AWS makes Trainium2 generally available, adds Bedrock and EKS support for enterprise AI training

Amazon Web Services made its Trainium2 accelerators generally available on Tuesday, opening access in select regions and integrating the chips with Amazon Bedrock and Amazon EKS. The company said Trainium2 delivers up to four times higher training performance and twice the energy efficiency compared with Trainium, citing internal tests. AWS positioned the rollout as a cost control move for customers training large language models as demand for Nvidia systems remains high.

The timing matters because enterprise buyers are locking 2027 training budgets before year end, and capacity on Nvidia H100 and Blackwell class systems continues to sell out. Nvidia still accounts for more than 80 percent of data center AI compute shipments by revenue, according to Omdia, which has pushed cloud providers to highlight alternatives.

AWS said customers can train and fine tune models with PyTorch and JAX using the Neuron SDK, with managed orchestration through EKS and SageMaker. Bedrock now supports Trainium2 for model customization on select first party models, and AWS said third party models will follow as vendors certify Neuron support. Swami Sivasubramanian, vice president for data and AI at AWS, said the chips target repeatable pretraining and fine tuning jobs where customers want predictable throughput and lower power draw.

AWS did not publish list prices for Trainium2 instances, but said the hardware is eligible for Compute Savings Plans and reserved capacity, which can provide meaningful discounts for one year and three year commitments. The company said early users have seen double digit percentage reductions in total training spend compared with comparable GPU clusters for mid sized transformer workloads, though it did not name customers.

In benchmarking, AWS referenced prior claims from re Invent 2023 that Trainium2 would deliver four times the performance of Trainium on LLM training tasks. Nvidia continued to lead recent MLPerf training and inference rounds across categories, and Google has expanded Cloud TPU v5e and v5p availability. That competitive context underscores how buyers now weigh a mixed hardware strategy across GPUs and custom silicon.

For enterprises, the Trainium2 release strengthens the case for heterogeneous AI stacks that balance performance, availability, and regional control. It also signals that cloud providers will lean on in house silicon, alongside Nvidia and AMD, to keep unit economics in check for large scale model training and continual fine tuning.