Nvidia’s grip on AI computing has long rested on two pillars: its hardware and its software. The GPUs get all the attention, but the real lock-in comes from CUDA, the proprietary programming platform that millions of developers have built their workflows around. OpenAI is now taking a direct shot at that second pillar.
The company is preparing to release a tool designed to let AI models run on non-Nvidia hardware, leveraging its open-source Triton language as a viable alternative to CUDA.
Triton’s quiet evolution
Triton isn’t new. OpenAI first released it back in July 2021 as an open-source language for writing high-performance GPU kernels in Python. The pitch was straightforward. CUDA is powerful but notoriously complex. Triton aims to deliver comparable performance with code that’s far more accessible to the average developer.
Since then, the project has been steadily gaining traction. It now serves as a backend for popular frameworks like PyTorch. The latest version, Triton 3.7, was released in 2026, signaling that OpenAI isn’t treating this as a side project.
The hardware diversification play
OpenAI’s software push doesn’t exist in a vacuum. The company has been actively exploring alternatives to Nvidia chips since 2025, driven in part by dissatisfaction with some of Nvidia’s inference chips. Inference is the process of actually running a trained AI model, as opposed to training it in the first place.
The company announced a partnership with AMD that includes a substantial 6GW of AMD-powered compute capacity. OpenAI has described this as incremental to its existing Nvidia engagements, not a replacement.
Beyond AMD, OpenAI has been in conversations with startups like Cerebras and Groq, both of which have designed specialized chips optimized for inference workloads. And the company is pursuing custom AI inference chips with Broadcom, with production plans noted for 2026.
What this means for investors
Nvidia’s CUDA ecosystem has millions of developers, years of institutional knowledge, and deep integration into virtually every major AI framework.
AMD has been enhancing its ROCm platform to improve compatibility with AI workloads. Open projects like ZLUDA have emerged to translate CUDA code to run on non-Nvidia hardware. And now the single largest consumer of AI compute on the planet is actively building tools to make Nvidia’s software advantage less relevant.
For AMD and the broader alternative chip ecosystem, OpenAI’s moves represent a potential inflection point. The biggest barrier to adoption for non-Nvidia hardware has always been software compatibility. If Triton matures into a genuine cross-platform standard, it removes the single largest objection that AI developers have when considering AMD or custom silicon.