OpenAI’s proprietary Jalapeño inference chip has demonstrated performance benchmarks surpassing competing accelerator designs.

OpenAI has unveiled benchmark results for its first custom AI inference chip, Jalapeño, demonstrating industry-leading gains in efficiency and speed. According to company data, Jalapeño delivers between 1.5 and 1.9 times higher throughput per kilowatt and up to 3.6 times lower latency compared to NVIDIA’s GB300-class systems. These improvements mark a significant advancement in AI inference capabilities, particularly for large language models such as GPT-OSS 120B, DeepSeek R1, and Kimi K2.5 1T.

First announced on June 24, 2026, in collaboration with Broadcom, Jalapeño represents OpenAI’s strategic entry into custom silicon development. Unlike general-purpose GPUs, the chip is purpose-built exclusively for running pre-trained AI models, including those powering ChatGPT. Optimized specifically for inference workloads, Jalapeño minimizes power consumption and reduces response times. While NVIDIA’s programmable GPUs and extensive ecosystem continue to dominate the broader AI hardware market, Jalapeño’s specialized architecture provides OpenAI with a distinct competitive advantage for its internal operations.

Released on August 25, the benchmarks underscore Jalapeño’s capacity to handle interactive AI tasks with unprecedented efficiency. On the largest tested model, Kimi K2.5, the chip achieved 1.5 times higher peak performance per watt and cut end-to-end latency by 3.4 times relative to competing systems. These metrics highlight its ability to manage high-demand applications, such as real-time chatbot interactions, while significantly lowering energy expenditures—a critical consideration as global AI adoption accelerates.

OpenAI’s engineering team credited artificial intelligence itself for accelerating Jalapeño’s development cycle. By leveraging internal AI tools, the team progressed from initial design to tapeout in just nine months. AI also played a direct role in optimizing the chip’s circuitry and programming, enabling faster deployment and enhanced performance. Notably, AI-generated implementations for specific model blocks outperformed human-written code by 1.5 to 1.8 times, further streamlining the development process.

Although Jalapeño will not be available for external sale, its integration into OpenAI’s infrastructure is expected to yield substantial operational benefits. Faster, more power-efficient inference reduces computational costs and expands serving capacity, directly improving the company’s operating leverage. With second-generation (Gen 2) and third-generation (Gen 3) chips already in development, OpenAI is doubling down on custom silicon as a long-term strategic moat.

In the broader semiconductor landscape, Jalapeño’s performance positions OpenAI as a formidable competitor to Google’s Tensor Processing Units (TPUs), though their use cases diverge. Google TPUs support both training and inference and are offered commercially through Google Cloud, whereas Jalapeño remains strictly an internal inference accelerator. Similarly, while NVIDIA GPUs retain their status as the industry standard for versatility and ecosystem compatibility, Jalapeño’s narrow specialization delivers superior efficiency for targeted workloads.

OpenAI plans to deploy Jalapeño at scale across its compute infrastructure by the end of the year. As the company continues to refine the platform and expand its capabilities, Jalapeño marks a pivotal step toward satisfying surging global demand for AI-powered applications while maintaining strict control over costs and environmental impact.