OpenAI Uses AI to Tape Out Custom Jalapeño Chip in 9 Months
OpenAI CFO Sarah Friar announced that the company used internal AI models to design and tape out its custom Jalapeño inference chip in just nine months. Additionally, OpenAI slashed the price of its lightweight Luna model by 80%, driving a tenfold increase in adoption while positioning its API cost below self-hosted open-weight alternatives. This highlights OpenAI's aggressive push toward vertical integration into custom silicon while undercutting open-weight models hosted on cloud infrastructure. It also signals a rapid shift in commercial strategy, as enterprise revenue now accounts for half of OpenAI's total income alongside experiments with outcome-based pricing. Reaching the tape-out stage signifies that the Jalapeño chip design is finalized and sent to a semiconductor foundry for production. OpenAI also reported a 32% enterprise revenue jump between June and July, while its developer tool Codex reached 25 million users.
## BACKGROUND
OpenAI developed Jalapeño as a custom inference chip in partnership with Broadcom to optimize LLM performance and reduce reliance on third-party hardware. Meanwhile, Luna is the fastest and lowest-cost tier in OpenAI's GPT-5.6 model lineup, designed to compete against open-weight models like Zhipu AI's GLM series that require costly cloud server hosting.