NVIDIA A100 Pricing: Cheapest On-demand GPU Instances

A100 rental starts at $1.09/GPU-hr on Thunder Compute in this October 2026 comparison of on-demand NVIDIA A100 cloud GPU pricing.

This guide covers on-demand rates for the 80GB and 40GB variants across major cloud and marketplace providers. It also recaps the specs, shows which models fit on a card, and compares the A100 to other GPUs.

NVIDIA A100 80GB Cloud Pricing

The pricing information in this guide is reviewed weekly.

Note

Thunder Compute is about 4.7x cheaper than Google Cloud for a single A100, about 3.1x cheaper than AWS, and about 1.8x cheaper than Crusoe Cloud.

NVIDIA A100 40GB Cloud Pricing

Cost to Run an A100 80GB for 2 Hours

Methodology

• On-demand only: No reserved-instance, commitment, or prepaid discounts.
• Same class of silicon: Listed rates are for NVIDIA A100 GPUs where the provider publishes A100 pricing, separated by memory variant.
• Public price lists: Each figure comes from the vendor's current pricing page on the date above.
• USD in U.S. regions: Rates in other regions can differ by 5-20%.
• We review this data weekly.

A100 Cost For Longer Workloads

Even small differences in hourly pricing add up quickly. A 100-hour A100 80GB workload costs $109.00 on Thunder Compute, $200.00 on Crusoe Cloud, $343.09 on AWS, and $502.77 on Google Cloud. These figures exclude storage, networking, and egress fees.

A100 40GB vs 80GB: Which Should You Rent?

Rent the 80GB variant in almost every case. It often costs the same or less than a 40GB card while doubling memory: Thunder Compute's 80GB A100 at $1.09/hr undercuts 40GB cards on Paperspace ($3.09/hr) and Google Cloud ($3.67/hr).

The extra memory buys larger batch sizes, longer context windows, and headroom to run bigger models without spilling to slower CPU offload. It also makes full fine-tuning and multi-tenant MIG partitioning far more practical on a single card.

The 40GB card is only worth it when priced well below the 80GB option, as on Vast.ai at about $0.70/hr. For short, small-model jobs, at $0.35/hr the RTX A6000 may be the better budget pick: compared to the A100 40GB it has more VRAM (48GB vs 40GB), but lower memory bandwidth (768GB/s vs 1,555GB/s).

A100 Specs at a Glance

The A100 80GB pairs 80GB of HBM2e at up to 2,039GB/s (SXM; the PCIe card is 1,935GB/s) with 6,912 CUDA cores and 432 third-generation Tensor Cores on NVIDIA's Ampere architecture. It supports FP16, BF16, TF32, and INT8, and tops out at BF16 for training, with no native FP8.

A single A100 can be partitioned into up to seven isolated Multi-Instance GPU (MIG) slices for multi-tenant inference. The card ships in 40GB and 80GB memory variants and in PCIe or SXM form factors, with SXM delivering higher bandwidth and a larger power envelope.

Read next

For the full four-variant comparison, precision breakdown, MIG detail, and DGX and HGX system configurations, see our NVIDIA A100 specs guide.

What Models Fit on an A100?

The A100 80GB runs a wide range of open-weight LLMs, image, and video models on a single card, while the 40GB variant covers the smaller ones. The table below lists approximate single-inference VRAM and single-card fit for each variant.

How the A100 Compares to Other GPUs

The A100 is faster and more memory-rich than the consumer RTX 4090 and workstation RTX A6000, but a generation behind the H100. The table below orients all four GPUs on the specs that drive workload fit.

A100 vs H100

The H100 can be 2-4x faster than the A100, but at a higher hourly rate it only lowers total job cost when its speedup reaches roughly 3x or more. For LoRA fine-tuning, sub-30B inference, and development work, the A100 at $1.09/hr is usually cheaper per job.

