AWS Trainium Customers

Performance and scale aren't just technical requirements - they're essential to achieving that mission. That's why we partnered with AWS as our primary cloud provider to build Project Rainier, one of the world's most powerful operational AI supercomputers. With almost a million Trainium2 chips training and serving Claude today, we're excited about Trainium3 and expect to continue to scale Claude well beyond what we've built with Project Rainier, pushing the boundaries of what's possible in AI.

James Bradbury, Head of Compute, Anthropic.

Our partnership with AWS gives us both. Trainium allows our customers to scale their usage of poolside at a price performance ratio unlike other AI accelerators. And, Trainium's upcoming native Pytorch and vLLM support will unlock even more innovation and flexibility for Trainium users, including poolside. Above all, AWS' customer focus shines through, and AWS were able to quickly iterate and use our feedback to adapt Trainium to our needs. We look forward to deepening our collaboration on all aspects of Trainium.

Joe Rowell founding Engineer

Trainium’s unique architecture—efficient memory hierarchy and high-throughput AI engines—proved ideal for Decart’s real-time video models, driving full utilization of the hardware. Early testing shows up to 4x higher frame throughput and 2x better cost efficiency compared to top GPUs, with latency reduced from 40 ms to 10 ms. This performance enables live, dynamic, and interactive video generation at scale—previously impractical on standard hardware. Through Bedrock, these capabilities will soon be directly accessible to AWS customers.

Dean Leitersdorf Co-founder & CEO
 

By adopting AWS Trainium, we reduced our LLM training costs by over 50 percent while maintaining consistent infrastructure availability. This has enabled us to create Japan's most accurate Japanese language model while staying well below budget. The infrastructure stability has also delivered unexpected productivity gains, allowing our team to focus on innovation rather than troubleshooting.

Tomofumi Nakayama CPO

In 2025 the AI community reached an inflection point with over 10 million AI Builders using and sharing millions of open models and datasets on Hugging Face. It’s now more important than ever to reduce the cost of running ever larger and more diverse open models, to make sure AI benefits everyone and every industry. At Hugging Face, we have been working hand in hand with the engineering teams at AWS building purpose-built AI chips since the first Inferentia1 instances became available. So today, we are incredibly excited about Trainium3, the next generation of AWS AI Chips that will power the most demanding AI applications, from MoE LLMs to Agents and video generation models. With Optimum Neuron, we are committed to bringing the high memory and cost efficiency benefits of Trainium 3 to the millions of users of Transformers, Accelerate, Diffusers and TRL, so they can build their own AI while controlling their costs.

PyTorch's vision is straightforward: the same code should run everywhere on any hardware platform. AWS's native Trainium support brings this hardware choice to researchers who need to rapidly experiment and iterate freely. With the launch of AWS Trainium3, PyTorch developers can research, build and deploy their ideas at higher performance, lower latency and better token economics, all while maintaining their familiar PyTorch workflows and staying within the ecosystem they already know.