Amazon drops Trainium3 chip and plans to play nice with Nvidia

Big tech continues to battle over power grids and AI accelerators, but AWS is doubling down on its home-cooked silicon.

On Tuesday at AWS re:Invent 2025, Amazon unveiled Trainium3, its third-generation AI chip, alongside a beastly new server system. The company also surprised the crowd by teasing its next-gen Trainium4, which will feature a very unexpected partnership with Nvidia.

A 3nm leap packed into 144-chip servers

Trainium3 is built on a cutting-edge 3-nanometer process. AWS claims the new silicon delivers more than four times the speed and four times the memory of Trainium2, covering both intensive AI training and high-demand inference workloads.

The real muscle comes from the new Trainium3 UltraServer. Each server rack packs 144 Trainium3 chips, and using Amazon homegrown networking tech, AWS can link thousands of them together.

That means a single AI application can scale up to 1 million Trainium3 chips simultaneously, a massive 10x jump over the previous generation.

Power efficiency also took center stage. With AI data centers devouring gigawatts of electricity, AWS designed the new system to consume 40% less energy than its predecessor.

Key enterprise customers are already on board. Amazon noted that Anthropic, Japanese LLM maker Karakuri, SplashMusic, and Decart have been running on Trainium3 hardware and seeing significant drops in inference costs.

Trainium4 extends an olive branch to Nvidia

AWS also gave attendees a sneak peek at Trainium4, which is currently in development.

The head-turner here is that Trainium4 will support Nvidia NVLink Fusion high-speed interconnect technology.

Nvidia CUDA remains the de facto software standard across the AI landscape, making direct hardware switches painful for developers. By adopting NVLink Fusion, AWS systems powered by Trainium4 will be able to interoperate directly with Nvidia GPUs while staying inside Amazon lower-cost server racks.

Amazon did not provide a release date for Trainium4, though we will likely hear more details by next year conference.

Why it matters

For AI builders and enterprises, compute and electricity bills represent existential expenses. Trainium3 delivers a massive bump in capacity alongside a 40% reduction in energy usage, offering teams like Anthropic a viable way to trim inference overhead right now without sacrificing performance.

Across the industry, cloud giants are realizing that raw silicon is only half the battle; scaling and power delivery matter just as much. Linking up to one million chips in a single cluster shows AWS is prepared for the next frontier of foundational model training.

We think teasing NVLink Fusion support for Trainium4 is a particularly sharp strategic maneuver. Instead of pretending it can kill Nvidia CUDA overnight, Amazon is building a bridge. By allowing customers to run Nvidia GPUs side by side with custom silicon in cheaper proprietary racks, AWS makes its cloud the most practical landing zone for enterprise AI workloads.