OpenAI says custom Jalapeño chip outperforms Nvidia processors

OpenAI says its custom-built Jalapeño inference chip outperforms Nvidia processors in benchmark tests regarding power efficiency and response speed.

This development represents a strategic move to lower the company's dependence on Nvidia, which currently dominates the AI hardware market. By designing its own silicon, OpenAI seeks to optimize the specific workloads required for large-scale AI inference while reducing operational costs.

Richard Ho, OpenAI's head of hardware, said the Jalapeño chip delivers higher AI work per unit of power and faster response speeds than existing Nvidia options. According to company data, the hardware outperformed Nvidia in two specific benchmark categories ^[1] — performance per watt and response speed.

The push for custom silicon began as early as June 2026 ^[2]. The company's goal is to achieve better efficiency for inference workloads, which are the processes used when an AI model generates a response for a user. Reducing the power required for these tasks is critical as AI models scale in size and user demand increases ^[3].

Reports on the chip's current status vary. Some reports indicate the chip was developed in June and has already shown superior results in tests ^[2]. However, other reports suggest OpenAI is still in the process of finalizing its first custom-chip design ^[4].

OpenAI has not detailed the exact production volume or the timeline for a full rollout of the Jalapeño hardware across its data centers. The company continues to operate from its hubs in San Francisco and New York ^[4]. The transition to proprietary hardware would allow OpenAI to tailor the physical architecture of its chips to the specific mathematical needs of its models, potentially bypassing the general-purpose limitations of commercial GPUs ^[3].