AWS Trainium2 (Trn2 / Trn2 UltraServer)

Amazon's second-generation custom training accelerator from Annapurna Labs, generally available December 2024. Each chip has 8 NeuronCores and 96 GiB of HBM with 2.9 TB/s bandwidth, delivering up to 1.3 PFLOPS dense FP8 / 5.2 PFLOPS sparse FP8.

A Trn2 instance bundles 16 chips for 20.8 PFLOPS dense FP8; the Trn2 UltraServer connects 64 chips via NeuronLink for up to 83 PFLOPS dense FP8. Anthropic's Project Rainier cluster uses "hundreds of thousands" of Trainium2 chips and reportedly provides 5x more exaflops than what was used to train Anthropic's then-current models. Trainium3 (3 nm) was previewed for late 2025.

Trainium2 already serves latency-optimized inference for Claude 3.5 Haiku and Llama 3.1 405B on Amazon Bedrock. The hardware track underpins Amazon's long-term frontier-model strategy — tightly coupled with the Anthropic partnership and the AGI org's training plans.