GPUBeacon
Language
NVIDIA

Cloud GPU · verified 2026-09-05

NVIDIA A100

Ampere80 GB

Ampere’s 80 GB workhorse. Slower than Hopper, cheaper than denial. Still the right hour when the software is stuck on A100 images or the H100 premium is a tax you will not earn back.

Author Valentin FourierLast modified

VRAM80 GB
From /GPU/hr$1.19
ArchitectureAmpere
VendorNVIDIA

In brief

What to know before you rent

Catalog listings

A100 Pricing and Availability

Published on-demand rates first. Spot and waitlists are labeled. Confirm the live rate before you provision.

ProvidersRegionBilling typeInterconnect/GPU/hr
Vast.aiIn stockUS-EastOn-demandPCIe$1.19Open provider
LambdaIn stockUS-WestOn-demandNVLink$1.29Open provider
RunPodIn stockUS-EastOn-demandNVLink$1.39Open provider
CoreWeaveIn stockUS-EastOn-demandInfiniBand NDR$1.79Open provider
AWSLimitedUS-EastOn-demandNVLink$2.48Open provider
Google CloudLimitedUS-WestOn-demandNVLink$2.93Open provider
AzureLimitedUS-EastOn-demandNVLink$3.67Open provider

Our reading

Who this GPU is for

I rent A100 when the stack is already Ampere, when H100 is sold out, or when the hour delta does not pay for a CUDA/driver move. I do not rent A100 to “save money” on a job that will run 3× longer. Time is GPU rent.

Rent it if Established training recipes, inference that fits 80 GB without FP8, budget Hopper-alternatives with NVLink.

Skip it if Greenfield FP8 training, people who needed H200 memory, anyone offered a 40 GB A100 as if it were 80.

How we compare hours

Datasheet

Technical specifications

NVIDIA A100 · Per GPU · SXM

Compute · dense

FP16 / BF16
312 TFLOPS624 with sparsity, SXM 80 GB
TF32
156 TFLOPS312 with sparsity
INT8
624 TOPS1,248 with sparsity
FP32
19.5 TFLOPS
FP64
9.7 TFLOPS19.5 TFLOPS Tensor Core
Precision support
FP16, BF16, TF32, FP32, FP64, INT8

Memory

Capacity
80 GB HBM2e40 GB SKUs exist, not this row
Bandwidth
2,039 GB/sSXM 80 GB
ECC
Yes

Silicon

Architecture
Ampere
Process
TSMC 7N
Shader cores
6,912CUDA cores
Matrix cores
4323rd-gen Tensor cores
Compute units
108SMs

Fabric and host

GPU interconnect
NVLink 600 GB/s
Host interface
PCIe 4.0 x16

Power

Board power
400 WSXM; PCIe ~250–300 W

Platform

Partitioning
MIG, up to 7 instances

Table is A100 80 GB SXM. PCIe 80 GB and any 40 GB card are slower or smaller. Lambda and hyperscalers still list A100 for a reason: the software already works.

Trade-offs

Pros and cons of A100

What this GPU does well, and when the hour is the wrong buy.

Strengths

  • 80 GB with NVLink and MIG, a real datacenter product, not a GeForce
  • Deep catalog, mature images, fewer Blackwell surprises
  • Often the cheapest *serious* 80 GB hour
  • SXM 400 W is easier to host than 700 W Hopper

Limits

  • No FP8. Transformer Engine is a Hopper story
  • 2 TB/s HBM2e vs 3.35 TB/s HBM3, bandwidth-bound jobs feel it
  • 40 GB vs 80 GB confusion is still how people get burned
  • NVLink 3 at 600 GB/s is a generation behind

Still rentable, not nostalgic

A100 is not a museum piece. It is the 80 GB SKU with the longest operational memory in this market. If your Dockerfile says pytorch 1.13 and A100, paying Hopper rent to feel modern is a personality choice.

If you are starting clean and the job is FP8-shaped, skip to H100. Do not build a new stack on Ampere because the hour is $0.40 cheaper and the step is 2× slower.

How I actually rent it

  1. 01

    Demand 80 GB

    40 GB A100 is a different GPU. Reject the swap.

  2. 02

    Check NVLink

    Multi-GPU Ampere without NVLink is a PCIe tax.

  3. 03

    Price vs H100

    Convert to $/epoch, not $/hour. Hopper often wins the bill.

FAQ

A100 questions

Who has the cheapest A100?+

As of September 5, 2026, the cheapest verified in-stock on-demand A100 tracked by GPUBeacon is $1.19 per GPU per hour from Vast.ai.

How much does the A100 cost per month?+

As of September 5, 2026, at 720 hours per month, one A100 costs an estimated $1,289 at the median on-demand price in this sample. Cheapest verified in stock: $857 per month on-demand.

Where can I rent an A100?+

GPUBeacon currently tracks A100 configs from 7 providers. The cheapest verified in-stock on-demand configs come from Vast.ai, Lambda, RunPod and CoreWeave. See the full price comparison above for every provider and config.

Are A100 prices going up or down?+

GPUBeacon publishes a catalog sample, not a 90-day price index. As of September 5, 2026, the median on-demand A100 in this sample is $1.79 per GPU per hour. The cheapest verified in-stock hour is $1.19.

Why choose the A100?+

Established training recipes, inference that fits 80 GB without FP8, budget Hopper-alternatives with NVLink.

When is the A100 not a good fit?+

Greenfield FP8 training, people who needed H200 memory, anyone offered a 40 GB A100 as if it were 80.

A100 80 GB or H100 80 GB?+

Same VRAM integer, different memory and Tensor generation. H100 if you can use FP8 or need bandwidth. A100 if the image is Ampere and the hour delta is real after wall-clock.

Is A100 40 GB fine?+

For some inference. Not as a silent substitute for 80 GB training. This review is the 80 GB SXM.

Keep comparing

Alternatives to NVIDIA A100

Same job, different envelope. Compare VRAM, bandwidth and the hour before you lock a badge.

NVIDIA H100

80 GB HBM3 · 64% more bandwidth than the A100

From$1.73/ GPU / hr

Read the verdict

NVIDIA L40S

48 GB GDDR6 · 58% less bandwidth than the A100

From$0.79/ GPU / hr

Compare

NVIDIA A10

24 GB GDDR6 · 71% less bandwidth than the A100

From$0.45/ GPU / hr

Compare

NVIDIA H200

141 GB HBM3e · 2.4x the bandwidth of the A100

From$3.00/ GPU / hr

Compare

Browse all GPUs

Sources

Where these numbers come from

TFLOPS and TDP are from the vendor datasheet. Hourly prices are from the GPUBeacon catalog sample. Open the originals, then confirm the live SKU with the provider before you rent.