OpenAI has released the first detailed performance results of its custom artificial intelligence chip, Jalapeño, showing significant gains in speed, power efficiency and response times. The company says the chip could help it serve more AI users while reducing the amount of power and hardware needed to deliver responses.
The results mark an important step in OpenAI’s push to build more of its own AI infrastructure instead of relying entirely on third-party processors. Jalapeño was developed in collaboration with Broadcom and is designed specifically for AI inference, the stage at which trained AI models process requests and generate answers.
OpenAI tested Jalapeño using InferenceX, a public benchmark from SemiAnalysis that measures the performance of AI systems while serving real requests. The company compared its chip with leading commercially available systems across different workloads, from high-throughput applications to highly interactive, low-latency tasks.
According to OpenAI, Jalapeño delivered between 1.5 and 1.9 times more AI work per watt at peak throughput than the comparison systems. It also recorded 1.7 to 3.6 times lower end-to-end latency. For highly interactive workloads, performance was between 2.1 and 4.1 times higher.
The testing covered three large language models: GPT-OSS 120B, DeepSeek R1 670B and Kimi K2.5 1T. OpenAI said Jalapeño performed strongly across all three, suggesting that its architecture can handle models developed both inside and outside the company.
The chip is rated at 700 watts, although OpenAI said its measured sustained power remained at or below 550 watts during the workloads tested. On Kimi K2.5 1T, one of the largest public models included in the test, Jalapeño delivered about 1.5 times higher peak performance per watt and 3.4 times lower end-to-end latency than the comparison system.
The significance of Jalapeño goes beyond simply making a faster processor. The company has designed the chip, memory, networking, software and rack-scale system together around the way modern AI models actually work.
AI inference has several stages. During prefill, a system processes the user’s prompt and is largely constrained by computing power. During decode, it generates the response token by token, making memory bandwidth particularly important. Moving information between chips can also introduce delays.
Jalapeño is designed to reduce those bottlenecks. OpenAI said it keeps important model data, including the KV cache used during response generation, closer to where it is needed. This reduces data movement and communication delays, allowing computing, memory and networking resources to work together more efficiently.
Another notable part of the project is the role played by AI itself in developing the chip. OpenAI said its AI models helped engineers move from initial design to tapeout in nine months by exploring implementations, speeding up design and verification and improving arithmetic circuits.
The company also used Codex with GPT-Astra to optimise software for three open-weight models that were not part of Jalapeño’s original production plan. OpenAI said selected AI-generated implementations for GPT-OSS attention and mixture-of-experts components ran 1.5 to 1.8 times faster than existing implementations written by human experts. The company clarified that these results apply to selected components rather than the complete models.
The move comes as demand for AI computing continues to grow. Faster and more efficient inference is increasingly important as users rely on AI assistants, coding tools and agentic systems that can perform multiple steps in sequence. Even small reductions in response time can become significant when an AI agent has to complete several actions.
OpenAI believes Jalapeño could ultimately help it deliver faster AI responses, support more users and lower the cost of running AI services. The company says better performance per unit of power is particularly important because AI infrastructure requires large amounts of electricity and computing capacity.
OpenAI plans to begin deploying Jalapeño inside its own computing infrastructure by the end of 2026. However, wider deployment is expected in 2027, with the company describing the current chip as the first generation of a broader hardware roadmap. Gen 2 is already deep in development, while Gen 3 is in its early design stage.
Importantly, OpenAI does not intend to abandon other chip suppliers. The company said it will continue widely deploying Nvidia accelerators and hardware from other partners for both AI training and inference. That means Jalapeño is better understood as an additional pillar of OpenAI’s AI infrastructure strategy rather than an immediate replacement for Nvidia hardware.
The first Jalapeño results therefore offer a glimpse of how OpenAI wants to approach the next phase of AI computing: by controlling more of the technology stack and tailoring hardware specifically for increasingly demanding AI workloads. If the performance gains hold up during large-scale deployment, the custom chip could play a significant role in making AI services faster, more efficient and potentially cheaper to operate.
