AWS's Project Rainier: the most powerful computer for training AI (www.aboutamazon.com)

AWS has turned on Project Rainier, a massive EC2 UltraCluster built around nearly half a million Trainium2 chips and already running customer workloads (notably Anthropic’s Claude). Deployed less than a year after announcement, the cluster uses Trainium2 UltraServers—each combining four Trainium2 servers (16 chips each) into 64‑chip nodes linked by high‑speed NeuronLinks—and scales across multiple U.S. data centers using Elastic Fabric Adapter (EFA) for cross‑server connectivity. AWS says the Trainium2 footprint is ~70% larger than any prior AWS AI platform, and expects Claude to run on more than 1 million Trainium2 chips by the end of 2025. For researchers and engineers this matters because it dramatically increases available training and inference capacity while lowering cost and latency for very large models. AWS’s vertical integration (chip design, server architecture, software stack and data‑center engineering) enables end‑to‑end optimizations—power delivery, network topology and orchestration—that improve performance and reliability at scale. Project Rainier also emphasizes efficiency and sustainability: AWS matches its data‑center electricity with renewables and reports a water‑use efficiency roughly twice the industry average (0.15 L/kWh). The result is a new template for hyperscale model development that could accelerate frontier AI work across science, healthcare and climate applications.