Rent NVIDIA A100 GPUs

Rent an A100 on Akash

As of October 20, 2026, on-demand A100s on Akash start at $1.72 per GPU-hour. Reserved clusters are custom-quoted by GPU count, term and workload.

Why teams still run the A100

• 80 GB of HBM2e at 2,039 GB/s — room for mid-size models and large batches.
• Multi-Instance GPU: up to 7 isolated instances of 10 GB each on one GPU.
• TF32 and BF16 Tensor Cores: 156 TFLOPS dense TF32 and 312 TFLOPS dense BF16 for training.
• FP64 Tensor Cores: 19.5 TFLOPS for HPC and simulation.
• Generational leap: NVIDIA rates A100 Tensor Core performance at up to 20× the prior Volta generation for AI workloads.

Is the NVIDIA A100 the right GPU for you?

Choose the A100 when your workload fits in 80 GB and doesn't need FP8.

• Choose the A100

  if you run BF16 or TF32 training, batch inference, or MIG-based multi-tenant workloads.

• Choose the H100

  if you need FP8 and the Transformer Engine. NVIDIA projects up to 4× faster GPT-3 training on H100. See H100 pricing →

• Choose the H200

  if you need 141 GB per GPU. See H200 pricing →

What is the NVIDIA A100 used for?

The NVIDIA A100 is used for AI training, inference, data analytics and HPC. Its 80 GB of memory and MIG support make it suited to:

• LLM inference with NIM

  Verified for current open models including gpt-oss, Llama 3.1/3.3 70B and Nemotron 3 through NIM microservices.

• Consolidated multi-model pipelines

  Run speech recognition, text-to-speech and vision-language models together on one A100 with MIG, with 32% higher throughput than time-slicing.

• AI training with TF32 and BF16

  Up to 20× the Tensor Core performance of the Volta generation, and 3× faster training on the largest recommender models.

• Multi-tenant inference

  Up to 7 MIG instances per GPU, delivering 7× BERT-Large inference throughput versus a single instance.

• Big data analytics

  2× faster Spark and RAPIDS analytics on the 80 GB version than on A100 40GB.

• HPC simulation

  19.5 TFLOPS of FP64 Tensor, with up to 2× faster Quantum Espresso on the 80 GB version.

Not ideal for: FP8 or FP4 inference. The A100 has no FP8 or FP4 Tensor Cores; choose H100, H200 or Blackwell.

Deploy an A100 in three steps

1. 1

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

2. 2

   Choose A100. Turn on GPU, select A100 and check current provider capacity.

3. 3

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

NVIDIA A100 specifications

The 40GB A100 uses HBM2 at 1,555 GB/s with up to 7 MIG instances of 5 GB.

NVIDIA A100 FAQ