Rent NVIDIA H100 GPUs

Rent an H100 on Akash

As of October 20, 2026, on-demand H100s on Akash start at $2.73 per GPU-hour. Reserved clusters are custom-quoted by GPU count, term and workload.

Why the H100 is still the standard

• FP8 Transformer Engine: 1,979 TFLOPS of dense FP8 per GPU for faster training and inference.
• Big leap over A100: NVIDIA projects up to 4× faster GPT-3 175B training and up to 30× faster inference on a 530B-parameter model versus A100.
• HPC: 67 TFLOPS of FP64 Tensor per GPU and DPX instructions, with NVIDIA citing up to 7× higher HPC performance than A100.
• Confidential computing: NVIDIA calls H100 "the world's first accelerator" with confidential computing.
• Broad support: supported across major AI frameworks and serving engines.

Is the NVIDIA H100 the right GPU for you?

Choose the H100 when your workload fits in 80 GB per GPU and you want proven Hopper performance.

• Choose the H100

  if you fine-tune or serve models that fit in 80 GB, need FP8, or want the most mature software stack.

• Choose the H200

  if you need more memory: 141 GB per GPU with the same compute. See H200 pricing →

• Choose the A100

  if you don't need FP8 and want an established Ampere GPU. See A100 pricing →

What is the NVIDIA H100 used for?

The NVIDIA H100 is used for AI training, fine-tuning, inference and HPC. Its FP8 Transformer Engine and 3.35 TB/s bandwidth make it suited to:

• Large language model training

  Up to 4× faster GPT-3 175B training than A100; Llama 3.1 405B was trained on more than 16,000 H100s.

• LLM inference with FP8

  1,979 TFLOPS of dense FP8 through the Transformer Engine; up to 5× A100 on Llama 2 70B.

• Reasoning-model serving

  Disaggregated serving with NVIDIA Dynamo more than doubles Llama 70B throughput on Hopper.

• RLHF and post-training

  Up to 5.6× faster RLHF with TensorRT-LLM; the standard GPU for Nemotron reasoning-model fine-tuning recipes.

• Enterprise AI with RAG, vision and speech

  NIM microservices plus MIG partitioning into up to 7 instances for multi-tenant deployments.

• HPC and genomics

  67 TFLOPS of FP64 Tensor and DPX instructions deliver up to 7× A100 on dynamic-programming workloads such as DNA sequence alignment.

Deploy an H100 in three steps

1. 1

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

2. 2

   Choose H100. Turn on GPU, select H100 and check current provider capacity.

3. 3

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

NVIDIA H100 specifications

H100 NVL (PCIe) has 94 GB per GPU and 3.9 TB/s; confirm which version providers offer.

NVIDIA H100 FAQ