OpenAI Cross-Chip Tool Challenges NVIDIA CUDA

According to recent industry reports, OpenAI is developing a new software system designed to enable artificial intelligence workloads to run seamlessly across heterogeneous hardware platforms, including chips from NVIDIA, AMD, and Google. Once officially released, the tool could pose a significant long-term challenge to NVIDIA’s widely adopted CUDA software ecosystem.

Citing The Information, OpenAI’s Head of Compute and Infrastructure, Sachin Katti, stated that the company is building software that enables AI computation to switch across different hardware architectures without requiring developers and product teams to account for underlying chip differences. He referenced Google’s Borg system as an architectural inspiration, noting that the AI industry is gradually moving toward a more hardware-agnostic computing model.

Katti also indicated that discussions are underway regarding whether the software should be open-sourced or publicly released. While OpenAI has expressed interest in making the technology broadly available, no formal timeline has been established.

Since its launch in 2006, NVIDIA’s CUDA platform has become a foundational pillar of GPU computing, combining compilers, libraries, and optimization tools to create a deeply entrenched software ecosystem. However, Katti suggested that advances in AI-driven automation may gradually weaken such ecosystem lock-in. In particular, the automatic generation of optimized kernels for multiple hardware architectures could significantly improve cross-chip compatibility and reduce dependency on a single vendor’s software stack.

Efforts to challenge CUDA’s dominance are not limited to OpenAI. As early as 2016, Meta introduced the PyTorch framework, which enables AI workloads to run across a wide range of hardware platforms. In addition, several startups have developed translation tools that convert PyTorch code into instructions optimized for underlying chip architectures.

According to Reuters, Google is also advancing an internal initiative known as “TorchTPU,” aimed at improving compatibility between its in-house Tensor Processing Units (TPUs) and the PyTorch ecosystem. Google is also considering open-sourcing parts of the project to accelerate TPU adoption in enterprise markets, with technical support from Meta, the original creator of PyTorch.