OpenAI and Broadcom have announced a new chip called Jalapeño, built specifically to run large language models in data centers. The two companies say the chip is the first in a planned series that will be refined over time.
According to Ars Technica, the chip is an ASIC, or application-specific integrated circuit, meaning it was designed from scratch for one purpose rather than adapted from a general-purpose processor. Broadcom says the design was based on detailed insights drawn from conversations with OpenAI researchers and was shaped by OpenAI's own roadmap for future models and products. The entire design and production process took nine months.
The central claim from both companies is that Jalapeño is better matched to the current demands of large language model inference than the chips that existing data centers rely on. Inference refers to the process of running a trained AI model to generate responses, as opposed to the training process itself, which requires different hardware configurations.
OpenAI says that early testing shows Jalapeño will deliver better performance per watt than current state-of-the-art chips, but the company also noted it has not finished measuring performance. A detailed technical report is expected to be released in the coming months.
The announcement comes as demand for AI inference capacity has grown sharply. Running large language models at scale requires substantial computing infrastructure, and both companies are positioning Jalapeño as a more efficient solution than repurposing chips originally designed for other workloads. Broadcom is an established silicon supplier with experience building custom chips for major technology companies, and the partnership with OpenAI marks a step toward the AI company controlling more of its own hardware stack.
No production timeline or deployment date was given in the announcement. Both companies described this first-generation chip as the beginning of a longer-term collaboration.
