Amazon Unveils AI Chip Trainium3

AWS unveils Trainium3 AI chip and plans for Trainium4.

Amazon Web Services (AWS) has introduced the latest version of its proprietary AI chip, Trainium3, and announced the development of the next iteration, Trainium4.

Trainium3 UltraServers now available: Enabling customers to train and deploy AI models faster at lower cost. #AWSreInvent ⬇️ https://t.co/YV8iUyKUB8

— Amazon News (@amazonnews) December 2, 2025

The firm announced the launch of the UltraServer system, which is based on the advanced 3-nm Trainium3 processor and internal networking technology.

Both solutions have demonstrated significant performance improvements in AI training and inference compared to the second-generation semiconductor.

The system shows a fourfold increase in performance and also has four times more memory. This allows not only for AI training but also for servicing AI applications during peak loads.

UltraServer consists of 144 Trainium3 chips. Systems can be combined into a structure that can accommodate up to 1 million semiconductors in total.

The energy efficiency of the new solution is improved by 40%. The company reported that its clients, including Anthropic, LLM Karakuri, SplashMusic, and Decart, are already using third-generation chips. As a result, they have significantly reduced computing costs.

Trainium4

Amazon has presented a roadmap for semiconductor development. The next-generation chip, Trainium4, is already in development. It will provide “another big leap” in performance and will support Nvidia’s high-speed NVLink Fusion technology.

Systems based on Trainium4 will be able to scale computing power by interacting with Nvidia graphics processors while continuing to use Amazon’s own, more cost-effective server rack technology.

AI Agents

Amazon Web Services announced three new AI agents. One of them can learn user preferences to autonomously operate for several days.

Each digital assistant performs different tasks:

• Kiro — an assistant for programmers. It asks the user for instructions, confirms or corrects assumptions, and then independently completes the task;
• AWS Security Agent — operates independently to detect security issues as code is written;
• DevOps Agent — automatically tests a new algorithm for performance issues or compatibility with other software, hardware, or cloud settings.

Kiro maintains “persistent context between sessions”—its memory does not expire. The agent can work for several hours or days with minimal human intervention.

Sovereign AI

Another major release from Amazon is AI Factories. This solution allows large corporations and government entities to deploy AI systems directly in their data centers.

“Clients provide power and data centers, while AWS installs the AI system, manages it, and can connect it to other company cloud services,” the firm stated.

The idea is to meet the needs of corporations and governments interested in having absolute control over their data.

Back in November, Amazon demanded that Perplexity remove a browser with an integrated AI agent from its online store.