NVIDIA B200 price: $5.50/hr
NVIDIA's Blackwell GPU: 180GB of HBM3e per GPU, FP4 Tensor Cores at 9 PFLOPS dense, and eight GPUs per HGX node. On OpenRelay it runs your container as a pod.
What the rate buys
The B200 runs as a pod: your container image on 1 to 8 whole GPUs of a single node, with SSH access and an HTTPS endpoint. Every pod gets a persistent volume at /workspace, 10 to 1,024 GB, that adds nothing to the hourly rate and survives a stop or restart. A stopped pod is not billed. The details are in the pods documentation.
B200 specifications
| Specification | B200 |
|---|---|
| Architecture | NVIDIA Blackwell |
| Memory | 180GB HBM3e |
| FP4 Tensor Core | 9 PFLOPS dense (18 sparse) |
| FP8 and FP6 Tensor Core | 4.5 PFLOPS dense |
| FP16 and BF16 Tensor Core | 2.25 PFLOPS dense |
| FP64 | 37 TFLOPS |
| NVLink | 5th generation, 1.8 TB/s per GPU |
| Compute capability | 10.0 (CUDA 12.8 or later) |
Source: NVIDIA HGX B200 specifications, per GPU (board totals divided by 8), read 2026-10-01.
What the B200 is for
Models that outgrow an H100
180GB is 2.25 times an H100's 80GB, so a 70B model in FP16, about 140GB of weights, fits on one GPU instead of two.
FP4 inference
FP4 Tensor Cores run at 9 PFLOPS dense, twice the FP8 rate. Hopper has no FP4 path at all.
Training on one node
2.25 PFLOPS of dense BF16 per GPU is a little over twice an H100's 0.99, and a pod holds up to eight B200s on one node.
Frontier-size open models
A 405B model in FP8 is about 405GB, three B200s. A 671B mixture-of-experts model at 4 bits is under 400GB, also three.
FP64 alongside AI
The B200 keeps 37 TFLOPS of FP64, which the B300 gives up, so simulation code that needs double precision belongs here.
Your own serving stack
A pod runs the container you bring, so vLLM, SGLang, TensorRT-LLM or your own server run as you configure them.
Frequently asked questions
How much does it cost to rent an NVIDIA B200?
$5.50 per GPU-hour, billed while the pod is running; a stopped pod is not billed. A full node of eight is $44.00 an hour, about $31,680 a month around the clock. The /workspace volume is included in the rate.
How many B200s does my model need?
Count the weights first. A 70B model is about 140GB in FP16: one B200. A 405B model is about 405GB in FP8: three, with 135GB left over. A 671B model such as DeepSeek-V3 is about 671GB in FP8, four GPUs, or under 400GB in FP4, three. The KV cache comes on top and grows with context length and batch size, so a long-context or high-traffic deployment wants a GPU or two more than the weights alone.
What does it cost to serve a 405B model on B200s?
Three B200s hold its 405GB of FP8 weights: $16.50 an hour, about $11,880 a month around the clock. A fourth GPU for the KV cache brings it to $22.00 an hour. A stopped pod is not billed, so a batch job pays only for the hours it runs.
One B200 or two H100s for a 70B model?
Both hold its 140GB of FP16 weights: one B200 at $5.50/hr or two H100s at $5.20/hr. The B200 has 2.25 PFLOPS of dense BF16 against the pair's 1.98, 20GB more memory for the KV cache, and no traffic between cards, since the model is not split across two GPUs.
Why 180GB and not the 192GB other sources quote?
192GB is the figure for the B200 die. Every shipping HGX B200 board carries 180GB per GPU: NVIDIA lists 1.4TB for the eight-GPU board. 180GB is what your container sees.
Should I pick the B200 or the H100?
The B200 has 180GB against 80GB, 4.5 PFLOPS of dense FP8 against the H100's 1.98, and FP4, which the H100 does not have. The H100 costs less per hour and is the better buy when the model fits in 80GB.
When is the B200 the better choice than the B300?
When the model fits in 180GB, or when the code needs FP64. The B200 runs at $5.50/hr against the B300's $7.49/hr, keeps 37 TFLOPS of double precision where the B300 has 1.25, and matches it on FP8 and BF16. The B300's extra memory and FP4 throughput pay off for the largest models and the longest contexts.
What is FP4, and should I serve my model in it?
FP4 stores each weight in 4 bits plus a scale per block of 16, about 4.5 bits a parameter or 56% of the memory of FP8, and the B200's Tensor Cores run it at 9 PFLOPS dense, twice their FP8 rate. The model has to be quantized to FP4 first, and how much quality it keeps depends on the model and the method, so compare it against FP8 on your own evaluation before you switch.
Can I train across more than one B200 node?
A pod runs on a single node: up to eight B200s with 1.4TB of memory between them, enough for fine-tuning and for training runs whose weights, gradients and optimizer state fit in it. If your run depends on NVLink between the GPUs, or needs several nodes, book a call and we will confirm the host and size it with you before you start.
Is this a GB200?
No. These are HGX B200 boards: eight B200 GPUs on an x86 host, so a container built for x86 runs unchanged. The GB200 pairs two Blackwell GPUs with NVIDIA's Arm-based Grace CPU and ships in NVL72 racks of 72 GPUs, which OpenRelay does not offer.
Will my CUDA container run on the B200?
Yes, if it is built against CUDA 12.8 or later. That is the release that added the B200's architecture (sm_100); an image built for an older CUDA cannot compile kernels for it.
How do I get B200 capacity?
Book a call. B200 nodes are allocated with our team: we size the GPU count and the term with you, and you run pods on them like on any other GPU we sell.
B200 capacity, sized with you
We allocate B200 nodes with you: how many GPUs, for how long, and on what terms. Book a call and we will size it.