OpenAI Says Its Jalapeño Chip Beats NVIDIA In Speed & Efficiency

OpenAI says its first custom AI chip, Jalapeño, can outperform NVIDIA’s GB200 and GB300 systems on AI inference speed and efficiency. Yet even as it promotes those gains, OpenAI says NVIDIA accelerators will remain part of its infrastructure strategy.
 

TL;DR

 

• OpenAI says Jalapeño delivers 1.5x to 1.9x more AI work per watt and 1.7x to 3.6x lower end-to-end latency than comparison systems.
• OpenAI models helped take Jalapeño from initial design to tapeout in nine months.
• Deployment is planned by year-end, while NVIDIA hardware will continue supporting training and inference.

 

OpenAI Says Jalapeño Delivers Faster, More Efficient AI Inference

Built with Broadcom, Jalapeño was designed specifically for the large language model workloads behind ChatGPT, Codex, the API, and future agentic products.

OpenAI tested the chip using SemiAnalysis’ public InferenceX benchmark across GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T. Across those models, the company says Jalapeño delivered 1.5x to 1.9x more AI work per watt at peak throughput and 1.7x to 3.6x lower end-to-end latency than the comparison systems.

For highly interactive workloads, OpenAI reported 2.1x to 4.1x higher performance. On Kimi K2.5 1T, the largest public model tested, Jalapeño reportedly produced about 1.5x higher peak performance per watt and 3.4x lower latency.

The chip is rated at 700 watts, although OpenAI said sustained power remained at or below 550 watts during the tested workloads.
 

 

AI Helped OpenAI Design Jalapeño In Nine Months

Jalapeño is also a test of OpenAI’s full-stack strategy. The company says its AI models helped engineers move from initial design to tapeout in nine months by exploring implementations, shortening design and verification loops, and optimizing arithmetic circuits.

The architecture was built to reduce data movement and communication delays across compute, memory, networking, and software. That is intended to help Jalapeño handle both prompt processing and token-by-token generation efficiently, especially for agentic workloads.

OpenAI also used Codex with GPT-Astra to optimize three open-weight models for Jalapeño within two months. For selected GPT-OSS attention and mixture-of-experts blocks, AI-generated implementations ran 1.5x to 1.8x faster than existing human-expert-written versions, although those gains did not apply to the full model.
 

 

Jalapeño Heads To OpenAI Data Centers While NVIDIA Remains In The Mix

OpenAI plans to begin deploying Jalapeño inside its compute infrastructure by the end of the year. A second generation is already deep in development, while a third is taking shape.

“Jalapeño is working first-party silicon with measured results, and it is the beginning of a multigenerational platform,” OpenAI said.

Still, Jalapeño is not replacing NVIDIA outright. OpenAI says rising AI demand will require compute from every available source, and it plans to keep widely deploying accelerators from NVIDIA and other partners for both training and inference.

That positions Jalapeño as OpenAI’s first major step toward controlling more of its AI stack, while keeping the external hardware ecosystem it still relies on firmly in play.