Amazon Web Services officially released its Trainium3 UltraServers on Tuesday at the re:Invent conference in Las Vegas, marking a significant push by the cloud giant to challenge Nvidia’s dominance in the artificial intelligence chip market with systems promising to cut AI training and inference costs by up to 50%.
Raw Performance Gains#
The new servers pack up to 144 Trainium3 chips built on 3-nanometer technology into a single integrated system, delivering 4.4 times more compute performance and 40% better energy efficiency than the previous Trainium2 generation. Customers testing the systems with OpenAI’s open-weight GPT-OSS model achieved three times higher throughput per chip and four times faster response times.
Early Adopters Report Major Savings#
Several companies are already deploying the new chips in production. Anthropic, which trains its Claude AI models on AWS infrastructure, is among the early adopters alongside Karakuri, Metagenomi, NetoAI, Ricoh, and Splash Music—all reporting cost reductions of up to 50% compared to GPU-based alternatives. Amazon Bedrock, the company’s managed foundation model service, is already serving production workloads on Trainium3.
Decart, an AI lab specializing in real-time generative video, achieved particularly striking results, generating video frames four times faster at half the cost of GPUs. “We need to demonstrate that we have a product that meets their performance requirements at an attractive price point,” Dave Brown, AWS president of compute and learning services, stated.
Scaling to Million-Chip Clusters#
The Trainium3 systems can scale up to 1 million chips across AWS’s EC2 UltraClusters 3.0, representing 10 times the capacity of the previous generation. This expansion builds on Project Rainier, AWS’s collaboration with Anthropic that currently operates with nearly 500,000 Trainium2 chips and is expected to reach 1 million chips by year-end.
Trainium4 Already in Preview#
AWS also previewed Trainium4, which will integrate Nvidia’s NVLink Fusion high-speed interconnect technology, enabling Trainium and Nvidia systems to work together seamlessly. The forthcoming chip promises at least six times the FP4 processing performance, three times the FP8 performance, and four times more memory bandwidth than Trainium3.
AWS simultaneously launched new P6e instances featuring Nvidia’s GB300 NVL72 platform, underscoring its strategy of offering both proprietary and third-party chip options as competition intensifies with Google’s TPU chips and Nvidia’s GB300 systems.