OpenAI's First AI Chip Jalapeño Announced, Collaborated with Broadcom but Testing Still Ongoing

OpenAI and Broadcom jointly announced the Jalapeño chip design on June 24, 2026, the first result of their partnership reached in October 2025. The processor is positioned as the "first Intelligence Processor," optimized for large language model inference and ChatGPT-like products.

Actual Progress in Hardware Control

Over the past two years, OpenAI has relied heavily on external GPU supply for training and inference. The launch of Jalapeño marks the beginning of bringing some compute nodes within its own design scope. The official blog disclosed that the chip went from initial design to tape-out in just nine months, reflecting close collaboration between the software team and Broadcom's silicon implementation team. Such a cycle is uncommon in traditional chip development, indicating that OpenAI has used its own models to optimize the chip design process.

OpenAI claims its performance per watt is "significantly better than the current state-of-the-art," with a full technical report to be released in the following months.

Signals of Supply Chain and Cost Structure Adjustment

Reducing dependence on a single supplier is a repeatedly emphasized motivation for this collaboration. Jalapeño is planned to be deployed first in its own data centers, with the initial launch scheduled for the end of 2026. If this timeline is executed, OpenAI will gain a second inference path beyond GPUs. Whether the cost structure of this path can be lower than existing solutions will directly affect the marginal profit of ChatGPT services.

Multi-generational computing platforms are another expression in the cooperation agreement between the two parties, implying that Jalapeño is not an isolated product but the foundation for subsequent iterations. If future generations of chips continue to evolve along the same architecture, OpenAI's voice at the hardware level will gradually accumulate. However, this also requires continuous investment in design, verification, and manufacturing coordination resources, balancing with the pace of existing model iterations.

Targeted Design for Inference Scenarios

Jalapeño is explicitly described as an accelerator built around the "future vision of LLM inference." This differs from the focus of training chips, paying more attention to the balance between throughput and power consumption.

If early data center performance meets expectations, OpenAI may consider opening some capacity to partners in 2027; if performance falls short, it may extend the procurement cycle of external GPUs. Both paths will affect the capacity planning of other industry participants.