OpenAI partners with Broadcom and TSMC to develop custom JalapeƱo AI inference chips for 2026 rollout
OpenAI is partnering with Broadcom and TSMC to develop custom JalapeƱo AI inference chips aimed at reducing infrastructure costs and scaling data center efficiency by 2026.
1. Strategic Shift to Custom Silicon
OpenAI has announced the development of the JalapeƱo chip, a custom application-specific integrated circuit (ASIC) designed specifically for large language model (LLM) inference. This initiative is a strategic effort to reduce the company's reliance on third-party hardware and mitigate the high costs associated with current infrastructure. With operational expenses for ChatGPT reaching an estimated US$14 billion this year, the company aims to improve efficiency and lower the financial burden of scaling its services. OpenAI collaborated with Broadcom for silicon engineering and networking integration, while TSMC will handle manufacturing and Celestica will manage board and rack assembly.
2. Architecture and Design Efficiency
The JalapeƱo processor is optimized to address data-movement bottlenecks that often hinder general-purpose hardware. By focusing on the specific requirements of LLM inference, OpenAI intends to push realized utilization closer to theoretical peak performance. The development process was notably accelerated, moving from initial design to manufacturing tape-out in nine months. To achieve this timeline, OpenAI utilized its own language models to automate and optimize portions of the hardware design. Early samples are reportedly already running frontier workloads, including an unreleased GPT-5.3-Codex-Spark model.
3. Vertical Integration and Future Rollout
By designing its own chips, OpenAI is transitioning toward a vertically integrated business model that encompasses chip architecture, software kernels, and network scheduling. This approach is intended to create a feedback loop where infrastructure efficiency gains allow for more responsive products, driving further user growth and reinvestment. The company plans to begin deploying the JalapeƱo hardware into data centers by the end of 2026. Broadcom CEO Hock Tan confirmed that the rollout will scale alongside partners like Microsoft to support gigawatt-scale data center integration, positioning OpenAI to compete more effectively with other major technology firms that have already established proprietary AI hardware pipelines.