Catalog
P/G instance families as on-demand when published. Spot and capacity blocks labeled or excluded from the on-demand rank.
Provider review / verified 2026-09-05
AWSGPU review
AWS sells P- and G-class GPU instances inside the rest of Amazon. Credits, IAM, PrivateLink, and a bill that includes everything except peace. On-demand exists. It is rarely the floor.
Models listed4
From /GPU/hr$0.81
Regions in sample1
Business modelHyperscaler
Our reading
Who this provider is for
I use AWS GPUs when the company already lives there, data in S3, SSO, the security questionnaire that would take longer than the training run. I do not use AWS to save money on an H100. The hour is the tax you pay for the rest of the platform. If a startup with no AWS estate picks P-instances because ‘nobody got fired,’ I will argue with them. If a bank does it, I will nod and go back to the catalog.
Rent it if Teams whose gravity well is already AWS, who need GPU next to the rest of the estate.
Skip it if Anyone whose only workload is a GPU and who can tolerate a specialist cloud. You are donating margin.
How we compare hoursStrengths
- The rest of the cloud: IAM, networking, storage, compliance paper
- On-demand, reserved, and spot, as long as you keep the units honest
- Regions everywhere the rest of your company already is
- The default that procurement already knows how to buy
Limits
- List on-demand GPU hours are often worse than neoclouds for the same NVIDIA name
- Quota theatre: you may not get the instance until a ticket says so
- Spot is not on-demand. Capacity blocks are not on-demand. Stop mixing them in slides
- Complexity is a cost. A P5 that needs a networking priest is not a pod
Same standard, every cloud
What we measured on AWS
Four checks we apply to every provider. This is not a score, and it is not a paid ranking.
Price signal
Usually the expensive honest number. Neoclouds exist because of this cell.
Product shape
Hyperscaler GPU as a SKU among thousands. The platform is the product.
Still verify
Quota, AZ, instance size (xlarge vs metal), and whether you are looking at spot.
- Model
- Hyperscaler
- Headquarters
- Seattle, US
- Setup
- EC2 P/G instances
- On-demand
- Yes
- Spot
- Yes, labeled, not mixed
- Why people stay
- The rest of AWS
- Why people leave
- The GPU hour
- Compare with
- CoreWeave, Lambda, GCP
Catalog listings
AWS catalog sample
Published on-demand hours first. Spot and waitlists are labeled. Confirm the live rate before you provision.
Nobody got fired, and nobody got an H100 either
AWS is where GPU quota goes to wait. I have watched teams spend a week on limits for a job that would have been a RunPod Secure pod after lunch. If your data cannot leave AWS, stay. If your data is a bucket you can copy, look at the specialist hour.
Spot is a different product
AWS Spot on GPU can look like a neocloud. It can also vanish. GPUBeacon will not let a spot P4 pretend to be on-demand. Checkpoint or pay.
When AWS is actually correct
When GPU is 10% of the architecture. When PrivateLink and KMS are not optional. When the model reads from a lake that will never get a SAS URL to Vast.ai. Then the ‘expensive’ hour is cheap in political capital.
How I actually rent it
- 01
Check quota before you dream
P-family limits are the real stock. The pricing page is fiction until quota is real.
- 02
Name the instance family
P4, P5, G6, not ‘an A100.’ Size and GPU count live in the SKU.
- 03
Compare one neocloud anyway
Lambda or CoreWeave, same GPU, on-demand. Know what the tax is.
FAQ
AWS questions
Is AWS more expensive than RunPod?+
Usually yes on on-demand GPU hours. You are buying AWS, not a GPU. Sometimes that is rational.
AWS vs CoreWeave?+
AWS if the estate is AWS. CoreWeave if the estate is the GPU cluster. Do not buy CoreWeave to run your email.
Should I use AWS Spot for training?+
Only with checkpoints and a stomach. It is not on-demand. We will not rank it as such.
Who should skip AWS GPUs?+
Indie labs whose data can move and whose security questionnaire is ‘please do not leak the weights.’
Sources
Where these numbers come from
Prices and SKUs are from the GPUBeacon catalog sample. Ranking rules are on the methodology page. Outbound links may be affiliate. They never change the hour we display.
GPUs
Keep comparing
Other providers to compare
Same fields, different product. Compare the GPU hour before you lock a brand.
