OpenAI Unveils Jalapeo AI Chip Developed With Broadcom for Efficient Inference

OpenAI has announced Jalapeo, a custom AI chip developed with Broadcom, marking a shift from solely building AI models to also designing and manufacturing hardware. The company aims to improve efficiency and reduce response times for AI inference tasks with this new chip.

Key Highlights

• OpenAI introduces Jalapeo, a custom AI inference chip developed with Broadcom.
• Jalapeo delivers up to 1.9 times more AI work per watt and up to 3.6 times lower latency.
• The chip will see limited deployment by late 2026 with broader rollout planned for 2027.

Jalapeo Chip Development and Purpose

OpenAI designed Jalapeo specifically for running AI inference, which is the process of an AI model generating answers to user requests. The chip was developed in collaboration with Broadcom, while Celestica handled the boards, rack systems, and production hardware. OpenAI stated that Jalapeo was built around the operational needs of its AI models.

By co-designing the chip, software, memory, networking, and serving systems, OpenAI seeks to boost performance and efficiency across the entire technology stack. The company highlighted that Jalapeo addresses the challenge of achieving high performance without increasing latency, a key requirement for interactive AI applications.

Performance Metrics and Efficiency Gains

OpenAI reported that Jalapeo delivers 1.5 to 1.9 times more AI work per watt compared to existing systems across three tested models. The chip also achieved 1.7 to 3.6 times lower end-to-end latency. For highly interactive workloads, performance was up to 4.1 times higher than comparison systems.

These improvements mean Jalapeo can process more requests while using less electricity and reducing wait times for users. For large-scale AI systems handling millions of requests, even small efficiency gains can lead to significant cost and energy savings.

Strategic Goals and Future Plans

Jalapeo is part of OpenAI’s broader strategy to control more of the technology stack behind its AI services. Rather than relying solely on third-party hardware, OpenAI wants to optimize chips, software, memory, networking, and serving systems together for better results.

The company completed Jalapeo’s development, from initial design to manufacturing tape-out, in just nine months. OpenAI used its own AI models to assist engineers during parts of the design and optimization process.

Despite these advancements, OpenAI does not plan to replace Nvidia hardware immediately. The company will deploy Jalapeo in small volumes by the end of 2026 and expand deployment through 2027. OpenAI also intends to continue working with other hardware partners, including Nvidia, as part of its overall compute strategy.