OpenAI and Broadcom Unveil Jalapeño, Its First Custom AI Inference Chip

OpenAI and Broadcom have unveiled a custom AI inference chip named Jalapeño, marking OpenAI's first foray into custom silicon. The chip is specifically optimized for large language model inference workloads and aims to reduce OpenAI's dependence on Nvidia GPUs.

Built as a massive reticle-sized ASIC, the Jalapeño chip was developed by Broadcom in an exceptionally fast nine-month cycle. According to multiple reports, the chip delivers significant improvements in inference cost and energy efficiency compared to current Nvidia GPU solutions, with potential cost reductions of approximately 50%.

This launch represents a strategic shift for OpenAI — from a pure model company to a full-stack AI infrastructure provider. By controlling its own silicon, OpenAI can directly optimize the compute layer for its specific workloads, reducing operating costs and improving inference efficiency over the long term.

Notably, OpenAI used its own models to accelerate the design and verification process for the Jalapeño chip, demonstrating a novel AI-designing-AI paradigm that dramatically shortened the traditional chip development timeline.

Analysts note that while Jalapeño will not immediately replace Nvidia GPUs for training workloads, its breakthrough in inference signals a shifting landscape in AI infrastructure. Broadcom's position in the custom chip market is further strengthened by this partnership.

With Jalapeño, OpenAI joins the ranks of Google (TPU), Amazon (Trainium/Inferentia), and Microsoft (Maia) in developing custom AI silicon, underscoring a growing industry trend where leading AI companies increasingly view chip customization as a core competitive advantage.