Rent NVIDIA H200 GPUs

Rent an H200 on Akash

As of October 20, 2026, on-demand H200s on Akash start at $4.45 per GPU-hour. Reserved clusters are custom-quoted by GPU count, term and workload.

Why the H200 stands out

• Memory: 141 GB of HBM3e per GPU — NVIDIA describes it as "nearly double the capacity" of the H100.
• Bandwidth: 4.8 TB/s, 1.4× the H100.
• Faster LLM inference: NVIDIA reports up to 1.9× higher Llama 2 70B inference throughput and 1.6× higher GPT-3 175B throughput than H100 in its published benchmarks. Results depend on model, batch size and configuration.
• Same power profile: NVIDIA notes the gains come "within the same power profile as the H100."
• Mature software: the Hopper architecture, with FP8 and the Transformer Engine, is supported across major frameworks.

Is the NVIDIA H200 the right GPU for you?

Choose the H200 when you need more memory than an H100 but want to stay on Hopper.

• Choose the H200

  if your model or KV cache doesn't fit in 80 GB, you serve 70B-class models with long context, or your stack is tuned for Hopper.

• Choose the H100

  if your workload fits in 80 GB per GPU. See H100 pricing →

• Choose the B200

  if you want Blackwell, native FP4 and 180 GB per GPU. See B200 pricing →

What is the NVIDIA H200 used for?

The NVIDIA H200 is used for LLM inference, fine-tuning and HPC on the Hopper architecture. Its 141 GB of memory and 4.8 TB/s bandwidth make it suited to:

• Large-model LLM inference

  Up to 1.9× higher Llama 2 70B throughput and 1.6× higher GPT-3 175B throughput than H100.

• Long-context serving

  One 8× H200 node holds a 405B-parameter model in FP8 and serves 128K-token inputs.

• Speculative decoding

  Up to 3.6× higher throughput with draft-target speculative decoding in TensorRT-LLM.

• LLM fine-tuning

  Fine-tune 70B models at over 15,000 tokens/s on an 8× H200 node with NeMo.

• RAG, speech and vision AI

  Deploy retrieval-augmented generation, computer vision and speech pipelines with NIM microservices.

• HPC and simulation

  Up to 110× faster time to results than CPUs on scientific workloads, in the same power profile as H100.

Deploy an H200 in three steps

1. 1

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

2. 2

   Choose H200. Turn on GPU, select H200 and check current provider capacity.

3. 3

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

NVIDIA H200 specifications

NVIDIA lists H200 specifications as preliminary. Dense tensor figures are NVIDIA's sparse figures divided by 2.

NVIDIA H200 FAQ