# Trainium2 / Inferentia2 Neuron adoption checklist (July 2026)

Verify instance SKUs, Regions, and on-demand rates on [EC2 accelerated computing](https://aws.amazon.com/ec2/instance-types/accelerated-computing/) and [Trn2](https://aws.amazon.com/ec2/instance-types/trn2/) before board math. AWS has published **~30–40% better price performance** for Trn2 vs current-generation P5e/P5en in launch materials — treat any wider % as a **workload-specific** measurement, not a guarantee.

## Fit gate (fail fast → stay on GPU)

- [ ] Model traces cleanly (stable shapes; no custom CUDA kernels you cannot replace)
- [ ] Framework path supported (PyTorch/JAX via Neuron; pin DLC tag)
- [ ] Team can own AOT compile + artifact versioning (model + Neuron SDK + shapes)
- [ ] Region has `trn2` / `inf2` capacity for the target size

## Instance pins (confirm current)

| Family       | Example size                              | Notes                                                       |
| ------------ | ----------------------------------------- | ----------------------------------------------------------- |
| Trainium2    | `trn2.48xlarge`                           | 16× Trainium2, **3.2 Tbps** EFA (not older 4.8 Tbps slides) |
| UltraServers | `trn2u.48xlarge`                          | UltraServer path — **preview** status can change            |
| Inferentia2  | `inf2.{xlarge,8xlarge,24xlarge,48xlarge}` | Up to 12 chips / 384 GB HBM on 48xlarge                     |

## Migration steps

1. Compile on Neuron DLC; store artifact in versioned S3
2. Equivalence test vs GPU baseline (eval set + latency/throughput)
3. Shadow traffic or A/B endpoint
4. Cost per 1M tokens / per training run vs current GPU SKU (**your** rates)
5. Only then flip SageMaker/EC2 default

## Monday smoke

```bash
aws ec2 describe-instance-types --region us-east-1 \
  --filters Name=instance-type,Values=trn2.48xlarge,inf2.xlarge,inf2.48xlarge \
  --query 'InstanceTypes[].{Type:InstanceType,Network:NetworkInfo.NetworkPerformance}' \
  --output table
```