AWS Trainium

AWS Trainium is a family of AI accelerator chips designed by Amazon Web Services (AWS) specifically to train and run large artificial intelligence models. Built by AWS's Annapurna Labs silicon division and offered to customers only through AWS cloud services rather than as standalone hardware, Trainium is marketed on cost and energy efficiency — AWS describes the goal as the best "token economics" for AI workloads at scale — and is the centrepiece of the company's effort to reduce dependence on NVIDIA graphics processors for frontier-scale training.^[1][2]

History

AWS began designing custom silicon for machine learning in the 2010s, releasing the Inferentia chip for AI inference in 2018. Trainium, its training chip line, was announced in December 2020, with the first Trn1 instances becoming available in 2022 under the Neuron software development kit. The second generation, Trainium2, arrived at general availability in December 2024 alongside the Trn2 instances, and was adopted for Project Rainier, a purpose-built cluster in New Carlisle, Indiana, constructed for the AI company Anthropic at a reported cost of about US$11 billion. Rainier came fully online in late 2025 with close to half a million Trainium2 chips — among the largest AI training clusters outside the GPU ecosystem — and was expanded toward one million chips, with Anthropic using it to train and serve its Claude models.^[3][4][5]

The third generation, Trainium3, was released to general availability in December 2025 alongside EC2 Trn3 UltraServers. It is AWS's first chip built on a 3-nanometre process, manufactured by TSMC, and each chip delivers 2.52 petaflops of FP8 compute with 144 gigabytes of HBM3e memory at 4.9 terabytes per second of bandwidth. A single Trn3 UltraServer links 144 chips through a new all-to-all interconnect, NeuronSwitch-v1, for an aggregate 362 FP8 petaflops and 20.7 terabytes of memory; AWS quotes up to 4.4 times the compute, about 4 times better energy efficiency and almost 4 times the memory bandwidth of the previous Trn2 UltraServer generation, with EC2 UltraClusters 3.0 scaling to hundreds of thousands of chips.^[1][2][6]

Key Concepts and Technology

Trainium chips pair a compute engine optimised for dense matrix mathematics with high-bandwidth memory and a dedicated interconnect fabric; AWS says the largest gains emerge at the system level, where hundreds of chips cooperate on a single training run. Trainium3 supports FP32, BF16 and low-precision MXFP8 and MXFP4 formats with hardware for structured sparsity and micro-scaling, features aimed at the mixture-of-experts and long-context architectures common in frontier models. Software support comes through the Neuron SDK, which integrates with PyTorch and JAX so that teams can port existing training code with relatively few changes. On Amazon Bedrock, AWS's managed model service, Trainium3 is offered as the fastest accelerator option, which AWS says delivers up to three times the performance of Trainium2 with more than five times the output tokens per megawatt of power — an efficiency metric that matters as data-centre energy use becomes a constraint on AI growth. Because the chips are not sold separately, customers access them through EC2 instances, Bedrock and SageMaker rather than purchasing hardware.^[1][2][3][6]

Applications and Impact

Trainium's principal customer is Anthropic, which has committed to training frontier Claude models on AWS silicon and, in an expanded partnership, secured capacity of up to five gigawatts for training and deployment, including roughly one gigawatt of Trainium2 and Trainium3 capacity due online by the end of 2026. AWS also runs internal workloads on the chips — the company has cited a 30 per cent reduction in training costs for its Amazon Search models — and positions Trainium as part of a broader custom-silicon trend alongside Google's TPUs, Microsoft's Maia and Meta's MTIA. The strategic logic is economic and political: custom accelerators reduce exposure to GPU supply and pricing, and AWS argues that cheaper training and inference expand what customers can attempt. Analysts treat the family as credible for large-scale training while noting that the wider software ecosystem and developer familiarity still favour NVIDIA's CUDA platform.^[2][3][5][6]

>See Also

• NVIDIA
• Tensor Processing Unit
• GPU Cluster
• High Bandwidth Memory
• Amazon Bedrock
• Sovereign AI