AWS launches Trainium3 instances at re:Invent with 42% better price-performance as Anthropic expands usage
AWS AI demand converts into backlog as custom silicon crosses a third of new capacity
At re:Invent in Las Vegas, AWS launched general availability of Trainium3 instances, claiming up to 42 percent better price-performance than the prior generation for training and inference. The company said Anthropic had expanded its Trainium footprint to cover a majority of its model training and that a new UltraCluster deployment scaled to more than 100,000 chips. AWS also announced a 30 percent price reduction for certain inference instance families using custom silicon. Jassy said custom silicon now accounts for more than 30 percent of new AI compute capacity added this year. The announcement included a multi-year agreement with a sovereign AI program in the Middle East and a new offering that lets enterprises reserve capacity on Trainium clusters. Nvidia remains the default option on most AWS AI instance families, and the keynote avoided direct comparisons on training at frontier scale.
Custom silicon crosses 30 percent of new AI capacity with a 42 percent price-performance claim and an Anthropic reference. The avoided comparison is training at frontier scale.
The keynote compared generations, not competitors. Anthropic expanding to a majority on Trainium is the loudest reference possible, and the sovereign deal is the same playbook as last year.
- Trainium3 delivers up to 42 percent better price-performance than the prior generation
- Custom silicon accounts for more than 30 percent of new AI compute capacity added this year
- Inference instance prices cut 30 percent on custom silicon
- UltraCluster scales to more than 100,000 chips