OpenAI shares Jalapeño test results: more intelligence per watt, higher throughput and lower latency in one custom inference architecture.