OpenAI Unveils Jalapeño: LLM-Designed AI Chip With 3.6x Lower Latency Than Nvidia GB300

OpenAI has unveiled Jalapeño, its own AI accelerator. A team of fewer than 100 engineers built it in 20 months. Inference latency is up to 3.6 times lower than the Nvidia GB300. OpenAI's own language models were among the tools used, helping to optimize the circuit designs.

The most telling detail: LLMs raised the performance of a test compute kernel from 0.31% to 88.94% of the theoretical maximum in about 40 hours. This is not "chips designing themselves." It is engineers with a tool that handles the most labor-intensive part of parameter tuning.