.svg){kind=spacer}
Details
- AWS introduced Trainium3 UltraServers at re:Invent 2025, powered by its first 3-nanometer AI chip for large-scale model training and inference.
- Each UltraServer can house up to 144 Trainium3 chips and scale to connect thousands, supporting deployments with up to one million chips.
- Trainium3 delivers 4.4 times more compute, is four times more energy-efficient, and offers four times the memory bandwidth of its predecessor, Trainium2.
- Major customers, including Anthropic, Karakuri, Metagenomi, NetoAI, Ricoh, Splash Music, Decart, and Amazon Bedrock, have reported up to 50% lower costs and significantly improved performance.
- Decart achieved four times faster generative video inference at half the cost of traditional GPUs, while AWS also announced further open-sourcing of its Neuron SDK and native PyTorch integration to simplify AI development.
Impact
This launch highlights AWS's escalating effort to challenge NVIDIA's lead in the AI chip market, offering disruptive cost and energy efficiency advantages for enterprise AI workloads. Backed by early customer successes, Trainium3 strengthens AWS's position in the AI infrastructure race and signals the growing importance of custom silicon in cloud computing strategies.
