OpenAI has unveiled its first custom inference processor, a chip named Jalapeño developed and manufactured in collaboration with Broadcom. The processor is designed for the computing patterns involved in serving trained AI models, extending OpenAI’s work from models and products into the hardware and systems beneath them.
Jalapeño remained in testing at the time of the announcement. OpenAI said early results showed materially better performance per watt than current leading alternatives, but the supplied report did not include benchmark tables, test conditions or independent validation. The company emphasized the operating cost of real-time coding models as a target workload.
Inference is the stage in which a trained model processes user requests and generates results. It differs from pre-training, the large-scale computation used to create a model. The new chip is tailored to inference, so more demanding training jobs are still likely to depend on graphics processors and other accelerators. Even so, serving models continuously is a major recurring expense, making incremental efficiency gains economically significant at large scale.
OpenAI and Broadcom publicly announced their partnership in October, after earlier speculation that OpenAI wanted to reduce its reliance on Nvidia hardware. Google and Amazon have also designed their own AI accelerators. Custom silicon can optimize around a known set of workloads, but it also requires substantial design, software and deployment investment and does not automatically replace general-purpose accelerators across every task.
OpenAI president Greg Brockman previously described the strategy as identifying workloads that existing hardware serves poorly and building around the company’s detailed understanding of those workloads. The company said AI models also assisted development of Jalapeño. The report does not specify which design stages used AI or how much time that assistance saved.
The processor is one component of a broader vertically integrated stack. OpenAI develops frontier models and agent products such as Codex, while also building data centers and working on chip architecture, software kernels, memory systems, networking, scheduling and deployment. Coordinating those layers could allow an optimization in one part of the system to improve product speed, reliability or cost elsewhere. It can also deepen the company’s capital requirements and dependence on manufacturing partners.
Jalapeño’s significance will depend on production yield, deployment scale, software compatibility and measured performance on real workloads. None of those questions is resolved by an unveiling. The confirmed development nonetheless marks OpenAI’s entry into purpose-built inference silicon and adds another major AI company to the industry push for hardware designed around its own models. Broadcom’s role also shows that custom does not necessarily mean wholly in-house: design can be tailored to one customer while fabrication and specialist development remain collaborative.
