Amazon shifts more inference onto Trainium and keeps Nvidia for frontier training
AWS is treating custom silicon as the house brand for high-volume serving and Nvidia as the scarce input for the next model generation.
Technology Editor
Close-up of a semiconductor circuit board
SEATTLE — AWS is treating custom silicon as the house brand for high-volume serving and Nvidia as the scarce input for the next model generation.
The split is showing up in customer quotes: Trainium for predictable token factories, Blackwell-class GPUs for the runs that still miss their loss curves on anything else.
James Whitaker
Technology Editor
Reports on semiconductors, cloud infrastructure, and the industrial politics of AI.
