Inside Amazon’s Trainium lab
Reporting on Amazon’s Trainium lab shows Anthropic, OpenAI and even Apple testing Trainium, indicating hyperscaler custom silicon is moving from experiment to enterprise adoption. The coverage frames Trainium as a credible alternative for big model training workloads. (webpronews.com) (bez-kabli.pl)
Amazon announced a $50 billion strategic investment in OpenAI and committed to supplying OpenAI with approximately 2 gigawatts of Trainium compute capacity under the same partnership. (aboutamazon.com) AWS says there are 1.4 million Trainium chips deployed across three generations, and the company reports Anthropic’s Claude is running on more than 1 million Trainium2 chips as part of Project Rainier. (techcrunch.com) Project Rainier itself was described by AWS as delivering nearly half a million Trainium2 chips in its initial deployment, with an explicit plan to scale past one million Trainium2 processors by year‑end. (aboutamazon.com) AWS’s Trainium3 Trn3 UltraServers can hold up to 144 chips, deliver up to 362 FP8 PFLOPs per system, and AWS advertises Trn3 as offering roughly 4.4x compute versus Trn2 with up to 4x better energy efficiency and customer‑reported cost reductions up to ~50% for some training and inference workloads. (aboutamazon.com) Anthropic engineers reportedly contributed to Trainium’s instruction set and interconnect design, and AWS engineers told touring journalists that in many cases porting a PyTorch model to Trainium can be done with a single line change and a recompile. (techcrunch.com) AWS says Trainium2 handles the majority of inference traffic on Amazon Bedrock, and TechCrunch and other outlets note AWS hosted a private tour of its Austin Trainium lab led by Kristopher King and Mark Carroll as the company scales production. (techcrunch.com)
