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 say