Get the latest on AWS AI Chips from re:Invent 2025
This article highlights the key announcements and innovations showcased at re:Invent 2025 across Trainium and Neuron, alongside our customers.
Authored by Armin Agha-Ebrahim, GTM Specialist, AWS AI Chips
This year at re:Invent, AWS raised the bar for purpose-built AI accelerators with a series of major announcements. We unveiled AWS Trainium3, our next-generation AI training accelerator, delivering significant gains in compute performance, memory bandwidth, and energy efficiency to power the next wave of frontier-scale model training on AWS.Alongside Trainium3, we introduced our second generation Trainium3 UltraServer and meaningful updates to the Neuron SDK, including native PyTorch support. With native PyTorch on Trainium, developers can now use a seamless, fully pythonic workflow to build, train, and optimize large language models—without changing how they write PyTorch code. Learn more about native PyTorch on Trainium here.Customers are already seeing the impact. Decart.ai showcased the power of Trainium3 for training frontier-scale video models, while Splash Music is using Trainium to empower a new generation of creators—achieving nearly 50% faster training times and 54% lower training costs compared to their previous infrastructure.This year’s keynotes and breakout sessions demonstrated how Trainium and Inferentia continue to enable faster, more cost-efficient generative AI experiences at scale. Attendees explored hands-on labs and technical deep dives covering best practices for training and inference, model optimization techniques, and real-world examples of how customers are building sustainable, production-grade AI systems in the cloud. learn more about whats new on neuron here: https://awsdocs-neuron.readthedocs-hosted.com/en/latest/about-neuron/whats-new.html Missed re:Invent 2025? We’ve got you covered. Below, you can explore all the keynote highlights and breakout session recordings for AWS Trainium and Inferentia. Dive in and enjoy!
Keynote Highlights:
AWS re:Invent 2025 Keynote with CEO Matt Garman
This year, Matt took the stage to unveil Trainium3 and the all-new Trn3 UltraServers, now generally available. He highlighted the technical innovations across the chip, system, and multi-server fabric that enable Trn3 UltraServers to deliver up to 4x more compute performance, 4x greater memory bandwidth, and roughly 40% better energy efficiency compared to Trainium2. These advances position Trainium3 as a unified AI platform for both training and inference, designed to efficiently scale across frontier AI models, multimodal workloads, and agentic systems—delivering scale, performance, and cost efficiency for customers.Matt also highlighted the scale and adoption of Trainium, noting that over one million Trainium chips are now running in production across AWS. He emphasized that the majority of inference on Amazon Bedrock is powered by Trainium, underscoring the platform’s central role in AWS’s AI ecosystem. Looking ahead, Matt shared AWS’s roadmap for purpose-built AI chips, confirming that Trainium4 is already in development with 6x performance, 4x memory bandwidth, 2x Memory capacity, and including support for NVIDIA NVLink interconnects, designed to take next-generation model training even further.
AWS re:inent2025 - Keynote with Peter DeSantis and Dave Brown
In this year’s keynote, Peter took the stage to unveil the next chapter of AWS’s purpose-built AI chips, delivering a technical deep dive into Trainium3 and the all-new Trn3 UltraServers. He highlighted architectural breakthroughs across the chip, system, and multi-server fabric layers, enabling Trainium3 to deliver massive gains in compute density, memory bandwidth, and energy efficiency for frontier-scale model training. These innovations are built to meet the rapidly growing demands of customers training ever-larger multimodal and LLM workloads.Peter was joined by Decart.ai CEO Dean, who shared how his team is leveraging Trainium3 to power their next-generation real-time vision models. They demonstrated how Trainium3 is already pushing the boundaries of model size, capability, and throughput, achieving 80% utilization efficiency and 4x the frames per second compared to GPUs.Peter also highlighted the addition of native PyTorch support for Trainium, offering developers a seamless, pythonic workflow for training LLMs with minimal code changes. He introduced the Neuron Explorer, a unified suite of developer tools for model bring-up, debugging, profiling, and optimization. With CLI, UI, and VSCode integration, Neuron Explorer provides ML engineers with a streamlined, end-to-end workflow to achieve peak performance on Trainium.To help developers get the most out of Trainium, we introduced the Neuron Explorer, a unified suite of tools for the entire ML development journey—from model bring-up to profiling, debugging, and optimization. With CLI, UI, and VSCode integration, Neuron Explorer helps engineers iterate faster, maintain context, and deliver high-performance, cost-efficient deep learning models. Explore Neuron Explorer here. Breakout Highlights
AWS re:Invent 2025 - End-to-end foundation model lifecycle on AWS Trainium (AIM351)
**Discover how Splash Music scaled AI on AWS Trainium3.**This session spotlights AWS Trainium3, delivering major gains in compute, memory, and bandwidth—unlocking up to 5.4× more concurrent users and 6x more tokens per user for agentic workloads. You’ll see how open-weight models like GPT-OSS, DeepSeek, and Kimi K2 run cost-effectively on Trainium, plus a live demo showing a 50% attention latency reduction using Neuron tools.The highlight: Splash Music’s CTO shares how they cut training costs by 54%, scaled from 4 to 100 nodes, and reached 750M streams in under a year with a custom model on Trainium3.Watch the full session to see the demos, tooling, and real-world results in action.
AWS re:Invent 2025 - Break through AI performance and cost barriers with AWS Trainium (AIM201)
At AIM 201, AWS highlighted how customers like Decart AI are using Trainium 3 to enable real-time, interactive AI that wasn’t previously possible. Decart AI showcased live video transformations with 2× the performance of Trainium 2, while Poolside shared how Trainium delivers efficient, scalable inference—not just training. With major gains in compute, memory bandwidth, and energy efficiency, Trainium 3 is built for agentic and domain-specific AI. Watch the session to see how customers are pushing the limits of real-time AI on Trainium.
AWS re:Invent 2025 - AWS Trn3 UltraServers: Power next-generation enterprise AI performance(AIM3335)
At AIM3335, AWS and Anthropic went deep on how Trainium 3 is purpose-built for the next generation of agentic, reasoning, and large-scale inference workloads. AWS leaders explained how Trainium 3 balances compute, memory, and interconnect to deliver sustained—not just peak—performance, while scaling rapidly to massive clusters like Project Rainier. Anthropic shared how Claude models are already running at scale on Trainium, and walked through real kernel-level optimizations that unlock faster inference, lower latency, and higher efficiency—showing how Trainium 3 pushes utilization past 90% for real workloads. If you want to see how silicon, systems, and software come together to power production AI at scale, watch the full session
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article