Amazon Says Bedrock Now Mostly Runs On Trainium

Amazon AMZN says more than half of its Bedrock AI service now runs on its own custom chips, marking a major shift toward in-house infrastructure at AWS.

AWS Chief Marketing Officer Julia White told The Information's TITV that BedrockAmazon's cloud service for hosting and training AI modelsrelies primarily on its Trainium processors, with Nvidia (NVDA) GPUs still used alongside them. It's the first time Amazon has disclosed how widely Trainium is deployed inside Bedrock, signaling a meaningful step toward reducing reliance on Nvidia's hardware.

Bedrock hosts models from partners like Anthropic, and until now, it was unclear how much of the service ran on Amazon's own chips. White's comments suggest that using Trainium helps AWS cut costs and boost margins, since the chips are purpose-built for AI training and inference, not general graphics work. AWS also offers Trainium-based servers at a lower price than GPU-powered ones, adding competitive leverage as AI cloud demand surges.

Building around Trainium could help Amazon control hardware supply, pricing, and performance as hyperscalers look to curb their dependence on Nvidia. Microsoft (MSFT) has also said it plans to rely more on its own custom silicon in future data centers.

Investors will watch for how far AWS extends Trainium adoption and whether cost savings show up in upcoming quarters.