OpenAI's Jalapeño reportedly beats Nvidia Blackwell and Rubin in inference
openai's first custom inference chip 'jalapeño' claims 1.5-1.9x better efficiency and 3.6x lower latency than nvidia blackwell and rubin, signaling a potential end to nvidia's 'cuda moat.'
openai revealed 'jalapeño,' its first in-house inference chip, at the hot chips conference, claiming it surpasses nvidia's blackwell and rubin in throughput per watt and latency. built with broadcom in just nine months using openai's own models for design and optimization, the chip delivered 1,400 tokens per second on gpt-oss 120b and 700 tps on deepseek r1 670b in internal tests. while the chip shows significant performance leads without speculative decoding, it currently exists only as an engineering sample, whereas rubin systems are already shipping to customers.