Trainium is Amazon's training and inference chip, and the silicon behind Anthropic's Project Rainier, a cluster of about 500,000 Trainium2 chips. Trainium3 moved to 3 nm with 144 GB of HBM3e per chip and up to 144 chips per UltraServer. AWS says Trainium4 will support Nvidia NVLink Fusion so it can share racks with GPUs.
Key Takeaways
- What it is — Hyperscaler custom silicon. ASIC with eight NeuronCores per Trainium3 chip, each with tensor, vector, scalar and GPSIMD engines.
- Best for — Teams already committed to AWS that want a cheaper path for large PyTorch training and inference.
- Watch out — AWS-only, and Neuron is a smaller ecosystem than CUDA. Expect kernel work for non-standard architectures.
At a glance
| Company | Amazon Web Services |
|---|---|
| Current chips | Trainium3 (3 nm) generally available since December 2025; Trainium2 in volume; Trainium4 announced |
| Architecture | ASIC with eight NeuronCores per Trainium3 chip, each with tensor, vector, scalar and GPSIMD engines |
| Memory | Trainium3: 144 GB HBM3e at 4.9 TB/s per chip |
| Scale-up | Trn3 UltraServer: up to 144 chips, 362 MXFP8 petaflops, 20.7 TB HBM3e; UltraClusters reach hundreds of thousands of chips |
| Software | AWS Neuron SDK: PyTorch, vLLM, Hugging Face Transformers, TorchTitan and Ray; NKI for custom kernels |
| Where to get it | AWS only (EC2 Trn2 / Trn3 instances and UltraServers, SageMaker HyperPod, EKS) |
| Pricing | Per instance-hour on EC2; AWS positions it on cost per token rather than publishing a per-chip rate |
What changed in 2026
- Dec 2025. Trainium3 and Trn3 UltraServers generally available at re:Invent; AWS claims up to 4.4x the compute of Trainium2. Trainium4 previewed with NVLink Fusion support.
- 2026. Project Rainier with Anthropic runs on about 500,000 Trainium2 chips. AWS names Anthropic, OpenAI, Databricks and Uber among Trainium customers.
Scores
Scored 1 to 5 on the same five axes as every vendor in the AI chip companies analysis.
| Performance | Scale-up | Software | Access | 2026 momentum | Total / 25 |
|---|---|---|---|---|---|
| 4 | 4 | 3 | 3 | 4 | 18 |
Verdict
The default custom chip for AWS-native teams. It wins on price for large, standard PyTorch jobs, and loses wherever you need multi-cloud portability.
Alternatives
In the same lane (hyperscaler custom silicon):
- Google TPU: large, steady training runs and high-volume serving of standard transformer models, by teams already on Google Cloud.
- Microsoft Maia: nobody buys Maia directly. It matters because it lowers the cost of the Azure models you already pay for.
In other lanes:
- AMD Instinct (merchant gpu)
- Cerebras (inference specialist)
- Groq (inference specialist)
- SambaNova (inference specialist)
- Tenstorrent (open hardware)
Head-to-head comparisons
- Google TPU vs AWS Trainium vs Nvidia GPU: The three ways to train and serve frontier-scale models in 2026, scored on scale-up domain, software, access and price.
- Cerebras vs Groq vs SambaNova: The SRAM-first inference specialists compared on speed, token price, deployment options and company risk.
New to custom AI silicon? Our partner site CTAIO explains how TPUs, pods and GPU clusters differ in TPU vs GPU.
Frequently asked questions
What is AWS Trainium?
Trainium is Amazon's training and inference chip, and the silicon behind Anthropic's Project Rainier, a cluster of about 500,000 Trainium2 chips. Trainium3 moved to 3 nm with 144 GB of HBM3e per chip and up to 144 chips per UltraServer. AWS says Trainium4 will support Nvidia NVLink Fusion so it can share racks with GPUs.
How do you get access to AWS Trainium?
AWS only (EC2 Trn2 / Trn3 instances and UltraServers, SageMaker HyperPod, EKS). Pricing: Per instance-hour on EC2; AWS positions it on cost per token rather than publishing a per-chip rate.
What software does AWS Trainium run?
AWS Neuron SDK: PyTorch, vLLM, Hugging Face Transformers, TorchTitan and Ray; NKI for custom kernels.
Is AWS Trainium a good alternative to Nvidia?
The default custom chip for AWS-native teams. It wins on price for large, standard PyTorch jobs, and loses wherever you need multi-cloud portability. It fits best for: teams already committed to AWS that want a cheaper path for large PyTorch training and inference. The main caveat: aWS-only, and Neuron is a smaller ecosystem than CUDA. Expect kernel work for non-standard architectures.
Sources
Facts verified 2026-10-31. Vendor performance figures are as stated by the vendor.
Ready to Find the Right AI Tools?
Browse our data-driven rankings to find the best AI tools for your team.