OpenAI has co-developed a custom 700W AI chip named Jalapeño with Broadcom. Benchmark claims suggest the chip significantly outperforms Nvidia's flagship GPUs in throughput per kilowatt and offers much lower latency for AI inference tasks.