OpenAI Launches First LLM Inference Custom Acceleration Chip Jalapeño
OpenAI has released the Jalapeño chip, the first custom acceleration chip designed specifically for large language model inference, with architecture and algorithm design handled by OpenAI in collaboration with Broadcom and Celestica to advance mass production.
The chip aims to enhance the speed of ChatGPT, Codex, API, and future intelligent agent products while reducing computing costs. It employs algorithm-hardware co-design to restructure the LLM core, data movement, and network architecture, achieving utilization close to hardware limits.
Thanks to OpenAI's own AI model assistance, the chip achieved the fastest record for ASIC development, taking only 9 months from concept to tape-out. The first batch of samples has successfully run workloads such as GPT-5.3-Codex-Spark in the lab, with energy efficiency significantly surpassing existing top devices.
Source: Public Information
ABAB AI Insight
OpenAI previously relied on external computing power from Nvidia and others. The self-developed Jalapeño reflects a vertical integration strategy similar to its approach in model training, significantly shortening the development cycle and optimizing energy efficiency, marking a move towards full-stack control for the AI giant.
Collaborating with Broadcom and Celestica accelerates industrial mass production on the capital path, with Jalapeño restructuring architecture to address LLM inference pain points, reducing reliance on general-purpose GPUs and providing cost advantages for future intelligent agent deployments.
Compared to the traditional chip development cycle of 18-24 months, OpenAI is currently in a leading phase of explosive innovation in self-developed AI hardware, with its model-assisted design capability becoming a core competitive advantage.
Essentially, this represents a technological replacement and industrial chain restructuring: custom ASICs are set to replace general-purpose GPUs in the dominant role of LLM inference, with capital from giants like OpenAI concentrating on proprietary hardware stacks, restructuring the AI computing power supply chain and driving structural breakthroughs in energy efficiency and costs.
ABAB News · Cognitive Laws
Model-assisted chip design accelerates the innovation cycle through vertical integration.
Algorithm-hardware collaboration maximizes utilization limits.
Self-developed hardware is a cost-reduction tool, defining the future of computing power.