GPU pricing
AWS GPU pricing: on-demand and spot cost per GPU-hour
AWS on-demand GPU rentals in this catalog start at about $0.42 per GPU-hour (T4g on g5g.xlarge in us-east-1). AWS bills GPU instances by the hour (per-second after the first minute). ByteCosts divides the instance list price by GPU count so an 8×H100 p5.48xlarge compares with a single L4 g6.xlarge. Filter by GPU model, instance family, region, and on-demand versus spot. Spot is interruptible; on-demand is the rate to size a steady job.
Open AWS GPU explorer - Filter instance family, region, and spot →
Cheapest AWS on-demand rate by GPU model
Lowest on-demand dollars per GPU-hour for each AWS GPU model in the committed catalog (717 instance rows).
| GPU model | Per-GPU/hr | Instance | GPUs | Region |
|---|---|---|---|---|
| T4g | $0.42 | g5g.xlarge | 1 | us-east-1 |
| T4 | $0.53 | g4dn.xlarge | 1 | us-east-1 |
| L4 | $0.80 | g6.xlarge | 1 | us-east-1 |
| A10G | $1.01 | g5.xlarge | 1 | us-east-1 |
| L40S | $1.86 | g6e.xlarge | 1 | us-east-1 |
| A100 | $2.74 | p4d.24xlarge | 8 | us-east-1 |
| RTXPRO6000 | $3.36 | g7e.2xlarge | 1 | us-east-1 |
| A100-80GB | $3.43 | p4de.24xlarge | 8 | us-east-1 |
| V100-32GB | $3.90 | p3dn.24xlarge | 8 | us-east-1 |
| H100 | $6.88 | p5.48xlarge | 8 | us-east-1 |
| H200 | $7.91 | p5en.48xlarge | 8 | us-east-1 |
| B200 | $14.24 | p6-b200.48xlarge | 8 | us-east-1 |
The decision this page helps you make
Filter AWS EC2 GPU instances by model, family, region, and on-demand versus spot. Source-backed dollars per GPU-hour from the committed ByteCosts catalog.
The practical question is which AWS GPU instance, region, and pricing mode delivers the required accelerator at a sustainable hourly rate. Use the same workload assumptions for every option so the comparison reflects billing differences instead of different inputs.
Start with these inputs
- Rate: On-demand or spot dollars per GPU-hour.
- Shape: GPU model, instance family, GPU count.
- Place: AWS region.
What the result includes
| Area | What ByteCosts shows |
|---|---|
| Rate | On-demand or spot dollars per GPU-hour |
| Shape | GPU model, instance family, GPU count |
| Place | AWS region |
How to use the result
- Run a realistic base case and a heavier-usage case before choosing a provider or plan.
- Compare alternatives with identical traffic, token, seat, runtime, and retry assumptions.
- Open the cited provider source before a purchase or production billing decision.
Formula
perGpuHour = instanceHourlyUsd / gpuCount. Spot per GPU-hour = spotUsd / gpuCount when a spot price is published.
Assumptions
- Visible prices come from the committed ByteCosts GPU catalog filtered to providerId=aws.
- Linux list prices only; Windows licenses and Inferentia/Trainium chips are excluded.
- Spot capacity is interruptible and not guaranteed.
- Committed-use, reserved, and private-rate discounts are excluded.
Example scenario
Take an 8-GPU instance list price and divide by 8 to get dollars per GPU-hour, then compare that rate with a one-GPU instance of a different family in the same region.
How to read the example
| Step | Example input | What to inspect |
|---|---|---|
| Instance/hr | AWS on-demand or spot list price | What AWS bills for the box |
| GPUs | Accelerator count on the instance | Divisor for a fair per-card rate |
| Per GPU-hour | Instance/hr divided by GPU count | Comparable rental rate |
Interpretation guide
- Compare instances only after normalizing to dollars per GPU-hour.
- Size steady jobs on on-demand; treat spot as a discount for interruptible work.
- Recheck the AWS price list for the target region before committing spend.
Limitations
AWS GPU pricing: on-demand and spot cost per GPU-hour is a planning tool, not a billing guarantee. It uses the visible assumptions and committed source-backed data available at the page’s last update.
Check the cited provider page and your own production logs before signing a contract, changing price, or committing infrastructure spend.
Frequently asked questions
What should I enter first in AWS GPU pricing: on-demand and spot cost per GPU-hour?
Start with rate: on-demand or spot dollars per gpu-hour. Add optional adjustments only after the base case is understandable.
Is the result a guaranteed invoice forecast?
No. It is a planning estimate based on the visible workload assumptions and source-backed public prices. Taxes, negotiated discounts, undocumented limits, and production behavior can change the final invoice.
Where do the prices and assumptions come from?
ByteCosts keeps provider source links, confidence information, and last-checked dates attached to pricing records. User-entered workload assumptions remain separate from published vendor facts.
AWS GPU pricing: on-demand and spot cost per GPU-hour. ByteCosts. https://bytecosts.com/tools/aws-gpu-pricing/