OpenAI Tests Its First Custom AI Chip Jalapeno
OpenAI shared the test results of Jalapeno, its first custom AI processor developed in collaboration with Broadcom. In inference tasks, the chip demonstrated higher power efficiency and faster response times compared to Nvidia's GB300 model.
AI company OpenAI announced that it has tested its first in-house processor named 'Jalapeno', developed in partnership with Broadcom. Focused on the AI inference phase, the chip is expected to significantly reduce data center infrastructure costs and start deployment later this year.
Performance and Efficiency Tests
Jalapeno was benchmarked against Nvidia's leading GB300 model on open testing systems. In the tests, the new chip excelled in two main categories: the amount of processing performed per unit of power consumed and response latency.
OpenAI Head of Chips Richard Ho stated that the processor delivered strong results at a power level as low as 700 watts. The company aims to reduce electricity consumption, one of the largest expense items in data centers, through this effort.
Design Focused on Inference Tasks
Jalapeno was specially designed for the inference phase, where trained models respond to incoming requests, rather than for model training. The tests were not conducted against Nvidia's newly shipping Vera Rubin series.
The chip was tested with models from DeepSeek, Moonshot AI, and OpenAI's unreleased advanced models. It was stated that as the workload increased, the advantages offered by the chip's design became even more pronounced.
Next-Generation Processor Plans and Partnerships
OpenAI leveraged its own AI models to accelerate the chip development process. While the company plans to finalize the design of the second-generation chip in the coming months, it reported starting conceptual work on the third generation as well.
Richard Ho emphasized that Nvidia will continue to be a core supplier and partner. The company also intends to keep working with existing infrastructure providers like Cerebras, while gradually reducing infrastructure costs through its own chip production.