Particle.news
Download on the App Store

AWS Launches Trainium 3, Deepens AI Stack With Nova Forge and On‑PremAI Factories

The company is pushing customers toward lower‑cost, customizable AI by pairing its new silicon with tools to build and run bespoke models in the cloud or on private infrastructure.

Overview

  • AWS says Trainium 3 delivers up to 4.4× higher performance, roughly 4× better energy efficiency, and nearly 4× more memory bandwidth versus its prior generation, figures not independently verified in the coverage.
  • Trainium 3 is tightly integrated with services like Bedrock and the Nova family of models, with industrial‑scale use reported by customers such as Anthropic.
  • AWS disclosed that Trainium 4 is in development with planned NVLink compatibility so customers can combine Amazon chips with Nvidia GPUs.
  • Amazon Nova Forge debuts to let organizations pretrain and customize their own frontier models, while Bedrock adds 18 options including Mistral, Gemma 3, Nvidia’s Nemotron and OpenAI’s GPT OSS Safeguard.
  • AWS unveiled turnkey AI Factories for high‑performance on‑prem deployments using Trainium and NVIDIA accelerators; pilots and early projects include Reddit, Siemens, Sony and a Saudi AI zone targeting about 150,000 chips.