OpenAI has unveiled its first custom-designed artificial intelligence chip, named Jalapeño, marking a significant step in the company’s efforts to strengthen its computing infrastructure and lessen its dependence on Nvidia’s graphics processing units (GPUs).
The chip was developed in collaboration with semiconductor firm Broadcom and is specifically designed for AI inference — the process through which AI systems generate responses to user queries.
As demand for increasingly powerful AI models continues to surge, companies such as OpenAI and Anthropic are facing mounting pressure to secure sufficient computing resources to support their services.
According to Broadcom Chief Executive Hock Tan, the Jalapeño chip delivers performance comparable to Nvidia’s Blackwell processors and Google’s Tensor Processing Units (TPUs). OpenAI’s Head of Hardware, Richard Ho, said the processor has been engineered to efficiently support large language models and is expected to remain effective across future generations of AI systems.
OpenAI plans to begin deploying the chip before the end of the year as part of a broader, multi-generation strategy to develop its own AI hardware. The company revealed that samples of Jalapeño are already operating in its laboratories, where they have successfully achieved targeted power and performance levels while running the GPT-5.3-Codex-Spark model.
Canadian electronics manufacturer Celestica will produce the server systems that will house the chips. Both the processors and servers are intended solely for OpenAI’s internal use.
The company said the chip design was completed in approximately nine months before being sent to Taiwan Semiconductor Manufacturing Company (TSMC) for production. OpenAI noted that artificial intelligence tools helped accelerate certain aspects of the design process.
The move mirrors similar efforts by major technology firms including Meta, Amazon and Google, all of which have partnered with companies such as Broadcom and Marvell to develop custom AI chips. Reports have also suggested that Anthropic is exploring the possibility of creating its own processor.
Despite growing demand for custom AI hardware, Tan acknowledged that profit margins remain under pressure due to the high memory requirements of AI chips.
These processors rely heavily on high-bandwidth memory, supplied primarily by South Korean manufacturers SK Hynix and Samsung Electronics, making them more expensive to produce than some of Broadcom’s other products, such as networking switches.
