OpenAI Unveils Jalapeño, Its First Custom AI Chip
3 min readOpenAI just stepped into the silicon business. On June 24, the company and chipmaker Broadcom unveiled Jalapeño, OpenAI’s first custom AI chip, a processor built specifically to run large language models faster and at lower cost. It is the clearest sign yet that OpenAI wants to control more of the hardware its models depend on.
Why OpenAI Built Its Own AI Chip
Until now, OpenAI has run almost entirely on Nvidia graphics processors, the same chips nearly every major AI lab competes to buy. That dependence is costly, and supply stays tight. Rivals moved early to escape it: Google designs its own TPUs, Amazon builds Trainium, and Meta has its own accelerators. With Jalapeño, OpenAI joins the club of AI giants designing chips tuned to their own software.
What Makes Jalapeño Different
OpenAI says it designed Jalapeño from scratch around how language models actually behave during inference, the stage where a trained model answers user prompts. It is a purpose-built inference accelerator, not a repurposed training processor or a general-purpose part. Broadcom handled the silicon implementation and networking, while manufacturer Celestica contributed board, rack, and system integration.
The most striking claim is speed of execution. OpenAI and Broadcom took Jalapeño from initial design to manufacturing tape-out in just nine months, which they call possibly the fastest development cycle ever for a high-performance advanced chip. Early testing, the company says, shows performance per watt substantially better than the current state of the art.
Why It Matters
Inference is where the real money goes. Every chatbot reply, coding suggestion, and image request runs through inference hardware, and at OpenAI’s scale those costs are enormous. A chip tuned for that exact workload could lower the cost of serving models and reduce reliance on a single supplier. It also boosts Broadcom, which has quietly become a central player in the custom AI chip race.
Jalapeño is only the first step in what OpenAI calls a multi-generation compute platform. Initial deployment is planned by the end of 2026 through data center partners including Microsoft, with more chips to follow. The open question is whether a first-generation design can deliver in production, where Nvidia still sets the bar.
For everyday users, none of this changes the chatbot experience overnight. But cheaper, more efficient inference is what makes ambitious AI features affordable at scale, and that is the real prize OpenAI is chasing.