OpenAI unveils Jalapeño, its first custom AI chip built with Broadcom to cut Nvidia reliance
The San Francisco company partnered with Broadcom to design an ASIC tuned for large language model inference, aiming to lower costs and reduce dependence on Nvidia GPUs.
What Jalapeño is
OpenAI’s first custom chip is an application-specific integrated circuit (ASIC) built for AI inference, the real-time processing that answers user queries on ChatGPT and Codex. Unlike Nvidia’s general-purpose GPUs, the Jalapeño Intelligence Processor (JIP) is tuned solely for large language model workloads. OpenAI president Greg Brockman said the move lets the company “serve more intelligence with greater efficiency and keep pushing advanced AI toward broader access.”
Performance and claims
Broadcom CEO Hock Tan told Reuters that Jalapeño matches the performance of Nvidia’s Blackwell chips and Google’s tensor processing units. OpenAI says early testing shows performance per watt “substantially better than current state-of-the-art,” though a full technical report is still months away. The chip is already running engineering samples of GPT-5.3-Codex-Spark in its labs at target power and performance levels.
A nine-month sprint, aided by AI
OpenAI and Broadcom moved from first design to tape-out in nine months, a pace the companies call the fastest ever for a high-performance ASIC. OpenAI’s own models helped accelerate parts of the design process. The wafer was manufactured by TSMC, and Celestica will build the surrounding server systems and racks. OpenAI plans to deploy Jalapeño in data centers by the end of 2026, with broader volume expected next year.
- Reuters reports OpenAI exploring in-house chip development
- OpenAI and Broadcom announce partnership for custom AI chips
- Jalapeño unveiled; engineering samples running in OpenAI labs
- Initial deployment in data centres planned
- Broader volume rollout begins
Breaking the Nvidia dependency
The chip is the clearest sign yet that the AI lab, one of Nvidia’s biggest customers, wants to diversify its compute supply. OpenAI already works with Amazon, AMD and Cerebras, but Jalapeño is its first wholly-owned silicon. “This is just the beginning of a multi-generation roadmap,” Tan said, with future chips likely to follow. The companies have a target of ten gigawatts of compute power by 2029.
Market context
Other tech giants, Amazon, Google, Meta and Microsoft, have also developed custom chips for AI. Anthropic is weighing its own design, sources told Reuters. Jalapeño’s arrival intensifies competition in the AI hardware race, where inference efficiency directly impacts product cost and user experience.
