According to OpenAI and Reuters on June 24, Jalapeño is an AI accelerator specifically optimized for large language model (LLM) inference. Inference refers to the process where a trained AI model responds to user queries or performs tasks. OpenAI stated that this chip is designed not as a general-purpose semiconductor but tailored to its service processing needs.
OpenAI described Jalapeño as being built to efficiently run current and next-generation LLMs. Initial tests have shown potential for performance improvements relative to power consumption, with specific metrics to be disclosed in a future technical report.
Broadcom is responsible for the chip implementation and networking technology, while Celestica is constructing the server system. Manufacturing is handled by Taiwan's TSMC. OpenAI noted that the design and manufacturing process took approximately nine months, during which it utilized its AI models to expedite certain design and optimization tasks.
Jalapeño is expected to be deployed within OpenAI's internal infrastructure by the end of this year. Reuters reported that the chip and server system will not be sold to external customers but will be used directly for OpenAI's service operations.
OpenAI confirmed that it is currently running prototypes in its lab, executing machine learning tasks at target power and performance levels using the GPT-5.3-Codex-Spark model.
Hock Tan, CEO of Broadcom, stated in an interview with Reuters that Jalapeño could compete with Nvidia's Blackwell chip and Google's tensor processing units (TPUs).
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.