OpenAI has unveiled Jalapeño, its first processor designed specifically to run large language models. The chip was developed in collaboration with technology company Broadcom and is being manufactured by Taiwanese chipmaker TSMC. It marks the first step in a long-term plan for OpenAI to build its own hardware infrastructure instead of relying exclusively on supplies from Nvidia.
The chip is designed exclusively for inference, the process in which a model handles a user query and returns a response. Whenever ChatGPT answers someone's question, generates code, or analyzes a document, inference is taking place. And this part of its operations costs OpenAI the most money.
OpenAI models also worked on the chip
Developing the chip from the initial concept to sending it into production took just nine months. That is an exceptionally short time by semiconductor industry standards. OpenAI acknowledged that parts of the design and optimization process were carried out directly by its own models. “The extent to which our models were able to accelerate development surprised even us,” OpenAI President Greg Brockman said in an interview with CNBC.
The chip was designed from the ground up by OpenAI engineers. Broadcom handled the silicon implementation, boards, and integration into server systems. The server racks will then be built by Canadian electronics manufacturer Celestica. Production samples of Jalapeño are already running in OpenAI's laboratories and processing real workloads, including the GPT-5.3-Codex-Spark model, while meeting target performance and power consumption levels.
Performance comparable to Nvidia chips, costs cut in half
According to initial tests, the chip is significantly more efficient than current top-tier alternatives. OpenAI has not yet released exact figures, and a more detailed technical report is expected in the coming months. However, Broadcom CEO Hock Tan provided a specific figure in an interview with Bloomberg: compared with standard graphics processors used for artificial intelligence, Jalapeño offers cost savings of approximately 50%.
Tan also said that Jalapeño's performance is comparable to Nvidia's Blackwell chips or the tensor processing units Google develops for its own use. Deployment in data centers is expected to begin before the end of this year. OpenAI plans to integrate the chips into large data centers operated by its investor Microsoft and other partners.
OpenAI is struggling to secure enough computing capacity
The entire project is driven by demand for computing power that is growing faster than OpenAI can secure it. “We simply cannot obtain computing capacity fast enough,” Brockman said. Tan confirmed this view. According to him, demand from his six custom chip customers is “insatiable.” And not just this year or next. “We are seeing the same, if not even greater, demand for 2028,” he added.
OpenAI has therefore been actively expanding its range of hardware suppliers in recent months. It signed an agreement with Amazon Web Services involving Trainium chips and is collaborating with AMD and Cerebras Systems. However, its own Jalapeño chip is in a different category because OpenAI designed it itself and will use it exclusively for its own operations.
From products and models to chips
Broadcom CEO Hock Tan and President Charlie Kawwas physically handed Jalapeño over to CEO Sam Altman and Greg Brockman. It was a symbolic moment that OpenAI itself described as a crucial step in its strategy to build the entire technology stack in-house. “By designing more layers of the stack ourselves, we can deliver more intelligence with greater efficiency,” Brockman wrote in the press release.
The two companies announced their partnership last October, after collaborating informally for approximately 18 months. At the time, they revealed plans to deploy 1.3 gigawatts of chip capacity in 2027. Tan now suggests that they intend to exceed that threshold. By 2029, the two companies plan to deploy chips with a total capacity of 10 gigawatts.
Jalapeño is not the ultimate goal, but the first in a planned series spanning multiple generations of processors. OpenAI is building its own technology foundation in which chips, data centers, and models form a unified whole optimized around a single goal: faster, more reliable, and more accessible artificial intelligence.
