On the path to摆脱 reliance on a single source of computing power, OpenAI has taken a highly symbolic step. On June 24 local time, OpenAI officially partnered with semiconductor giant Broadcom to launch its first customized AI inference chip - Jalapeño. This move marks that OpenAI is no longer satisfied with leading only in software and algorithms, but has begun to deeply participate in the design of underlying hardware architecture.

Jalapeño is a custom application-specific integrated circuit (ASIC) specifically designed for large language model (LLM) inference tasks. The project demonstrated astonishing development efficiency, taking only 9 months from the initial blueprint design to the final chip fabrication delivery. To achieve this goal, OpenAI even used its own AI models to accelerate the chip design process. In terms of division of labor, OpenAI was responsible for the core architecture design, Broadcom undertook the complex silicon implementation and network hardware support, while Canadian electronics manufacturing service provider Celestica was in charge of the subsequent board and rack integration work.

In terms of performance, although OpenAI is still evaluating the final data, early laboratory test results show that Jalapeño demonstrates excellent performance per watt when running key machine learning tasks such as GPT-5.3, Codex, and Spark. Through carefully designed architecture, the chip effectively reduces data transmission loss and achieves an optimal balance between computing, memory, and network resources, resulting in much higher actual utilization than industry standards.

Greg Brockman, President and co-founder of OpenAI, stated that as the world enters an economy centered around computing, Jalapeño is a crucial part of the company's strategy to build a full-stack infrastructure. By mastering the underlying technology stack independently, OpenAI not only significantly reduces expensive operational costs, but also drives more powerful intelligence with higher efficiency. Broadcom CEO Pichai Chen revealed that this is just the beginning of their cross-generation R&D roadmap.

In fact, this self-developed chip layout is an inevitable choice for OpenAI to cope with intense market competition. In the past, OpenAI relied heavily on Microsoft's Azure cloud computing cluster, but now, as the demand for inference computing power in the industry has rapidly exceeded training needs, a diversified computing foundation has become the lifeline to maintain commercial leadership. Google has already proven the huge cost advantages brought by software-hardware collaboration through its self-developed TPUs, and OpenAI's entry will undoubtedly further intensify global competition in the AI industry's infrastructure.