Overview
- AWS said Trainium3 is now available in UltraServers with 4.4x more compute, roughly 4x memory bandwidth, and 4x better energy efficiency than Trainium2, plus support for up to five times more AI tokens per megawatt.
- Executives disclosed Trainium2 has a multi‑billion‑dollar revenue run rate with more than one million chips deployed, and that Bedrock inference largely runs on Trainium.
- Anthropic was highlighted as a heavy user with over 500,000 Trainium2 chips on Project Rainier for training next‑generation Claude models.
- New AI Factories let customers run AWS stacks in their own data centers using Trainium or Nvidia GPUs alongside services such as SageMaker and Bedrock, presented as a private AWS region experience.
- AWS introduced agentic developer and security tools including Bedrock AgentCore policies and evaluations, the Kiro IDE, Security Agent (preview), and IAM Policy Autopilot, as analysts cautioned many enterprises have yet to see AI ROI.