OpenAI Claims In-House Jalapeno Chips Outperform Nvidia in Early Testing

OpenAI has announced that its newly developed in-house artificial intelligence processor, named Jalapeno, has outperformed Nvidia's current market-leading chips in internal and public testing. The breakthrough highlights OpenAI's efforts to independently develop AI infrastructure and cut down on high operational costs. According to OpenAI's chip chief Richard Ho, the Jalapeno processor outperformed Nvidia's GB300 chip in two main areas: the amount of AI workload it could process per watt of power and its speed in returning responses.

Created in collaboration with custom chipmaker Broadcom Inc., Jalapeno operates efficiently at a low voltage of 700 watts, thereby delivering both high throughput and low latency. This custom AI chip, which can outperform Nvidia processors, was tested using smaller open-source models alongside third-party models from DeepSeek and Moonshot AI.

The semiconductor is specifically designed for the inference phase of AI, the operational stage where trained models respond to user prompts, as opposed to Nvidia's technology, which excels in the initial training phase. Although OpenAI plans to roll out the new chips later this year, it still considers Nvidia a crucial supplier and partner.

OpenAI is also advancing the development of subsequent versions of its hardware ecosystem, with a second generation chip reportedly nearing completion and a third generation already in the conceptual stage.