NeuralStackAmazon Web Services detailed the Trainium3 rollout timeline on July 10. General availability begins September 15 in US-East-1 and expands to all US regions by October 31. The chips provide 2.8 times the training performance of Trainium2 while maintaining the same 700-watt TDP envelope.
Each Trn3 instance packs 16 Trainium3 accelerators connected via a new 3.2 Tbps interconnect fabric. AWS reports early benchmark results showing 41 percent cost reduction on 70 billion parameter model training compared with equivalent GPU clusters.
Development of Trainium3 began in 2024 following internal testing of Trainium2 limitations at scale. The design incorporates custom networking silicon developed jointly with Broadcom.
AWS continues to offer NVIDIA H200 and Blackwell instances alongside its custom silicon. The company claims over 200,000 Trainium2 chips are already in production use by customers.
Why this matters
Trainium3 strengthens AWS's ability to offer lower-cost training alternatives to GPU-heavy competitors. Customers gain more predictable pricing and reduced dependency on NVIDIA allocation cycles.
The move pressures Google and Microsoft to accelerate their own custom silicon roadmaps. AWS training revenue share is expected to rise from 12 percent to 19 percent by end of 2027.
Long-term, custom ASICs could reshape cloud economics and reduce overall industry reliance on merchant GPUs. Amazon's vertical integration strategy gains further validation with this launch.