AWS launches Trainium3, expands Nova to reduce Nvidia reliance

AWS unveiled Trainium3 at re:Invent 2025, the third generation of its custom artificial intelligence chip, alongside an expansion of the Amazon Nova model family. The move reinforces a strategy shared by other major cloud providers: designing their own silicon to contain the cost of running AI and reduce their reliance on Nvidia GPUs.

The most significant infrastructure announcement is the Amazon EC2 Trn3 UltraServer. Each system can integrate up to 144 Trainium3 chips, built on a 3-nanometer process—a technology that allows more transistors to be packed into the chip and, in principle, delivers higher performance with lower energy consumption.

More capacity per server, less training time

AWS says its Trainium3-powered UltraServers offer up to 4.4 times more computing capacity and four times greater energy efficiency than the Trainium2 generation. It also reports a threefold improvement in per-chip performance and up to a fourfold reduction in latency—the time a system takes to respond to a request.

These comparisons matter because training advanced models involves more than simply buying fast chips. Thousands of processors must be connected, enormous volumes of data moved between them, and power consumption kept within manageable limits. An UltraServer brings together the hardware and networking needed for those chips to work as a single system.

AWS says the improvement can cut training times from months to weeks for some projects. It is an ambitious claim that depends on the model, the software, and each customer’s existing infrastructure, but it illustrates the main appeal of custom accelerators: lowering the total cost of building and serving models, not just improving a speed benchmark.

Anthropic is among Trainium’s customers, alongside companies such as Karakuri, Metagenomi, Ricoh, and Splash Music. Amazon Bedrock, AWS’s managed service for using AI models, is already running production workloads on Trainium3, according to the company.

Nova 2 extends the bet into software

Amazon has also expanded the Nova family with four new models focused on reasoning, multimodal processing, conversation, coding, and agentic tasks. A multimodal model can work with more than one type of information, such as text and images, while an agent is a system capable of chaining steps together to complete a task.

The most notable release is Nova Forge, an “open training” offering that lets organizations start from model checkpoints taken at intermediate stages of an Amazon model’s training. They can then combine their own data with datasets curated by AWS, rather than being limited to fine-tuning a fully closed model at the end of the process.

The approach aims to address a common tension for businesses: general-purpose models are useful, but they do not always understand an organization’s internal language, processes, or data. Access to earlier training stages provides more room for customization, though it also requires expertise, well-governed data, and computing resources.

AWS has also announced Nova Act to automate tasks in web interfaces. The company says its early customers have achieved 90% reliability in browser workflows. That figure should be interpreted cautiously: web automation depends heavily on how many steps a process involves and whether the interface changes. Even so, the goal is clear: for AI not just to draft or respond, but to operate digital tools.

An alternative, not a break with Nvidia

Trainium does not turn AWS into a competitor that can do without Nvidia. The company itself has announced AWS AI Factories, AI infrastructure installed in customers’ data centers that will combine Nvidia GPUs, Trainium chips, and AWS services. For companies with data-sovereignty requirements or existing space and power at their facilities, that combination may be more practical than moving everything to the public cloud.

Competition is therefore shifting from individual models to the full stack: chips, networking, data centers, development tools, and models. Nvidia remains central because of the maturity of its GPUs and software ecosystem, but AWS wants customers to be able to choose Trainium when the savings and availability justify the effort of adapting their workloads.

For Bedrock users and teams already working on AWS, the immediate consequence is a wider range of infrastructure and models within the same provider. The real test will be whether Trainium3 can deliver those cost and performance advantages in everyday applications, not just in re:Invent demonstrations.