OpenAI considers releasing internal chip abstraction software...supports running AI across multiple chips

OpenAI is considering open-sourcing or broadly sharing an internal chip-abstraction layer that lets its AI run across different server chips without hardware-specific tweaks.

Infrastructure chief Sachin Katti said OpenAI is building software to optimize agentic AI across chip types and wants it to be usable more widely.

The company has recently added chip agreements with Amazon, Cerebras, and AMD, while also developing its own AI chip.

OpenAI has already secured samples of Nvidia’s upcoming Vera Rubin chip for future training.

If released, the software could make AI workloads easier to move across hardware and weaken Nvidia’s CUDA lock-in.