Calculator
One p5.48xlarge left on costs $1,320.96 a day. Price nine GPU instance types below. Then check if CostMon's default alert rule would catch a forgotten node.
Rates verified us-east-1 only. Other regions differ; see the sources below
Rate card
us-east-1, on-demand, Linux. Every rate comes from the AWS Price List Bulk API. The sources section below cites each one.
| Instance | GPUs | GPU memory | $/hour | $/day (24h) | $/month (730h) |
|---|---|---|---|---|---|
| g6.xlarge | 1x NVIDIA L4 | 24 GB | $0.80 | $19.32 | $588 |
| g5.xlarge | 1x NVIDIA A10G | 24 GB | $1.01 | $24.14 | $734 |
| g6e.xlarge | 1x NVIDIA L40S | 48 GB | $1.86 | $44.66 | $1,359 |
| g5.12xlarge | 4x NVIDIA A10G | 96 GB | $5.67 | $136.13 | $4,141 |
| g6e.12xlarge | 4x NVIDIA L40S | 192 GB | $10.49 | $251.82 | $7,660 |
| p4d.24xlarge | 8x NVIDIA A100 | 320 GB HBM2 | $21.96 | $526.98 | $16,029 |
| p5.48xlarge | 8x NVIDIA H100 | 640 GB HBM3 | $55.04 | $1,320.96 | $40,179 |
| p5en.48xlarge | 8x NVIDIA H200 | 141 GB HBM3e* | $63.30 | $1,519.10 | $46,206 |
| p6-b200.48xlarge | 8x NVIDIA B200 | 1432 GB HBM3e | $113.93 | $2,734.39 | $83,171 |
* AWS lists the p5en.48xlarge GPU memory per GPU. Every other row states a total.
Free calculator
The math runs in your browser. Nothing you type leaves this page. The rates are us-east-1 on-demand prices from the rate card above.
Nine EC2 GPU instance types, us-east-1 on-demand rates.
How many of this instance type you're running.
Enter 24 for a node that runs all day.
Hours of training, serving or batch work. Idle hours do not count.
Your usual daily "Amazon Elastic Compute Cloud - Compute" line, before this node.
"Cost of idle hours" is what the unused hours cost. It is not a promised saving. You decide if a workload can stop.
No minimum from 0% to 200% over baseline catches this node. The step is too small for your daily EC2 line.
The baseline is 14 days of your daily EC2 spend. A fixed pattern moves each day by up to ±5%. Day 15 adds the node for 24 hours. After 14 days, the node becomes part of the baseline. So the alert fires on the first days only.
The link carries your numbers.
Where it hides
A training run finishes, and the node keeps billing. Nobody set a reminder to stop it.
The cost lands in the same EC2 compute line as every other instance. A p5.48xlarge left on for one week adds $9,246.72 to that line.
A SageMaker real-time endpoint bills for every hour it exists. A notebook instance bills the same way. A GPU endpoint made for a demo keeps billing until someone deletes it.
SageMaker has its own service line, "Amazon SageMaker". That line is smaller than the EC2 line. So the same node makes a bigger percent step there. A percent-based rule sees that step sooner.
A GPU node group can stay at peak size after the job ends. A wrong autoscaler setting can cause this. A pod disruption budget that blocks scale-in can cause it too.
These nodes bill as normal EC2 instances. A small group hides in the EC2 line. A large group makes a step that the default rule catches.
Price idle cluster capacity on the Kubernetes cost page →A training job's volumes and snapshots keep billing after the compute stops. Each one is small next to the compute cost. They add up over every run that nobody cleaned up.
Find orphaned volumes and old snapshots →A Capacity Block reserves GPU capacity for a fixed future window. You pay the full price when you book it. The Price List shows $0.00/hour on-demand for this option. So the rate table on this page does not quote a Capacity Block rate.
Ways to cut it
Stop training and dev nodes outside working hours. A node that stops for 12 hours each night bills for half the hours.
Inference on a small or mid-size model often fits on one g6 or g6e GPU. Test those first, before a p4d or p5. The rate card above shows the price gap per hour.
A job that saves checkpoints can resume after a Spot interruption. A job without checkpoints loses its work and starts again. Use Spot only for the first kind.
A Savings Plan or Reserved Instance pays off only when usage stays steady. Size it against the floor of usage that always runs.
Check the break-even point in the commitment calculator →These steps do not promise a specific saving. Each fleet runs a different mix of work.
Your AI bill can also include model API calls. See how to cut an LLM API bill, or compare model prices in the LLM pricing table.
Sources
Every $/hour rate above comes from AWS's own pricing data. We checked each rate again on 2026-09-28.
Source for every $/hour rate and GPU memory figure. Filter: OnDemand, Linux, Shared tenancy, no pre-installed software, CapacityStatus=Used.
Confirms the NVIDIA L4 GPU in g6.xlarge.
Confirms the NVIDIA A10G GPUs in g5.xlarge and g5.12xlarge.
Confirms the NVIDIA L40S GPUs in g6e.xlarge and g6e.12xlarge.
Confirms the NVIDIA A100 GPUs in p4d.24xlarge.
Confirms the NVIDIA H100 GPUs in p5.48xlarge and the NVIDIA H200 GPUs in p5en.48xlarge.
Confirms the NVIDIA B200 GPUs in p6-b200.48xlarge.
FAQ
No. CostMon reads AWS Cost Explorer data grouped by linked account and service. It shows the daily "Amazon Elastic Compute Cloud - Compute" line and the "Amazon SageMaker" line. It does not show an instance ID, an instance type or a tag.
The default rule watches the whole EC2 compute line. A p5.48xlarge adds a large step to that line, so the rule fires. A g5.xlarge adds about $24 a day. A large EC2 line absorbs that step, and the rule stays quiet.
You pay for a Capacity Block up front when you book it. The AWS Price List shows $0.00/hour on-demand for it. A per-hour rate would mislead you, so this page does not quote one.
AWS lists the p5en.48xlarge GPU memory per GPU: 141 GB. The other rows show the total for all GPUs. Eight H200 GPUs hold 1,128 GB in total.
The rule compares each day with the 14 days before it. After 14 days, the node is part of that baseline. So the alert fires only in the first days after the node starts.
Yes. You can create a rule scoped to "Amazon Elastic Compute Cloud - Compute". That rule can use a lower minimum percent over baseline. The calculator shows the highest whole minimum that still catches your node. It also tells you when no minimum does.
A monthly invoice tells you what already happened. CostMon builds a daily baseline, so a spike stands out while you can still act on it.