AI workloads are evolving rapidly, with customers demanding higher performance, lower latency, and dramatically better price-performance for training and serving next-generation agentic, reasoning, multimodal, and video generation models. These models push hardware to its limits across memory bandwidth, compute density, and interconnect throughput. AWS built Trainium3—our first 3nm AI accelerator—to deliver the best token economics and support frontier-scale architectures. Trainium3 provides 2.6 PFLOPs of FP8 compute, 144 GB of HBM3e, 4.7 TB/s memory bandwidth, and up to 2.6× performance per chip over Trn2. Customers can scale to 144 chips per UltraServer, or to hundreds of thousands of chips through EC2 UltraClusters 3.0, enabling unprecedented scale for LLMs, MoE, RL, reasoning, and multimodal workloads. Trainium adoption has already scaled to millions of chips in production, and Trainium3 builds on that momentum to serve the next wave of AI use cases—including real-time experienc...