OpenAI is building the software infrastructure to run AI workloads across chips from multiple providers, a strategic effort to reduce dependency on Nvidia’s CUDA ecosystem.
The initiative focuses on low-level runtime systems and compiler technology that allow AI models to operate efficiently on hardware from various vendors.
In January 2026, OpenAI signed a deal with Cerebras for up to 750 MW of compute capacity dedicated to inference. In October 2025, OpenAI announced a multiyear collaboration with Broadcom to develop custom AI accelerators targeting 10 GW scale, with mass production expected in 2026.
AMD is also involved, with reports of multi-gigawatt deals for MI450-class accelerators. OpenAI is exploring integration with Google TPUs and AWS Trainium, positioning its infrastructure across nearly every major chip architecture.
The software layer addresses CUDA’s lock-in by abstracting hardware differences. OpenAI is hiring engineers for compute roles focused on building platform structures for diverse hardware environments.
For Nvidia, this shifts negotiating dynamics as OpenAI can threaten to shift workloads. For AMD and Broadcom, it creates market opportunities. Cerebras gains validation from OpenAI’s commitment. The risk is execution complexity-performance penalties from heterogeneous hardware optimization could offset cost savings.