[AINews] Hot Chips: OpenAI’s Jalapeño, Cerebras CS-5, Groq 3 LPX, Apple M6
OpenAI's Jalapeño chip claims 1.5–1.9x better performance per watt than NVIDIA GB200/GB300.
OpenAI debuted benchmark details for Jalapeño, its custom inference chip, at Hot Chips 37, claiming it outperforms NVIDIA's GB200/GB300 Blackwell systems by 1.5–1.9x on performance per watt. Unlike OpenAI's earlier Broadcom ASIC partnership, Jalapeño is a full custom chip, signaling a major vertical integration push. This marks a potential structural shift in the inference hardware stack, with performance-per-watt now the competitive axis rather than raw throughput.