OpenAI unveiled its proprietary Jalapeño chip, claiming it delivers significantly faster inference speeds and superior performance-per-watt compared to rival accelerators.

OpenAI has released new performance results for Jalapeño, its first custom inference chip designed to deliver more AI work per watt while significantly reducing response times. Unlike traditional approaches that treat processors as standalone hardware, OpenAI engineered Jalapeño as part of an integrated system combining chips, memory, networking, and software. The accelerator is purpose-built for AI inference—the stage where trained models generate outputs—rather than the training phase used to develop those models.

OpenAI tested Jalapeño against three public models, including Kimi K2.5 1T, the largest model evaluated. Across the suite, the company reports that Jalapeño delivers 1.5 to 1.9 times more AI work per watt at peak throughput, 1.7 to 3.6 times lower end-to-end latency, and 2.1 to 4.1 times higher performance for highly interactive workloads compared to its reference system. On Kimi K2.5 1T specifically, Jalapeño achieved approximately 1.5 times higher peak performance per watt and 3.4 times lower end-to-end latency. OpenAI notes these are proprietary benchmark results and advises interpreting them within the context of its specific testing methodology and comparison systems.

The chip’s architecture addresses the distinct phases and bottlenecks of AI inference. During the prefill phase, the system processes user prompts through relatively compute-intensive operations. The decode phase generates responses token by token and is primarily constrained by memory bandwidth. A third challenge involves communication overhead, where processing units often idle while waiting for data to move between different processors or memory resources. To mitigate these issues, Jalapeño keeps critical model information, including the KV cache, physically close to where it is needed. OpenAI emphasizes that the networking layer is integral to the architecture, enabling workloads to remain within a tightly connected system and minimizing unnecessary data movement. This approach yields a more balanced accelerator capable of efficiently handling both prefill and decode tasks, which is particularly important given that large AI models can contain hundreds of billions or even trillions of parameters requiring distribution across multiple chips.

For conventional chatbots, faster inference translates directly to quicker response times. However, the efficiency gains are expected to have a substantially larger impact on AI agents, which execute multiple sequential actions. By optimizing hardware around the specific characteristics of modern language models and AI agents, OpenAI aims to generate more useful AI output while consuming less electricity and shortening user wait times.

Notably, OpenAI claims that artificial intelligence played a direct role in designing the processor itself. AI tools assisted engineers in exploring alternative implementations, compressing design and verification cycles, and optimizing arithmetic circuits. This accelerated workflow enabled the team to move from initial design to tapeout—the stage where a chip design is finalized and submitted for manufacturing—in just nine months. Following hardware completion, AI was also leveraged to optimize the accompanying software stack.

Despite Jalapeño’s capabilities, OpenAI explicitly stated it will continue deploying accelerators from Nvidia and other partners for both training and inference. Consequently, Jalapeño should be viewed as an incremental addition to OpenAI’s computing capacity rather than an immediate replacement for commercial GPUs. Strategically, however, the chip grants OpenAI greater control over a segment of its infrastructure stack, allowing future hardware to be more closely tailored to its proprietary models and workloads. Jalapeño represents the first generation of a multigenerational hardware roadmap, with second- and third-generation chips already in development. Before large-scale deployment, OpenAI plans to begin internal rollout by the end of 2026 while continuing production qualification, software development, and performance testing across additional models.