Details
- OpenAI announced Jalapeño, its first custom AI chip, designed in-house and brought to production with Broadcom.
- Jalapeño is purpose-built for large language model inference workloads that power ChatGPT, Codex, the OpenAI API, and future agentic products.
- The chip is optimized to process user queries efficiently, focusing on performance-per-watt gains by reducing data movement and tightly integrating compute, memory, and networking.
- OpenAI positions Jalapeño as a foundational part of its full-stack infrastructure strategy, extending control from models and software down to the silicon.
- Engineering prototypes are already running large-scale model workloads in lab environments, with early tests indicating competitive performance against leading accelerators such as Nvidia Blackwell and Google’s TPUs.
- OpenAI plans a phased rollout of Jalapeño at significant scale in partner data centers, including those operated with Microsoft, starting later this year.
- The collaboration with Broadcom leverages its networking silicon and manufacturing capabilities to deliver racks of Jalapeño-based systems tailored to OpenAI’s inference-heavy services.
- By building its own chip, OpenAI aims to secure more predictable compute capacity, lower operating costs, and reduce dependence on general-purpose GPU supply constraints.
- Jalapeño is framed as the first generation in a multi-year, multi-generational custom silicon roadmap to support upcoming versions of OpenAI’s frontier LLMs and agent platforms.
- The announcement underscores how AI companies are increasingly investing in proprietary hardware to match the scale and specificity of their workloads.
Impact
OpenAI’s move into custom silicon with Jalapeño signals a strategic shift toward vertically integrated AI infrastructure, similar to Google’s TPU approach but focused tightly on its LLM-powered services. Owning an inference-optimized chip can reduce cost per query and ease reliance on scarce GPUs, potentially reshaping bargaining dynamics with Nvidia and cloud providers while pressuring rivals to pursue comparable hardware strategies.
