OpenAI and Broadcom unveil "Jalapeño," a custom chip built for LLM inference