Customer Stories¶
Since its launch, AWS Trainium/Inferentia has been rapidly adopted by companies across a wide range of industries and scales. From AI startups to global enterprises, customers who have experienced outstanding price-performance in large-scale language model training and inference workloads are accelerating their transition to AWS-designed chips. In particular, companies such as Anthropic, OpenAI, and ByteDance are achieving business results through large-scale Trainium/Inferentia deployments.
- Anthropic — Operates more than one million Trainium chips (~1 GW) to train and deploy Claude models. Trn2-based Claude Sonnet 4.5 delivers 1.4x higher output tokens/s than competitors. Additional 5 GW partnership signed (April 2026)
- OpenAI — Agreed to secure 2 GW of Trainium for next-generation Frontier and stateful runtime workloads
- ByteDance — Cut the cost of analyzing tens of billions of video content items per day by 50% with Inferentia2
- Amazon Bedrock — AWS also uses Trainium at scale to run its own services (more than 50% of total token throughput is processed on Trainium)
Explore first-hand testimonials below from customers who have achieved performance improvements and cost optimization with Trainium/Inferentia.
🏆 Featured Stories¶
-
Anthropic — Trainium 2, 3
Training and serving Claude models
-
Decart AI — Trainium 3
Achieved 4x faster performance than GPUs and 80% MFU for real-time video generation
-
Splash Music — Trainium
Cut music generation model training costs by 50%
-
Ricoh — Trainium
Cut Japanese-specialized model training costs by 50% and reduced training time by 25%
-
ScatterLab - Inferentia
Halved model serving costs
-
HyperConnect - Inferentia
Reduced model serving costs to one quarter
-
Klleon - Inferentia
Cut digital human generation model inference costs by 50%
-
Coming Soon
More stories coming soon