A new in-house chip promises faster inference for large language models, but deployment and supply chain limits will test OpenAI’s push for hardware independence.

OpenAI unveiled its first custom artificial intelligence chip, developed in collaboration with Broadcom, to accelerate the processing and deployment of its own infrastructure.

In the field of AI, industry leaders such as OpenAI and Anthropic are facing a shortage of computing power to run the most powerful chatbots and development tools. Some companies, including OpenAI, are considering creating internal chips to reduce costs and offer an alternative to Nvidia GPUs, which are typically used for AI.

The chip, named Jalapeño, was designed by OpenAI engineers in close collaboration with Broadcom to perform the inference task – processing data to answer a user’s query in a chatbot like ChatGPT.

According to OpenAI’s head of hardware, Jalapeño is designed for fast and efficient operation with large language models that power many AI applications.

Deepening OpenAI’s Infrastructure Self-Sufficiency

It will be productive, in our view, for all future versions of large language models.

– Richard Ho

Jalapeño is planned to be deployed by the end of the year; this marks the first step in a multi-stage plan for chip development.

Celestica in Canada will build server systems that work with the chips and are dedicated exclusively to OpenAI.

According to tests in OpenAI laboratories, chip samples have already demonstrated the expected power and performance metrics when paired with the GPT-5.3-Codex-Spark model.

Developers spent about nine months finalizing the chip design, after which it was sent to production at TSMC in Taiwan, partly aided by the use of AI to speed up certain stages of the manufacturing process.

Although earlier rumors circulated about the possibility of creating a proprietary chip, the industry is facing increased demand for memory, which affects the margin on Broadcom’s custom chips compared with the company’s other products. Memory supplies for such chips are provided by SK Hynix and Samsung Electronics.

Anthropic is also considering developing its own chip, according to sources disclosed in April.

This development underscores the growing role of hardware in scaling AI and companies’ efforts to reduce dependence on third-party manufacturers, while increasing data processing speed in future AI applications.