OpenAI Builds Its Own ‘Jalapeño’ AI Chip to Run ChatGPT Faster and Cheaper

On Wednesday, OpenAI announced its first custom AI chip, Jalapeño, as the first stage in its plan to develop first-party infrastructure for running LLMs. Called an “Intelligence Processor,” the chip was developed with help from Broadcom and Celestica to help run services like ChatGPT and code-generation tools.

The companies say Jalapeño delivers better performance and efficiency when running already-trained models, rather than when training them. It’s optimized for LLMs and the way they process information, enabling higher throughput and lower power consumption compared with current AI silicon running similar inference tasks.

Early statements from Broadcom’s leadership say Jalapeño performs on par with Nvidia’s latest Blackwell chips and Google’s TPUs for these workloads, with around 50% better cost efficiency. That said, no technical specs like process node, clock speeds, or memory configuration have been released yet.

OpenAI plans to deploy Jalapeño in its data centers starting at the end of 2026, with expansion in subsequent years. The chip will run internal services such as ChatGPT and the OpenAI API, and may also support third‑party models hosted by OpenAI.

But what is the need for custom silicon for an AI company? OpenAI says it wants more control over performance, scaling, and cost, and wants to reduce its reliance on external vendors—especially as component costs rise across the industry.

Disclosure: Ziff Davis, ExtremeTech’s parent company, filed a lawsuit against OpenAI in April 2025, alleging it infringed Ziff Davis’s copyrights in training and operating its AI systems.