OpenAI Releases First Results of Its Own AI Chip Jalapeño

OpenAI has published the first testing results of its proprietary AI chip, Jalapeño, and announced that the chip demonstrates significant advantages in both energy efficiency and processing speed compared to commercial AI hardware.

The company tested Jalapeño using the public InferenceX benchmark and compared its results against leading AI systems currently on the market. Testing was conducted on three models: GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T.

According to OpenAI, Jalapeño performed 1.5-1.9 times more AI work per watt at peak load across all three models, while full processing latency was 1.7-3.6 times lower compared to the systems being benchmarked. Particularly for interactive workloads, the chip demonstrated 2.1-4.1 times higher throughput.

In the case of the largest publicly tested model, Kimi K2.5 1T, Jalapeño achieved approximately 1.5 times higher throughput per watt and 3.4 times lower full latency.

The chip's nominal power consumption is 700 watts, though during testing its sustained consumption did not exceed 550 watts.

According to OpenAI, when using its own internal, more advanced AI models, Jalapeño's advantages increased further, which may suggest that its architecture is particularly efficient for larger and more complex AI workloads.

The chip development process took only nine months from initial design to preparation for production, or tapeout. OpenAI used its own AI models in the chip design and optimization process as well.

According to the company, using Codex and GPT-Astra, the team deployed three open-weight models to high-performance levels in just two months. In specific GPT-OSS attention and mixture-of-experts components, implementations created by AI were found to be 1.5-1.8 times faster than versions written by humans.

OpenAI plans to deploy Jalapeño in its own computing infrastructure by year-end. Additionally, the company is already working on a second generation of the chip, with development of the third generation already underway.

Despite creating its own chip, OpenAI states that it will continue to use AI accelerators from Nvidia and other partners alongside Jalapeño, both for model training and inference.