AMD has acquired Taalas to boost AI inference performance by etching machine learning models directly into silicon. Early demonstrations of these model-specific integrated circuits reveal processing speeds of up to 17,000 tokens per second.