[AWS re:Invent] AWS Challenges Nvidia's Dominance... Unveils AI Chip with 4x Performance

Unveiling of Proprietary AI Chip "Trainium 3" with Reduced Power Consumption
Next-Generation Chip Already in Development, Expected to Launch Next Year
"AI Agents Will Become Ubiquitous in Everyday Life and Work"

Amazon Web Services (AWS), the world's largest cloud computing company, has unveiled a new artificial intelligence (AI) chip that boosts computing performance while reducing power consumption. AWS, which has long invested in its own chip design, is now seen as making a direct challenge to Nvidia, which currently dominates the AI semiconductor market, as the era of AI agents unfolds.

On December 2 (local time), Matt Garman, Chief Executive Officer (CEO) of AWS, introduced the new model of its proprietary AI chip, "Trainium 3," during his keynote address at AWS's annual event, re:Invent 2025, held in Las Vegas, Nevada, USA. Trainium is a semiconductor designed by AWS specifically for AI model training.

Matt Garman, Chief Executive Officer (CEO) of Amazon Web Services (AWS), announced the company's proprietary artificial intelligence (AI) chip, "Trainium 3," at AWS's annual event "re:Invent 2025," held on the 2nd (local time) in Las Vegas, Nevada, USA. Provided by AWS

Trainium 3 is manufactured using a 3-nanometer (nm, one-billionth of a meter) process. Compared to its predecessor, it delivers more than four times the computing performance while reducing power consumption by approximately 40%. Lower power consumption translates to reduced operational costs. AWS explained that using this product can reduce AI model training and operational costs by up to 50% compared to using Nvidia graphics processing units (GPUs). CEO Garman emphasized, "Trainium 3 delivers the industry's best price-to-performance ratio for large-scale AI training and inference."

The next-generation chip, Trainium 4, was also announced for future release. This chip is expected to offer six times higher computing performance and four times the memory bandwidth compared to Trainium 3, and is scheduled to launch next year. AWS stated that Trainium 4 will support Nvidia's chip interconnect technology, NVLink. Given Nvidia's overwhelming market share, this feature is intended to enhance convenience for customers who prefer Nvidia GPUs. However, it also appears to be a strategic move to make it easier for customers who have primarily used Nvidia chips to eventually switch to AWS's AI chips.

CEO Garman said, "AWS has long recognized that AI chips would be a game changer," adding, "That is why we began investing in custom chips ten years ago." AWS entered proprietary chip development by acquiring Israeli semiconductor company Annapurna Labs in 2015, and has since designed Trainium as well as Inferentia, an AI chip specialized for inference.

A new service called "AI Factories," which implants enhanced AI infrastructure into customers' data centers, was also announced. This service allows companies or organizations to deploy AWS's AI infrastructure within their own data centers. AWS offers a variety of AI infrastructure options, including Nvidia GPUs and Trainium chips. By integrating these into customers' data centers, companies can mitigate concerns about sensitive data leakage during AI training processes.

During his keynote, CEO Garman stressed that the emergence of AI agents capable of performing users' work will mark a "turning point" that changes the course of the era. He stated, "AI agents are replacing traditional AI assistants," and predicted, "The advent of AI agents will bring about an innovation as significant as the initial emergence of AI."

He also introduced new features added to "Bedrock AgentCore," a tool for developing AI agents. AgentCore is a fully managed platform that supports the entire process of building, deploying, and operating AI agents, and was officially launched last October. With AgentCore, companies can select their preferred models and frameworks to build customized AI agents.

Matt Garman, Chief Executive Officer (CEO) of Amazon Web Services (AWS), is delivering the keynote speech at AWS's annual event 're:Invent 2025' held on the 2nd (local time) in Las Vegas, Nevada, USA. Provided by AWS