OpenAI unveiled its first proprietary AI chip, marking a strategic shift toward in-house silicon that challenges Nvidia’s dominance in the accelerator market.
OpenAI has unveiled Jalapeño, its first custom AI inference chip developed in partnership with Broadcom. The announcement marks a strategic shift for the company, which is now designing the hardware that runs its models alongside building the models themselves. CEO Sam Altman summarized the milestone in a brief post: “we made a chip and it is fast.”
According to OpenAI, Jalapeño delivers between 1.5x and 1.9x more AI work per watt than comparison systems across benchmarks including GPT-OSS 120B, DeepSeek R1, and Kimi K2.5 1T. The chip also achieved 1.7x to 3.6x lower end-to-end latency in internal testing. These results were measured against commercial systems built around Nvidia accelerators using the InferenceX benchmark. While the Nvidia comparison setups were rated at 1,200W or 1,400W depending on the test configuration, Jalapeño operates at 700W. For highly interactive AI workloads, OpenAI reported performance gains of up to 4.1x. The company notes, however, that these are self-reported benchmark results on specific workloads and do not indicate that Jalapeño has replaced Nvidia across broader AI computing.
Jalapeño is purpose-built for inference—the stage where an AI model processes and responds to user requests. This includes standard queries, coding agents, AI assistants, and API calls. Rather than pursuing a general-purpose processor, OpenAI engineered the chip specifically around these inference workloads. The company collaborated with Broadcom on the silicon and Celestica on the boards, racks, and production systems. By optimizing the entire stack—including the chip, memory, networking, and software—OpenAI aims to address two critical industry challenges: surging computational demand and escalating electricity costs.
OpenAI plans to begin deploying Jalapeño within its own computing infrastructure by the end of 2026. Importantly, the company will continue to widely deploy Nvidia and other accelerators for both training and inference tasks. Jalapeño represents the first step in a multi-generation chip program. OpenAI confirmed that a second generation is already deep in development, with a third generation currently taking shape.
Notably, AI itself played a central role in Jalapeño’s creation. Leveraging AI tools to explore implementations, optimize circuits, and accelerate engineering and testing, OpenAI moved from initial design to tape-out in just nine months. The overarching objective is straightforward: extract more AI work from the same power and hardware footprint. If successfully scaled, this approach could yield faster ChatGPT responses, more capable AI agents, and reduced operational costs as global demand continues to surge.