Accelerators pricing
AWS Trainium and Inferentia pricing
Updated 2026-10-19
9 accelerator instance sizes priced in us-east-1. Per hour, credits last longer than on comparable GPU instances; see the comparisons below.
All Trainium and Inferentia instances (us-east-1)
| Instance | Accelerators | Accel. memory | On-demand $/h | $/chip-hr | 3-yr RI no upfront $/h | Hours per $25k |
|---|---|---|---|---|---|---|
| inf1.xlarge | Inferentia ×1 | n/a | $0.228 | $0.228 | $0.11 | 109,649 h |
| inf1.2xlarge | Inferentia ×1 | n/a | $0.362 | $0.362 | $0.174 | 69,061 h |
| inf1.6xlarge | Inferentia ×4 | n/a | $1.18 | $0.295 | $0.567 | 21,186 h |
| inf1.24xlarge | Inferentia ×16 | n/a | $4.72 | $0.2951 | $2.27 | 5,295 h |
| inf2.xlarge | Inferentia2 ×1 | 32 GB | $0.7582 | $0.7582 | $0.3275 | 32,973 h |
| inf2.8xlarge | Inferentia2 ×1 | 32 GB | $1.97 | $1.97 | $0.8501 | 12,704 h |
| inf2.48xlarge | Inferentia2 ×12 | 384 GB | $12.98 | $1.08 | $5.61 | 1,926 h |
| trn1.2xlarge | Trainium ×1 | 32 GB | $1.34 | $1.34 | $0.5123 | 18,605 h |
| trn1.32xlarge | Trainium ×16 | 512 GB | $21.50 | $1.34 | $8.20 | 1,163 h |
Single-accelerator inference: inf2.xlarge vs GPU instances (us-east-1)
inf2.xlarge gives 32,973 hours per $25k vs g5.xlarge 24,851 h (credits last +33% longer on inf2.xlarge), g6.xlarge 31,064 h (credits last +6% longer on inf2.xlarge), g4dn.xlarge 47,529 h (credits last -31% longer on inf2.xlarge)
| Instance | Accelerators | Accel. memory | On-demand $/h | $/chip-hr | 3-yr RI no upfront $/h | Hours per $25k |
|---|---|---|---|---|---|---|
| inf2.xlarge | Inferentia2 ×1 | 32 GB | $0.7582 | $0.7582 | $0.3275 | 32,973 h |
| g5.xlarge | NVIDIA A10G ×1 | 24 GB | $1.01 | $1.01 | $0.4346 | 24,851 h |
| g6.xlarge | NVIDIA L4 ×1 | 24 GB | $0.8048 | $0.8048 | $0.3694 | 31,064 h |
| g4dn.xlarge | NVIDIA T4 Tensor Core ×1 | 16 GB | $0.526 | $0.526 | $0.227 | 47,529 h |
Hours per $25k = how long credits last at the on-demand price. Throughput per hour differs by model and needs the AWS Neuron SDK; benchmark before switching.
Large-model inference: inf2.48xlarge vs GPU instances (us-east-1)
inf2.48xlarge gives 1,926 hours per $25k vs g5.48xlarge 1,535 h (credits last +25% longer on inf2.48xlarge), g6e.12xlarge 2,383 h (credits last -19% longer on inf2.48xlarge)
| Instance | Accelerators | Accel. memory | On-demand $/h | $/chip-hr | 3-yr RI no upfront $/h | Hours per $25k |
|---|---|---|---|---|---|---|
| inf2.48xlarge | Inferentia2 ×12 | 384 GB | $12.98 | $1.08 | $5.61 | 1,926 h |
| g5.48xlarge | NVIDIA A10G ×8 | 192 GB | $16.29 | $2.04 | $7.04 | 1,535 h |
| g6e.12xlarge | NVIDIA L40S ×4 | 192 GB | $10.49 | $2.62 | $4.53 | 2,383 h |
Hours per $25k = how long credits last at the on-demand price. Throughput per hour differs by model and needs the AWS Neuron SDK; benchmark before switching.
Training node: trn1.32xlarge vs GPU instances (us-east-1)
trn1.32xlarge gives 1,163 hours per $25k vs p4d.24xlarge 1,139 h (credits last +2% longer on trn1.32xlarge), p5.48xlarge 454 h (credits last +156% longer on trn1.32xlarge)
| Instance | Accelerators | Accel. memory | On-demand $/h | $/chip-hr | 3-yr RI no upfront $/h | Hours per $25k |
|---|---|---|---|---|---|---|
| trn1.32xlarge | Trainium ×16 | 512 GB | $21.50 | $1.34 | $8.20 | 1,163 h |
| p4d.24xlarge | NVIDIA A100 ×8 | 320 GB | $21.96 | $2.74 | $9.37 | 1,139 h |
| p5.48xlarge | NVIDIA H100 ×8 | 640 GB | $55.04 | $6.88 | $23.78 | 454 h |
Hours per $25k = how long credits last at the on-demand price. Throughput per hour differs by model and needs the AWS Neuron SDK; benchmark before switching.
By chip
Related cost calculators
- EC2 Pricing Calculator · Compare Instance Types & Costs
- Amazon SageMaker AI Pricing Calculator · On-Demand Studio, Training & Inference
Questions
Do AWS credits last longer on Trainium/Inferentia than on GPUs?
inf2.xlarge gives 32,973 hours per $25k vs g5.xlarge 24,851 h (credits last +33% longer on inf2.xlarge), g6.xlarge 31,064 h (credits last +6% longer on inf2.xlarge), g4dn.xlarge 47,529 h (credits last -31% longer on inf2.xlarge). The hourly price is lower; whether cost per token is lower depends on your model's throughput on Neuron.
Does trn2 have an on-demand price?
Not on the AWS trn2 product page we fetched on 2026-10-14: it is sold through EC2 Capacity Blocks, so we do not publish a trn2 price page.
Prices and credit amounts are estimates and may change. Check AWS for current terms.