EVERYTHING
According to The-decoder, OpenAI and Broadcom have joined forces to launch Jalapeño, which OpenAI describes as its first Intelligence Processor. This custom accelerator is built from the ground up for modern large language model (LLM) inference rather than being a modified version of existing general-purpose hardware.
Strategic Hardware Partnerships
The development of Jalapeño involves a complex supply chain where each partner provides specific expertise. OpenAI leads the chip design, while Broadcom contributes silicon manufacturing and networking technology, including its Tomahawk series. Additionally, Celestica has been contracted to manage boards, racks, and overall system integration.
The collaboration aims to provide a full-stack solution that allows for faster and more cost-effective model deployment. Key details regarding the project include:
Performance and Technical Specifications
OpenAI claims that Jalapeño delivers performance per watt that is "substantially better" than current state-of-the-art hardware. While these figures are self-reported and awaiting independent verification, the architecture focuses on minimizing data movement to push utilization closer to theoretical maximums. Engineering samples are currently running machine learning workloads in laboratory settings, including the GPT-5.3-Codex-Spark model.
The move into custom silicon follows years of OpenAI focusing primarily on software and model development. By controlling the hardware layer, the company intends to mitigate the high costs associated with third-party infrastructure. Broadcom CEO Hock Tan noted that the first deployment will be a massive undertaking involving Microsoft and other key partners to support the next generation of AI scaling.
The transition to custom ASICs highlights a growing trend where leading AI labs seek to decouple from general-purpose GPU dependencies to achieve better efficiency. As Jalapeño moves toward production, it will serve as a benchmark for how vertically integrated companies can optimize the underlying infrastructure for specific intelligence tasks.