On Wednesday, OpenAI introduced its inaugural custom-designed inference processor, which was created and produced in conjunction with Broadcom. Named Jalapeño, this innovative processor was tailored to meet the distinct requirements of OpenAI’s inference systems. The development of the chip was supported by OpenAI’s own AI models, according to the company.

Although testing is still ongoing, OpenAI reports that initial findings indicate a much improved performance-per-watt compared to existing state-of-the-art alternatives.

The collaboration was formally revealed in October, though rumors about OpenAI’s chip initiatives have circulated for some time as a means to lessen the company’s reliance on Nvidia’s GPUs. Both Google and Amazon have crafted custom chips aimed at achieving a similar objective, frequently referred to as “AI accelerators” — silicon specifically created to enhance machine learning workloads.

Greg Brockman, president of OpenAI, discussed the organization’s strategy for chip creation on its in-house podcast, shortly following the announcement of the Broadcom partnership.

“We possess a profound comprehension of the workload,” Brockman stated during the episode. “We’ve been seeking out particular workloads that are not adequately served, [and pondering] how can we develop something that will accelerate what’s achievable?”

Jalapeño is specifically engineered for inference, the method of executing pre-existing AI models based on user inputs. In the announcement, OpenAI highlighted the chip’s minimal operating expenses when executing real-time coding models. It is probable that more resource-intensive activities like pre-training will still depend on Nvidia hardware, but even slight reductions in inference costs could significantly enhance the company’s profit margins.

Improving that inference system may become an essential element in the financial aspects of AI in the future — and it is expected to occur at every tier of the stack. OpenAI is already creating agentic products such as Codex and the models that drive them, along with data centers to operate those models. Transitioning to purpose-built chips enables the company to advance even further in that journey, as articulated in its announcement.

“OpenAI is not solely developing cutting-edge models or constructing products on top of them; it is also designing the underlying infrastructure: chip architecture, kernels, memory systems, networking, scheduling, deployment systems, and product experience,” the company stated. “Because OpenAI functions across the stack, each layer can be enhanced towards the same objective: making its models quicker, more dependable, and more cost-effective for users.”