OpenAI Jalapeño AI chip delivers strong perf per watt for LLM inference after nine months of design with Broadcom, expanding full-stack infrastructure.