OpenAI runs the model on a non-Nvidia chip for the first time

GPT-5.3-Codex-Spark is a stripped-down, energy-efficient version of the Codex family, optimized for fast programming and seamless interoperability. The model is currently available as a test for ChatGPT Pro subscribers.

This is the first time the company that owns ChatGPT has put a model into actual operation on a non-Nvidia chip platform. The hardware behind this change is Cerebras’ third generation Wafer Scale Engine. Unlike a conventional GPU cluster assembled from many small chips, Cerebras’ architecture uses a giant silicon slab containing hundreds of thousands of AI processing cores, helping to minimize latency and maximize data transfer speed, which is a weakness in real-time response tasks.

OpenAI said that when running on Cerebras’ system, GPT-5.3-Codex-Spark can reach processing speeds of more than 1,000 tokens per second. The model is fine-tuned to help programmers edit specific pieces of code or run targeted tests quickly.

Although it begins to diversify hardware, OpenAI still affirms its dependence on Nvidia. In early February, CEO Sam Altman wrote on X that the company “loves working with Nvidia” and called it the best chipmaker in the world. Altman’s move quells rumors that OpenAI is dissatisfied with the performance of some Nvidia chip lines. In fact, OpenAI’s most powerful models are still trained and operated primarily on Nvidia’s infrastructure.

Joining hands with Cerebras is part of OpenAI’s plan to expand the hardware ecosystem. Last month, the company confirmed signing an agreement to deploy 750 MW of computing infrastructure with Cerebras lasting until 2028. In addition, OpenAI has also agreed to buy the equivalent of 6 GW of chips from AMD over many years, and teamed up with Broadcom to design its own AI accelerators and custom networking components.

The Codex model line has now reached over one million weekly active users. OpenAI plans to expand access to GPT-5.3-Codex-Spark to a broader group of users after thoroughly evaluating performance and market demand.