OpenAI revealed new specifications and benchmarks for its Jalapeño processor, designed to accelerate artificial intelligence inference, at the Hot Chips conference. The first generation of the accelerator outperformed Nvidia’s Blackwell system in terms of power efficiency and speed, and took just 9 months from concept to release thanks to the use of AI.
OpenAI revealed new specifications and benchmarks for its Jalapeño processor, designed to accelerate artificial intelligence inference, at the Hot Chips conference. The first generation of the accelerator outperformed Nvidia’s Blackwell system in terms of power efficiency and speed, and took just 9 months from concept to release thanks to the use of AI.
The details of the tests and the architecture of the chip were reported by OpenAI representatives and the head of the company’s hardware department, Richard Ho. The testing was carried out using the InferenceX benchmark from SemiAnalysis on public models GPT-OSS 120B, DeepSeek R1 670B and Kimi K2.5 1T, as well as on OpenAI’s internal algorithms.
In testing on the Kimi K2.5 1T, the Jalapeño processor delivered 1,5x higher peak performance per watt and 3,4x lower end-to-end latency than the Nvidia Blackwell benchmark. The chip has a TDP of 700W, but under real-world workloads, its stable power consumption remained at or below 550W.
The chip’s architecture is specifically designed for the specifics of language models and interactive AI agents. Jalapeño is optimized for two key stages of inference: computationally intensive pre-filling and memory-limited decoding. Thanks to local storage of model state and a key-value (KV) cache, the developers were able to minimize data movement and network latency.
«The bottom line is that the results show a very significant leap in performance compared to state-of-the-art technologies. Jalapeño can process more AI tasks per unit of power, while providing faster responses. The system is very efficient at serving a large number of clients with low latency,» said Richard Ho.
The chip was created with the direct participation of artificial intelligence, which helped optimize arithmetic schemes and shorten the development cycle. In addition, engineers used Codex and GPT-Astra to adapt three large models with open scales to the new architecture in two months.
Despite developing its own hardware, OpenAI emphasizes that it does not plan to abandon external suppliers — the company will continue to purchase accelerators from Nvidia and other partners for training and inference.
The first version of Jalapeño will be deployed on OpenAI’s infrastructure in small quantities in late 2026, with larger-scale deployment expected in 2027. It is reported that this is just the beginning of a multi-generational platform: the second generation (Gen 2) is already in active development, and conceptual details are being formed for Gen 3.
The Jalapeño project was first announced in October last year, and details emerged in June 2026. OpenAI is developing the processor in close collaboration with Broadcom, using its own AI models for both design and programming of the chips.
