Rent NVIDIA B300 and HGX B300 servers from $5,190 per month

To rent NVIDIA B300 capacity without a one-year contract, you have two options with us: a single B300 in a dedicated virtual machine at $5,190 per month, or a bare-metal NVIDIA HGX B300 server with eight GPUs at $39,900 per month. Either way, every GPU is a whole physical GPU assigned only to you, billed monthly in advance with a one-month minimum. B300 stock is limited; the configurations card on this page shows what is available now.

What the B300 is and what changed from the B200

The B300 is NVIDIA's Blackwell Ultra GPU, the update of the Blackwell B200. The main change is memory: 288 GB of HBM3e per GPU, against 180 GB usable on the B200, so 60% more capacity at a similar bandwidth of about 8 TB/s. An HGX B300 board carries eight of them, roughly 2.3 TB of HBM3e in one server.

The second change is FP4. NVIDIA lists 13.5 PFLOPS of dense FP4 per B300, against 9 PFLOPS for the B200, while its published dense FP8 (4.5 PFLOPS) and BF16 (2.25 PFLOPS) figures are the same for both GPUs. The GPUs still connect over fifth-generation NVLink at 1.8 TB/s per GPU. In short, the B300 pays off when the extra memory saves you GPUs, or when you serve or train in FP4. If neither applies, the B200 offers the same vendor-listed FP8 and BF16 figures for less per month. Our B200 vs B300 comparison covers the details, and our guide to FP8 and FP4 explains when lower precision is safe for your model.

B300 workloads and memory math

Memory is usually the deciding factor, so start with the arithmetic. Weights take roughly parameters × bytes per parameter (2 for BF16, 1 for FP8, 0.5 for FP4). Add the KV cache, which grows with context length and the number of concurrent requests, and keep 10–30% headroom for activations and runtime overhead.

Inference

• 70B model in BF16: about 140 GB of weights. One B300 still has roughly 100 GB or more for KV cache after headroom, so you can serve in BF16 with long contexts. On a single 141 GB H200, the same model does not fit with useful headroom.
• 70B model in FP8: about 70 GB, which leaves most of the GPU for KV cache and high concurrency.
• 671B mixture-of-experts model such as DeepSeek-V3, in FP8: about 671 GB of weights. That needs at least three B300s for the weights alone, and it fits one HGX B300 server with well over 1 TB left for KV cache. Eight 80 GB GPUs could not hold the weights at all. Our DeepSeek hardware requirements guide covers other precisions.

Fine-tuning and training

Full fine-tuning with the Adam optimizer needs roughly 16–20 bytes per parameter before activations. For a 70B model that is about 1.1–1.4 TB, which fits in the 2.3 TB of one HGX B300 server when the state is sharded across all eight GPUs with FSDP or DeepSpeed ZeRO-3, leaving roughly 0.9–1.2 TB for activations. On a single B300, LoRA on a 70B model in BF16 works: about 140 GB of frozen weights plus small adapters and activations stays well under 288 GB.

For pre-training, what matters is keeping GPUs busy for weeks. A monthly server keeps your data, checkpoints and environment on local NVMe (eight 3.84 TB drives, about 30 TB) for the whole run, with no preemption. To run these numbers for another model, use our VRAM calculator.

One B300 or an 8-GPU HGX B300 server?

Pick the single B300 when one model fits in 288 GB and you scale by adding independent copies: inference replicas, LoRA runs, evaluation jobs. The plan includes 28 vCPUs, 384 GB of RAM, 3.5 TB of NVMe and a 25 Gbps port.

Pick the HGX B300 server when one job must span several GPUs. The eight GPUs share one NVLink domain at 1.8 TB/s per GPU, which is what tensor parallelism and sharded training need; split the same model across machines and every layer waits on much slower network links. The server adds two Xeon 6776P processors with 128 cores in total, 4 TB of DDR5 for data loading and CPU offload, and eight 800G ConnectX-8 adapters for multi-node work.

Cost points the same way once you need eight GPUs: eight single B300 plans would total $41,520 per month, against $39,900 for one HGX B300 server. Up to seven single plans cost less than the server, but they are separate machines with no NVLink between them. If one node is not enough, the rack-scale GB300 NVL72 puts 72 Blackwell Ultra GPUs in a single NVLink domain.

Software on a B300 server

You get root access over SSH with your key, on Ubuntu 22.04 or 24.04 with NVIDIA drivers, CUDA, Docker and the NVIDIA Container Toolkit preinstalled. Debian or Rocky Linux are available on request. Common choices are PyTorch for training and vLLM, SGLang or TensorRT-LLM for serving, usually run from containers, so an environment you already use elsewhere moves over as an image. A first session on an HGX B300 might look like this:

nvidia-smi              # eight GPUs, driver and CUDA versions
nvidia-smi topo -m      # how the GPUs are connected (NVLink)
docker run --gpus all --ipc=host -p 127.0.0.1:8000:8000 \
  vllm/vllm-openai:latest \
  --model <your-model> --tensor-parallel-size 8

Binding the inference endpoint to localhost and reaching it through an SSH tunnel keeps it off the public internet until you add your own authentication. Before you size a deployment around FP4, check that your framework version and model checkpoint support it on Blackwell Ultra. For training across all eight GPUs, see multi-GPU training with NCCL.

NVIDIA B300 price per month and per GPU-hour

• HGX B300, 8 GPUs $39,900 per month, about $6.83 per GPU-hour
• Single B300 $5,190 per month, about $7.11 per GPU-hour

The per-GPU-hour figure is the monthly price divided by the number of GPUs and by 730 hours, so it assumes the GPU works around the clock. To compare with hourly billing, divide our monthly price by the hourly rate you would pay elsewhere: the result is the number of hours per month above which the flat price costs less. For example, if an hourly B300 costs you 1.5 times our equivalent rate, the single plan comes out ahead after about 487 hours, two-thirds of the month. For a two-day experiment, hourly billing is the better deal; for a server that trains or serves for most of the month, monthly is. Our monthly vs hourly pricing comparison works through more cases.

The price covers the server, 50 TB of outbound traffic per month on the 8-GPU server (20 TB on the single GPU) with no egress fees within that allowance, a public IPv4 address and an IPv6 /64, DDoS mitigation, 24/7 support from engineers and a 99.9% monthly availability SLA with service credits. There is no setup fee. Prices are in US dollars and exclude VAT and other taxes where applicable. Once you have an active server, you can ask by ticket for volume pricing on 6–12 month terms or several servers.

How ordering works

1. On the deployment page, choose the configuration you want, set up the OS image, optional preinstalled software and your login (root password or SSH key), and create a free account or log in, all in the same form.
2. You pay the first month from your prepaid balance, which you top up in crypto (BTC, ETH, XMR, USDT on TRON or SOL) from $100 to $10,000 per top-up, with +5% bonus credit for every full $100, up to +50% from $1,000. If your balance is short, you top up in the same form: the order is held for 3 hours and paid automatically as soon as the top-up is credited. Card payments are coming soon.
3. Root SSH access is delivered within 10 minutes of payment, with the IP and SSH details in your dashboard.

Renewals are paid automatically from your balance each month, with an invoice for every payment. Most orders need only an account and a funded balance; large orders, such as several 8-GPU servers, may need verification first. For sizing before you order, use the VRAM calculator and the guides linked on this page.