OpenAI on Wednesday unveiled its first custom-built processor aimed specifically at inference workloads, developed and manufactured in collaboration with Broadcom. The chip is named Jalapeño, and OpenAI said that its own AI models contributed to the chip’s design.
The partnership was officially announced in October, though speculation about OpenAI building its own chips has circulated for some time as a way to reduce dependence on Nvidia GPUs. Google and Amazon have similarly developed custom silicon, often referred to as "AI accelerators," to speed up machine learning tasks.
Purpose and early results
Jalapeño is optimized for inference — running pre-trained models to respond to user prompts. OpenAI highlighted in its announcement that the chip offers low operating costs when running real-time coding models. Early test results, according to the company, show significantly better performance-per-watt compared with current state-of-the-art alternatives.
Despite these gains on the inference side, more computation-intensive work such as model pre-training is likely to remain on Nvidia hardware.
Why this matters
OpenAI president Greg Brockman discussed the company’s chip strategy on its in-house podcast after the Broadcom partnership was announced. Brockman said OpenAI has a deep understanding of its workloads and is looking for specific, underserved tasks where custom hardware can accelerate capabilities.
Reducing inference costs can materially affect the economics of running AI services, because inference often involves continuous, high-volume requests. By designing chips tailored to those workloads, OpenAI seeks to lower costs and improve performance for production models.
In its announcement, OpenAI emphasized that it is optimizing across the entire stack — from chip architecture and kernels to memory systems, networking, scheduling, deployment systems, and user-facing product experience — so that each layer supports the same goals: making models faster, more reliable, and more affordable for users.
Current status
Jalapeño is still in testing and OpenAI describes the performance figures as early results. The company positions the chip as a step toward deeper vertical integration of its infrastructure, enabling further optimization across models, products, and the hardware they run on.
