OpenAI and Broadcom have unveiled the OpenAI Jalapeño chip, a custom-built artificial intelligence accelerator optimized for large language model inference. The first-generation processor represents a shift in the company’s strategy to build a full-stack infrastructure for its models. Notably, early testing indicates that the accelerator delivers performance per watt that exceeds current industry standards.
The development of the hardware was completed in collaboration with Broadcom and Celestica. Specifically, Celestica assisted with board and rack system integration, while Broadcom provided silicon implementation and networking technologies. The chip is designed to run machine learning workloads, and engineering samples are currently running in laboratories at target production frequencies, including the GPT-5.3-Codex-Spark model.
Technical Specifications of the OpenAI Jalapeño chip
The architecture of the OpenAI Jalapeño chip focuses on reducing data movement while balancing compute, memory, and networking resources. Consequently, this design allows the hardware to achieve utilization rates close to its theoretical peak performance. The chip integrates Broadcom’s Tomahawk networking silicon to support large-scale production environments.
According to the companies, the processor is not a general-purpose accelerator but a specialized design built for interactive large language model (LLM) products. It is designed to support the workloads of ChatGPT, Codex, and the OpenAI application programming interface (API).
Rapid Development Cycle and AI Integration
The custom silicon was co-developed from initial design to manufacturing tape-out in nine months. This timeline represents one of the fastest application-specific integrated circuit (ASIC) development cycles in the semiconductor industry. Furthermore, OpenAI used its own artificial intelligence models to accelerate parts of the design and optimization process.
“Jalapeño was designed from the ground up for LLM inference using detailed insights from our close collaboration with OpenAI researchers. We optimized the architecture around the kernels, memory movement, networking, and serving patterns that matter most for frontier AI models.”
Richard Ho, Lead of OpenAI’s Hardware Program
Infrastructure Strategy and Industry Partnerships
The introduction of custom hardware is part of a broader effort to manage the physical infrastructure required for advanced computing. By designing its own chips, the company aims to improve compute efficiency and lower operational costs. This efficiency directly impacts the speed and reliability of consumer and enterprise services.
“Jalapeño is part of our long-term full-stack infrastructure strategy to make compute more abundant. By designing more of the stack ourselves, we can serve more intelligence with greater efficiency.”
Greg Brockman, President and Co-Founder of OpenAI
Future Deployment and Scalability
The companies plan to deploy the OpenAI Jalapeño chip at gigawatt scale with data center partners. Initial deployments are scheduled to begin by the end of 2026. Microsoft and other partners will participate in the infrastructure rollout over multiple generations.
This hardware initiative aims to make advanced models more accessible by reducing the cost of inference. As a result, future updates to services like ChatGPT and Codex will benefit from faster response times and improved availability during high-demand periods.
Source: OpenAI
