The new chip is designed to run LLMs that support reasoning, which typically require more compute to generate each response.