At re:Invent 2025, AWS unveils the Trainium3 generation for token factories

During its annual event, AWS unveiled the Trainium3 UltraServers chip, designed for training and inference of artificial intelligence models. This launch represents a strategic milestone in a chip-short market, which is critical for advancing AI. Trainium3 boasts a significant performance leap, enhanced energy efficiency, and powerful aggregation capabilities.

The second day of the AWS re:Invent 2025 conference in Las Vegas kicked off with a slew of announcements led by AWS CEO Matt Garman. This year’s edition highlights the accelerating landscape of generalized AI. The announcements span the emergence of next-generation AI agents, the industrialization of infrastructure through AWS AI Factories, and the scaling of large-scale AI development. At the heart of this initiative, the new Trainium3 UltraServers chip equips the hyperscaler with new capabilities for training and inferring massive AI models, balancing power, cost control, and carbon footprint reduction.

The Trainium3 UltraServers chip standardizes AWS’s commitment to delivering a hardware platform tailored to the evolving needs of generative AI and industrial applications. Each chip provides up to 2.52 petaflops of FP8 precision computing, combined with 144 GB of HBM3e memory and a bandwidth of 4.9 TB/s. The modularity of this offering allows for the aggregation of up to 144 chips within a single UltraServer, achieving a total of 362 petaflops and 20.7 TB of memory with bandwidth reaching 706 TB/s.

A Foundation for Industrializing Giant AI Training

According to AWS, the Trainium3 generation delivers 4.4 times the computational power and four times the energy efficiency of its predecessor while quadrupling memory bandwidth. This density improvement relies on a NeuronSwitch-v1/NeuronLink-v4 internal network interconnection that doubles inter-chip bandwidth and supports distributed training of large models, a critical challenge in contemporary AI.

With the Trainium3 UltraServers, AWS targets companies grappling with increasingly complex AI models for both training and large-scale inference. The integration of HBM3e memory on each chip (144 GB at 4.9 TB/s) addresses the growing need for rapid and simultaneous access to massive datasets. According to AWS, the architecture allows for the federation of up to 144 chips, forming compute clusters that boast 20.7 TB of global memory and 706 TB/s of bandwidth. This design enhances data throughput, reduces latency, and enables dynamic load adaptation based on model requirements.

Performance and energy efficiency are among the key advantages touted by AWS. The Trainium3 chip promises to significantly reduce the cost per model, utilizing performance levels suitable for training giant AI architectures. The new NeuronSwitch-v1/NeuronLink-v4 interconnection enhances calculation synchronization between chips while maintaining low latency. For developers, native integration with Amazon Bedrock AgentCore ensures seamless service continuity, while the enhanced features of Nova 2, Nova Forge, and Nova Act streamline model management, optimization, and testing within industrialized environments.

An Optimized Software Chain for Trainium3

In addition to the hardware announcement, AWS unveiled the rollout of the Nova 2 generation and a broadened software ecosystem around Nova Forge and Nova Act. These solutions industrialize the AI model lifecycle, from development to production, fully leveraging the computational and memory capabilities of the Trainium3 platform. Nova Forge, in particular, automates model optimization and accelerates production timelines, while Nova Act focuses on governance, analysis, and performance evaluation in operational contexts.

The new “Frontiers” AI agents embody AWS’s shift towards intelligent automation in software development. Leveraging Trainium3 and Nova, these agents assist developers in code generation, performance analysis, and anomaly detection, aiming to shorten development cycles while ensuring the reliability of large-scale AI applications.

AWS AI Factories: Industrializing and Democratizing AI

The introduction of AWS AI Factories presents an integrated offering to transform data centers into AI factories by combining the power of Trainium3 UltraServers, the Nova software suite, and automation tools to create a unified operational environment. This industrialization aims to lower entry barriers for businesses while optimizing costs, flexibility, and energy consumption. According to Garman, this new generation of infrastructure paves the way for an accessible, industrial-grade AI that is user-driven, striking a balance between technological performance and operational mastery.

The competition surrounding specialized chips for AI training and inference transcends mere innovation; it represents an industrial battle for control over current strategic resources. The chronic shortage of dedicated processors—whether GPU, NPU, or in-house solutions like Trainium3—weakens the entire AI value chain, from model designers to cloud platform providers. This tension has sparked intense demand among players whose business models hinge on the ability to absorb or meet skyrocketing demand: software vendors, tech giants, service companies, and even innovative startups are competing to secure access to these infrastructures. This race is far from trivial, particularly as most AI innovations rely on seamless, massive access to computing power.

The shortage drives consolidation, vertical integration, and the bundling of offerings, as each player seeks to bolster their technical capabilities. It is the principal engine of competition among major players, each striving to secure access to the essential resource that will dictate the future growth of AI.

Scroll to Top