OpenAI launches Jalapeño chip to make artificial intelligence run faster

The development of its own hardware marks a new era for OpenAI, where they want to control not only the models and software themselves, but also the processor cores, memory, and network infrastructure. The chip, called OpenAI Jalapeño, is designed to eliminate the classic dilemma of running AI models: choosing between high throughput and low latency.

In measurements conducted using the open-source software tool InferenceX, the OpenAI Jalapeño chip's performance was compared to existing commercial solutions from Nvidia, specifically the Nvidia Blackwell platform. Testing on publicly available models such as the OpenAI GPT-OSS 120B, DeepSeek R1, and Moonshot AI Kimi K2.5 showed that the newcomer achieves 1.5 to 1.9 times better efficiency per watt of electricity consumed at peak throughput. At the same time, the overall latency was reduced by 1.7 to 3.6 times. The aforementioned advantages were even more pronounced for highly interactive tasks.

Although the OpenAI Jalapeño is officially rated at 700 W, measurements showed that actual power consumption under load remained at or below 550 W. For developers, the key metric is primarily the amount of work done per unit of energy, not just raw power. When processing data, the new accelerator demonstrated a high number of processed tokens per kilowatt of energy compared to the current leading systems on the market.

The chip's architecture is tailored to the specific stages of language model processing, where the requirements for computing power and memory bandwidth change rapidly. The new system was, of course, developed as a whole, connecting the processor, memory and network at the level of entire server racks. Another interesting fact is that OpenAI used its own artificial intelligence tools to design the chip, which shortened the development cycle to just nine months. Artificial intelligence helped optimize even individual mathematical circuits and accelerated the adaptation of code for new models.

Of course, the data currently comes mainly from controlled tests, and the actual stability and cost-effectiveness of the system will only be demonstrated when it is used in a mass production environment. OpenAI also does not plan to completely abandon existing partners. They will continue to use equipment from Nvidia, AMD, Microsoft, Amazon AWS and others, as they want to maintain a diversity of sources.

The start of integrating the OpenAI Jalapeño chip into its own data centers is planned by the end of 2026. At the same time, OpenAI is already intensively planning the second and third generations of this hardware, with the aim of ensuring even higher speed and lower costs of implementing its services in the long term.