Rent NVIDIA B200 GPUs

Rent a B200 on Akash

As of October 21, 2026, on-demand B200s on Akash start at $5.00 per GPU-hour. Reserved clusters are custom-quoted by GPU count, term and workload.

What Blackwell brings

• Native FP4: 9 PFLOPS of dense FP4 per GPU, a format Hopper GPUs don't support.
• More memory: 180 GB of HBM3e per GPU, up from 141 GB on H200.
• More bandwidth: 8 TB/s per GPU, 1.67× the H200.
• Faster NVLink: 5th-gen NVLink at 1.8 TB/s per GPU, double the H200.
• Strong FP64: about 37 TFLOPS per GPU, so B200 still handles scientific computing.
• Confidential computing: NVIDIA calls Blackwell "the first TEE-I/O capable GPU in the industry"; full protection requires a compatible TEE-I/O host and configuration.

NVIDIA projects DGX B200 system performance at 3× the training and 15× the inference performance of DGX H100. These are system-level projections; results vary by workload and configuration.

Is the NVIDIA B200 the right GPU for you?

Choose the B200 when you need Blackwell compute and FP4 but your workload fits in 180 GB per GPU.

• Choose the B200

  if you train or fine-tune large models, serve FP4 or FP8 models at scale, or need strong FP64 alongside AI.

• Choose the B300

  if you need up to 288 GB per GPU for trillion-parameter models or very long context. See B300 pricing →

• Choose the H200

  if your workload fits in 141 GB and your stack is tuned for Hopper. See H200 pricing →

What is the NVIDIA B200 used for?

The NVIDIA B200 is used to train, fine-tune and serve large AI models. Its native FP4, 180 GB of memory and 8 TB/s bandwidth make it suited to:

• Reasoning-model inference in FP4

  Serve reasoning models in NVFP4 with accuracy within about 1% of FP8 and far more tokens per GPU. An 8× B200 node delivers over 30,000 tokens/s on DeepSeek-R1.

• Mixture-of-experts serving

  Run large MoE models with multi-token prediction and NVFP4 for big throughput gains over FP8.

• Agentic AI with long inputs

  Disaggregated serving with NVIDIA Dynamo gives up to 4× faster per-user token speed for code generation, summarization and agent workflows on 8× B200.

• Low-cost token generation

  Around $0.02 per million tokens on gpt-oss-120b, roughly 4.5× cheaper than Hopper.

• Large-model training and post-training

  Around 3× the training performance of an H100 system, with NVFP4 now speeding up training as well as inference.

• Confidential AI inference

  Run private inference with confidential computing on and keep 96–98% of normal throughput.

Deploy a B200 in three steps

1. 1

   Add your app. Pick a template or paste a Docker image.

2. 2

   Choose B200. Turn on GPU, select B200 and check current provider capacity.

3. 3

   Deploy. Request quotes, accept one and start your workload.

NVIDIA B200 specifications

Per-GPU figures are NVIDIA's HGX B200 system figures divided by 8.

NVIDIA B200 FAQ