Amazon Web Services (AWS) has announced a powerful addition to its AI training chip lineup, the next-generation Trainium3 chip. At re:Invent 2025, the company unveiled not only the Trainium3 processor but also the Trainium3 UltraServer system, designed for large-scale infrastructure. AWS also officially confirmed that its next step, the Trainium4 processor, is currently in development.
Trainium3 chip unveiled
The newly introduced Trainium3 delivers a fourfold increase in both AI training and inference performance, and a fourfold increase in memory capacity, compared to its predecessor, the Trainium2. Each UltraServer contains 144 Trainium3 chips.
Shiftdelete.netAWS offers customers the ability to connect thousands of these systems. At maximum scale, up to 1 million Trainium3 chips can be used in a single deployment. This represents a deployment size ten times larger than the previous generation.
Another key feature of Trainium3 is its energy efficiency. According to AWS, these new systems offer higher throughput while consuming approximately 40 percent less energy.
Companies such as Anthropic, Japanese LLM startup Karakuri, SplashMusic, and Decart are among the early adopters of Trainium3. These customers report faster inference, faster iterations, and lower billable processing times in their models.
AWS has also shared the first details about its next-generation chip, Trainium4. While no official release date has been announced yet, Trainium4 is expected to deliver a significant leap in performance. Its most critical feature is its support for Nvidia NVLink Fusion technology.
This support demonstrates that AWS positions its hardware not as a competitor to Nvidia solutions, but as a complementary element in a hybrid AI ecosystem. This allows companies using CUDA-based applications to easily integrate AWS’s specialized hardware into their systems without having to change their software infrastructure.
The partnership encompasses connectivity technology, data center infrastructure, open model support, and rack-scale AI system deployment. NVLink Fusion support will be incorporated into future AWS platforms, including Trainium4, Graviton CPUs, and Nitro virtualization hardware. The company did not provide pricing or availability information for the new chips.
