OpenAI is expanding beyond AI model development and moving into custom hardware as well. Its latest project, Jalapeo, is a specialised chip designed to run AI models more efficiently and was developed in collaboration with Broadcom.
In a blog post, OpenAI said Jalapeo is intended to handle more AI workloads while consuming less power and reducing response times. The company tested the chip with models including GPT-OSS 120B, DeepSeek R1 and Kimi K2.5 1T.
OpenAI said one of the key challenges in AI hardware is delivering high performance without increasing latency. Jalapeo has been designed specifically to address both requirements, with the company describing it as a chip capable of combining high performance with low latency.
What makes Jalapeo different?
The chip is focused on inference, which refers to the process of a trained AI model receiving a request and generating a response. Whenever someone asks ChatGPT a question or uses an AI coding assistant, inference is taking place.
Jalapeo has been developed specifically for this workload. OpenAI worked with Broadcom on the chip, while Celestica contributed to the boards, rack systems and production hardware. According to OpenAI, the chip was designed around the way its AI models operate.
The company said that jointly designing the chip, software, memory, networking and serving infrastructure around its models allows it to improve efficiency and performance throughout the technology stack.
In testing across the three models, OpenAI said Jalapeo delivered between 1.5 and 1.9 times more AI work per watt than the comparison systems. It also achieved 1.7 to 3.6 times lower end-to-end latency. For highly interactive workloads, performance was reportedly as much as 4.1 times higher.
In practical terms, the goal is to process more AI requests using less electricity while reducing the time users wait for responses. At the scale of millions of AI queries, even relatively small efficiency gains could translate into substantial cost savings.
Is OpenAI moving away from Nvidia?
Jalapeo is part of OpenAI’s broader effort to gain greater control over the technology powering its AI services. Rather than developing models and relying entirely on third-party hardware, the company is looking to optimise chips, software, memory, networking and serving infrastructure as a single system.
OpenAI said Jalapeo went from the initial design stage to manufacturing tape-out in just nine months. Its own AI models were also used to assist engineers during parts of the chip’s development and optimisation.
However, Jalapeo is not an immediate replacement for Nvidia hardware. OpenAI plans to begin deploying the chip in limited quantities by the end of 2026 and expand its use during 2027. The company also intends to develop future generations while continuing to work with hardware partners such as Nvidia.
A useful next step would be to compare Jalapeo with Nvidia’s latest AI chips to understand where OpenAI’s custom approach could have a real advantage.
