Rent NVIDIA B200 servers: a single GPU or an HGX B200

If you want to rent NVIDIA B200 GPUs for a training run or a production endpoint, we have them in stock in two shapes: a single B200 in a dedicated virtual machine at $4,590 per month, and a bare-metal NVIDIA HGX B200 server with eight GPUs at $34,900 per month. Each GPU is whole and assigned only to you, with no slicing or time-sharing, and the price is flat for the month.

What the B200 changes compared with Hopper

The B200 is the Blackwell data-center GPU that followed NVIDIA's Hopper H100 and H200. Against the H200, NVIDIA's published figures show four changes:

• Memory: 180 GB of HBM3e per GPU instead of 141 GB, at 8 TB/s instead of 4.8 TB/s, about 1.7 times the bandwidth.
• Compute: 4.5 PFLOPS of dense FP8 and 2.25 PFLOPS of dense BF16, about 2.3 times Hopper's 1,979 and 989 TFLOPS.
• FP4: 9 PFLOPS dense, a format Hopper does not support in hardware.
• NVLink 5: 1.8 TB/s per GPU, twice the 900 GB/s of NVLink 4.

One detail matters for sizing: the B200 package physically carries 192 GB of HBM3e, but 180 GB per GPU is the capacity listed for HGX B200 systems, so plan with 180 GB. For a side-by-side view, see our H200 vs B200 comparison.

B200 workloads and memory math

A quick rule: weights need roughly 2 bytes per parameter in BF16, 1 in FP8 and 0.5 in FP4, plus KV cache and 10–30% headroom. Applied to the B200:

Serving models

• 70B in BF16: about 140 GB, leaving roughly 40 GB for KV cache and overhead on one GPU. That works for moderate context lengths and concurrency; on a 141 GB H200 it does not.
• 70B in FP8: about 70 GB, so roughly 110 GB stays free for KV cache, which means more concurrent requests or longer contexts per GPU.
• 405B-class model in FP8: about 405 GB, more than two B200s can hold, but one HGX B200 server (1.44 TB) runs it with roughly 1 TB left for KV cache. In BF16, at about 810 GB, it still fits.

Training and fine-tuning

Full fine-tuning with Adam takes roughly 16–20 bytes per parameter before activations. A 32B model therefore needs about 512–640 GB, comfortable on one HGX B200 with the state sharded over eight GPUs. A 70B model needs about 1.1–1.4 TB, which is at the edge of the server's 1.44 TB: it can work with CPU offload into the 2 TB of DDR5, but an HGX B300 (2.3 TB) or two servers give real headroom. LoRA on a 70B model in BF16 fits on a single B200 with about 40 GB for adapters and activations, enough if you keep sequence lengths and batch sizes moderate.

Our walkthrough on fine-tuning a 70B model compares these options, and how much VRAM LLMs need explains the KV cache side of the math.

Single B200 or HGX B200 server?

A single B200 comes with 26 vCPUs, 256 GB of RAM, 2.75 TB of NVMe and a 25 Gbps port. It suits one model per GPU: a 70B endpoint in FP8, a LoRA run, an evaluation pipeline. To scale, you add more single plans.

The HGX B200 server is the choice when one model or job spans GPUs. Its eight B200s share one NVLink domain at 1.8 TB/s per GPU, so tensor parallelism across all eight is practical, and gradient synchronization in sharded training never crosses a network. Around them sit two Xeon 8570 processors with 112 cores in total, 2 TB of DDR5, eight 3.84 TB NVMe drives and eight 400G ConnectX-7 adapters.

The price follows the same logic. Eight single B200 plans would total $36,720 per month, against $34,900 for the HGX B200, so at eight GPUs the server costs less and adds NVLink. Up to seven GPUs, separate single plans cost less, as long as your workload does not need them linked.

Running PyTorch and vLLM on the B200

Servers come with root access over SSH and Ubuntu 22.04 or 24.04, with NVIDIA drivers, CUDA, Docker and the NVIDIA Container Toolkit already installed; Debian or Rocky Linux are available on request. Blackwell needs recent framework builds, so use current PyTorch, vLLM, SGLang or TensorRT-LLM releases, ideally from container images you pin by version. A typical check of the interconnect and an eight-GPU training launch:

nvidia-smi topo -m                             # NVLink between all eight GPUs
NCCL_DEBUG=INFO torchrun --nproc_per_node=8 train.py

With NCCL_DEBUG=INFO, NCCL logs the transports it sets up, a quick way to confirm that traffic stays on NVLink. Our docs on PyTorch and vLLM and multi-GPU NCCL setups go further.

What it costs to rent NVIDIA B200 servers

• HGX B200, 8 GPUs $34,900 per month, about $5.98 per GPU-hour
• Single B200 $4,590 per month, about $6.29 per GPU-hour

The hourly equivalent divides the monthly price by the GPU count and by 730 hours. For an 8-GPU server, the break-even point against hourly billing is $34,900 divided by eight times the per-GPU-hour rate you would pay elsewhere. Put as utilization: if that rate is 1.25 times our $5.98, monthly wins once you use the server more than 80% of the month (about 584 hours); at twice our rate, anything above 50% (365 hours) does. If you only need B200s for a few days, hourly billing is the better deal.

Predictability is the other argument. FullGPU's October 2026 survey of public price lists found the median on-demand B200 price per GPU-hour up roughly 20% over the past year. On a monthly plan, your price is fixed for each paid month, and any change for future months comes with at least 30 days' notice. Every plan includes 20 TB (single GPU) or 50 TB (8-GPU) of outbound traffic per month with no egress fees, public IPv4 and an IPv6 /64, DDoS mitigation, 24/7 support and the 99.9% availability SLA, with no setup fee. More context on supply is in our 2026 GPU shortage analysis.

From request to root access

Send a request with the configuration and quantity you need. We reply with a written confirmation of the configuration, monthly price, location and delivery date; nothing is binding until you pay the first month. You then pay by card or in crypto (see accepted coins), with the same USD price and invoice either way, and receive access details by email once the server is provisioned. If we miss the confirmed delivery date, you can cancel for a full refund. For volume pricing or 6–12 month terms, use the contact form.