OpenAI is expanding beyond AI model development into hardware, with the company designing a custom chip called Jalapeo to run AI models more efficiently. Developed in collaboration with Broadcom, the chip is aimed at delivering higher AI performance while consuming less power and reducing response times.

In a blog post, OpenAI said Jalapeo was built specifically for AI inference, the process through which a trained model processes a request and generates a response. The company tested the chip using GPT-OSS 120B, DeepSeek R1 and Kimi K2.5 1T, and said it was able to combine high performance with low latency.

Jalapeo Focuses On AI Inference

Inference takes place whenever an AI system responds to a user’s request. From answering questions in ChatGPT to powering AI coding agents, the process requires hardware to repeatedly process model workloads and generate outputs.

OpenAI said Jalapeo was designed specifically around these requirements rather than as a general-purpose AI accelerator. The company developed the chip with Broadcom, while Celestica worked on the boards, rack systems and production hardware.

OpenAI said the chip was designed alongside the software, memory, networking and serving systems used with its models. “By co-designing the chip, software, memory, networking and serving systems around our models, we can improve performance and efficiency across the entire stack,” the company said.

The company’s approach is aimed at addressing a central challenge in AI infrastructure: increasing computing performance without adding to the time users have to wait for a response.

“Jalapeo is designed to deliver high performance and low latency at the same time,” OpenAI said.

Chip Shows Higher Performance Per Watt

OpenAI said Jalapeo delivered between 1.5 and 1.9 times more AI work per watt than the comparison systems across the three models tested. The company also reported end-to-end latency reductions ranging from 1.7 to 3.6 times.

For workloads that require highly interactive responses, OpenAI said Jalapeo’s performance was up to 4.1 times higher.

The company said the improvements could have practical implications for large-scale AI services. Processing more requests with lower electricity consumption, while reducing response times, could translate into significant efficiency gains for systems handling millions of requests.

The focus on inference also reflects the growing importance of the computing infrastructure required to serve AI models after they have been trained. Rather than concentrating solely on developing increasingly capable models, OpenAI is seeking to optimise how those models operate once they are deployed at scale.

OpenAI Seeks Greater Control Of AI Infrastructure

Jalapeo forms part of OpenAI’s broader effort to control more of the technology stack supporting its AI systems. The company is seeking to optimise hardware and software together instead of developing models independently of the infrastructure used to run them.

OpenAI said Jalapeo went from the initial design stage to manufacturing tape-out in nine months. Its own AI models were also used to assist engineers during parts of the chip’s design and optimisation process.

However, the development does not signal an immediate move away from Nvidia or other hardware suppliers. OpenAI plans to deploy Jalapeo in small volumes by the end of 2026 and expand its use through 2027.

The company also intends to develop future generations of the chip while continuing to work with hardware partners including Nvidia as part of its wider computing strategy. Jalapeo therefore represents an expansion of OpenAI’s hardware capabilities rather than an immediate replacement for the company’s existing infrastructure.

Long or Short, get news the way you like. No ads. No redirections. Download Newspin and Stay Alert, The CSR Journal Mobile app, for fast, crisp, clean updates!

App Store –  https://apps.apple.com/in/app/newspin/id6746449540 

Google Play Store – https://play.google.com/store/apps/details?id=com.inventifweb.newspin&pcampaignid=web_share