The chip is designed for inference - the stage of AI computing in which a trained model generates responses for users - rather than for training new models, sources said.