb per second per GPU in NVIDIA’s reference design. It fits into standard data center racks and is what most providers rent as 8-GPU instances or bare-metal servers.
The GB300 NVL72 is a liquid-cooled rack with 72 Blackwell Ultra GPUs and 36 Grace CPUs in a single NVLink domain of 130 TB per second, with 20 TB of HBM3e in total. NVIDIA positions it for test-time scaling inference and AI reasoning. It is sold almost entirely through reservations, by the rack or in fractions of one, rather than on demand.
How B300 capacity is sold
B300 capacity comes in the same three forms as other GPUs, but the balance is tilted toward commitment:
- On-demand listings exist at a few providers, usually for single 8-GPU nodes and in a few regions. A listed price is not a guarantee of capacity, because popular regions sell out.
- Reserved clusters of several nodes with InfiniBand networking are the most common way to get B300 for training. Terms vary by provider and often include minimum durations, minimum cluster sizes, and payment up front.
- Committed contracts of a year or more carry the lowest rates and are how most capacity is allocated. Lead time between signing and delivery is part of the deal.
For the buying models in general, including AWS Capacity Blocks and Google Cloud flex-start, see GPU procurement in a shortage.
Snapshot: B300 providers in early November 2026
The following were public on-demand list prices per GPU per hour on providers’ own pricing pages in early November 2026. They exclude VAT, vary by region, and will be out of date by the time you read this. Use them to see the shape of the market, not to plan a budget.
| Provider | Listed price | Where |
|---|---|---|
| FPT AI Factory | $6.99 | Southeast Asia |
| Hyperstack | $7.40 | Canada |
| Civo | $7.49 | Global |
| Scaleway | €7.50 | Paris |
| Crusoe | $8.96 | Iceland |
| Nebius | $9.50 | Paris, UK |
| Together AI | $9.99 | US |
| Verda | $11.57 | Helsinki |
| AWS p6-b300 | $17.80 | US (London through Capacity Blocks) |
| CoreWeave, Gcore, IREN, Lambda, Voltage Park | On request, reserved, or contract | US, Canada, Europe, and more |
Several providers on the list run B300 in Europe, including Scaleway in Paris, Nebius in Paris and the UK, Verda in Helsinki, Crusoe in Iceland, and AWS in London through Capacity Blocks, which matters when training data has to stay in Europe.
What to check in a B300 offer
- Term and minimums. On demand, reserved for weeks, or committed for a year; the minimum number of nodes; prepayment and cancellation terms.
- Delivery date. When the nodes are handed over, and what happens if delivery slips.
- Price unit. Per GPU or per 8-GPU node, and whether CPU, memory, local NVMe, and networking are included.
- Interconnect. InfiniBand or RoCE between nodes and the bandwidth per GPU, which decides whether multi-node training scales.
- Location and data residency. The region and the legal entity that operates it.
- Hardware failure. How fast a failed GPU or node is replaced, and whether the term is extended for downtime.
- Handover. Virtual machines in a cloud account, or bare-metal servers with SSH access, and how they join the platform your workloads run on.
Running B300 next to the rest of your GPUs
B300 nodes rarely run alone. Training on a B300 reservation at one provider, inference on L40S or B200 instances in another region, and servers in your own data center is a common setup. Cloudfleet joins them into one Kubernetes cluster: reserved nodes join through the Cloudfleet API, Terraform, or a single CLI command as soon as the provider hands them over, Cloudfleet installs the NVIDIA driver and container toolkit, and workloads target B300 nodes with the cfke.io/accelerator-name label. See training and inference in one Kubernetes cluster for that setup.
Because B300 availability changes faster than any article, the most reliable way to get it is to ask the providers directly. The Cloudfleet team works closely with most infrastructure providers and can collect current B300 offers for your GPU count, term, interconnect, and region, including reserved capacity that never appears on a pricing page. You contract and pay the provider directly, with no markup. For the difference between that and a GPU marketplace, see GPU marketplace vs managed GPU Kubernetes.
Frequently asked questions
How much does it cost to rent an NVIDIA B300?
In early November 2026, public on-demand list prices ranged from about $7 to $18 per GPU per hour, with GPU clouds below AWS. Reserved and committed capacity costs less per hour but is quoted individually and changes often. Check the current list on the GPU Kubernetes page before you plan.
Where can I rent B300 GPUs?
GPU clouds such as Crusoe, Nebius, Scaleway, Verda, CoreWeave, Lambda, and Together AI, and AWS with p6-b300 instances. Availability differs by region and changes quickly, and most capacity is sold as reservations rather than on demand.
Is B300 available on demand?
At some providers, for single nodes and in a few regions. Most B300 capacity, and nearly all GB300 NVL72 capacity, is sold through reserved clusters and committed contracts, so plan for a reservation if you need more than a few nodes.
What is the difference between B200 and B300?
The B300, also called Blackwell Ultra, has more memory (2.1 TB against 1.4 TB on an 8-GPU board), 50% more dense FP4 compute, and twice the network bandwidth. It has far lower FP64 throughput, so double-precision workloads run better on the B200.
How much memory does the B300 have?
About 270 to 288 GB per GPU, depending on how the provider reports it, and 2.1 TB on an 8-GPU HGX B300 board according to NVIDIA. The GB300 NVL72 rack has 20 TB of HBM3e across 72 GPUs.