The A100 also lacks FP8 and the Transformer Engine, which the H100 uses to accelerate high-throughput inference on large models. Where those features matter, the H100 can pull ahead on cost per token despite the premium.

Read next

For benchmarks, a per-job cost model, and a use-case decision guide, see our full H100 vs A100 comparison.

A100 vs RTX A6000

The RTX A6000 carries 48GB of GDDR6, more than an A100 40GB, but its 768GB/s bandwidth is well under half the A100 80GB's 2,039GB/s. For memory-bound LLM training and inference, that gap makes the A100 substantially faster, and the A100 adds MIG partitioning that the A6000 lacks.

The A6000 is the better value for single-workstation development, rendering, and smaller models, where its large, cheaper memory matters more than raw bandwidth. For sustained training or large-model inference, the A100 pulls ahead.

Read next

For the full spec and workload breakdown, see our RTX A6000 vs A100 comparison.

A100 vs RTX 4090

The RTX 4090 delivers strong raw compute and native FP8, but its 24GB of VRAM and lack of NVLink cap the model sizes it can fine-tune. The A100 80GB fits far larger models and full fine-tunes on a single card, and scales across GPUs over NVLink.

For budget QLoRA and small-model fine-tuning, the 4090 is cost-effective where you can access one. For larger models, longer context, or multi-GPU runs, the A100 is the more capable and more available cloud option.

Read next

For benchmarks and memory analysis, see our A100 vs RTX 4090 comparison.

NVIDIA A100 on the Major Clouds

All four major hyperscalers offer the A100, but at 3-5x the rate of specialist clouds. Each summary links to a full instance-level breakdown of that cloud.

NVIDIA A100 on AWS EC2

AWS offers the A100 80GB through the p4de.24xlarge instance and the 40GB A100 through the older p4d.24xlarge.

Egress fees apply on top of the hourly rate after the free tier, so data-heavy workloads cost more than the sticker price suggests.

Read next

See the full lineup in our AWS EC2 GPU instances guide.

NVIDIA A100 on Microsoft Azure

Azure lists the A100 80GB through its NC A100 v4-series, starting at $3.67/GPU-hr for Standard_NC24ads_A100_v4 in the eastus region and scaling to 2- and 4-GPU sizes. The 8-GPU ND A100 v4 series targets distributed training over InfiniBand. Rates range $3.67-$4.78/GPU-hr across US regions.

Read next

See the full lineup in our Azure GPU instances guide.

NVIDIA A100 on Google Cloud

Google Cloud offers the A100 through the A2 machine family, with a2-ultragpu-1g (80GB) at $5.03/GPU-hr and a2-highgpu-1g (40GB) at $3.67/GPU-hr in us-central1. A2 shapes scale from 1 to 16 GPUs. Google Cloud is the most expensive hyperscaler for the 80GB card.

Read next

See the full lineup in our Google Cloud GPU instances guide.

NVIDIA A100 on Oracle Cloud

Oracle offers the A100 40GB through the BM.GPU4.8 bare-metal shape at $3.05/GPU-hr and the A100 80GB through the BM.GPU.A100-v2.8 bare-metal shape at $4.00/GPU-hr. Bare-metal access means no hypervisor overhead, which suits sustained multi-GPU training runs.

Read next

See the full lineup in our Oracle Cloud GPU pricing guide.

Run A100 GPUs on Thunder Compute

Thunder Compute offers the A100 80GB at $1.09/hr on-demand, among the lowest rates tracked, with per-minute billing and no minimum commitment. The VS Code and Cursor extensions connect your editor directly to a running instance, with no SSH configuration required.

See A100 availability and pricing on Thunder Compute →

Last Thoughts on NVIDIA A100 Pricing

The A100 remains one of the most cost-effective GPUs for fine-tuning, inference, and development in 2026, with on-demand rental from $1.09/hr. Specialist clouds price it 3-5x below the hyperscalers, so the provider matters as much as the GPU.

FAQ