OpenAI and Broadcom Unveil Jalapeño, a Custom AI Inference Chip Challenging Nvidia

OpenAI and Broadcom on Wednesday unveiled Jalapeño, their first custom AI accelerator chip designed specifically for large language model inference. The application-specific integrated circuit (ASIC) targets workloads behind ChatGPT, Codex, the OpenAI API, and future agentic systems, according to the companies.

Jalapeño is built for inference only, not training, and is positioned as a purpose-built alternative to general-purpose GPUs from Nvidia and AMD. The chip's development was accelerated using OpenAI's own AI models, the companies said in a joint statement.

Design and Availability

The chip is a custom ASIC optimized for the unique computational patterns of transformer-based LLMs. Both OpenAI and Broadcom indicated that Jalapeño could be made available to external AI firms, though no pricing or timeline for general availability was provided. Key details include:

• Designed from the ground up for current and future large language models.
• Built in partnership with Broadcom, leveraging their chip design expertise.
• Targets inference workloads for ChatGPT, Codex, and OpenAI API deployments.
• Employs a specialized architecture to reduce latency and energy consumption compared to GPUs.
• Development was sped up using OpenAI's own models to simulate and optimize chip performance.

Implications for the AI Chip Market

The launch represents the clearest signal yet that OpenAI, one of Nvidia's largest customers, is seeking to reduce its dependence on general-purpose GPUs. By developing a custom inference chip, OpenAI can tailor hardware to its software stack, potentially lowering costs and improving performance for its growing user base. While Nvidia remains dominant in AI training, specialized ASICs like Jalapeño could reshape the inference market. The chip may also open a new revenue stream for OpenAI if offered to third parties.

Broadcom, a major supplier of custom chips for networking and infrastructure, gains a high-profile partner in the AI space. The partnership could expand as OpenAI develops future silicon for training or other workloads. Analysts expect the first Jalapeño-powered deployments to begin in late 2026.