GPU rental pricing
AWS GPU pricing: rental cost per hour by GPU and region
Direct answer
AWS rents GPU instances from about $0.42 per GPU-hour on-demand (T4g in us-east-1), or roughly $0.06 per GPU-hour on spot. This page tracks 661 instance rows across 12 GPU models (T4, A10G, T4g, L4, L40S, V100, and more) and 23 regions, from the ByteCosts GPU catalog. Per-GPU rate is the instance hourly price divided by GPU count, so multi-GPU boxes can look cheaper per card. Spot and preemptible capacity is not guaranteed and can be reclaimed, so size on-demand for steady jobs and treat spot as a discount for interruptible work.
Compare GPU rental across providers - Cheapest per-GPU-hour by accelerator →
AWS GPU rental at a glance
- Cheapest on-demand: $0.42 per GPU-hour (T4g, g5g.xlarge, us-east-1)
- Cheapest spot: $0.06 per GPU-hour (T4g)
- 12 GPU models and 23 regions tracked
- 661 instance rows, updated June 15, 2026
Cheapest AWS rate by GPU model
Lowest on-demand per-GPU-hour rate for each GPU model AWS offers, from the ByteCosts catalog. Per-GPU rate is the instance hourly price divided by GPU count.
| GPU model | Per-GPU/hr | Instance | GPUs | Region | Instance/hr |
|---|---|---|---|---|---|
| T4g | $0.42 | g5g.xlarge | 1 | us-east-1 | $0.42 |
| T4 | $0.53 | g4dn.xlarge | 1 | us-east-1 | $0.53 |
| L4 | $0.80 | g6.xlarge | 1 | us-east-1 | $0.80 |
| A10G | $1.01 | g5.xlarge | 1 | us-east-1 | $1.01 |
| L40S | $1.86 | g6e.xlarge | 1 | us-east-1 | $1.86 |
| A100 | $2.74 | p4d.24xlarge | 8 | us-east-1 | $21.96 |
| V100 | $3.06 | p3.16xlarge | 8 | us-east-1 | $24.48 |
| A100-80GB | $3.43 | p4de.24xlarge | 8 | us-east-1 | $27.45 |
| V100-32GB | $3.90 | p3dn.24xlarge | 8 | us-east-1 | $31.21 |
| H100 | $6.88 | p5.48xlarge | 8 | us-east-1 | $55.04 |
| H200 | $7.91 | p5en.48xlarge | 8 | us-east-1 | $63.30 |
| B200 | $14.24 | p6-b200.48xlarge | 8 | us-east-1 | $113.93 |
When AWS GPUs are cost-effective
Headline per-hour price is only half the story for GPU rental. The variables that move a training or inference bill are GPU utilization (idle reserved capacity is pure waste), whether your job tolerates spot interruption, region and egress fees, and the storage and networking attached to the instance.
AWS tends to win when you can keep the GPU busy, batch interruptible work onto spot, and pick a region close to your data. It tends to lose when capacity sits idle between jobs, when egress moves large datasets out repeatedly, or when a smaller or older GPU would clear your throughput target at a lower rate.
Frequently asked questions
How much does a AWS GPU cost per hour?
AWS rents GPUs from about $0.42 per GPU-hour on-demand for T4g (the g5g.xlarge instance in us-east-1), up to higher rates for newer accelerators. See the table above for the cheapest rate per GPU model. Rates are list rental prices before committed-use discounts.
Is spot or on-demand cheaper on AWS?
Spot (interruptible) capacity on AWS can drop to about $0.06 per GPU-hour, well below on-demand, but it can be reclaimed at any time. Use spot for checkpointed, restartable jobs and on-demand for steady serving where an interruption would be costly.
Where does ByteCosts get AWS GPU prices?
Rates come from the SkyPilot open GPU catalog, which mirrors provider pricing, normalized into the ByteCosts GPU index and dated. The provider's own pricing API is the ground truth; the catalog is used for discovery and cross-checking, not re-export. Verify the provider source before production commitments.
Cite this page
AWS GPU rental pricing. ByteCosts. Updated June 15, 2026. https://bytecosts.com/gpu/aws/