OpenAIs Next Battle With Nvidia Is Not About Chips

OpenAI is considering releasing software that would make it easier to run AI on chips from any manufacturer — and if it does, Nvidia has the most to lose.

The Information reported Monday that OpenAI is weighing a public release of internal tooling its infrastructure team built to run AI workloads efficiently across hardware from Nvidia, AMD, and Google. OpenAI's head of compute infrastructure, Sachin Katti, a former Intel AI executive who joined OpenAI in 2025, is leading the effort, according to sources cited by the outlet.

The strategic weight of that hire is not subtle. Katti spent years inside Intel's AI hardware organization. He knows Nvidia's defensive architecture better than almost anyone now sitting inside a company with every incentive to dismantle it.

Nvidia's real advantage is not raw chip performance. It is CUDA, the software layer that sits between AI models and the hardware they run on. Over two decades, Nvidia built CUDA into the default environment where AI developers write, optimize, and deploy their work. PyTorch — the most widely used AI framework — is deeply integrated with CUDA. Moving a workload off Nvidia chips means rewriting code for a different software stack, eating into the performance gains from whatever hardware you are moving to. That friction is Nvidia's moat.

OpenAI's move, if it materializes, would reduce that friction directly. And OpenAI is not working in isolation. It signed a deal with AMD to deploy up to 6 gigawatts of AMD Instinct GPUs starting in the second half of 2026. It partnered with Broadcom in October 2025 to co-develop custom AI accelerators — a 10 gigawatt commitment — with chips targeted for the same window. Its hardware team is working with TSMC on chiplet architectures and custom inference chips, also for 2026 production. The common thread across all of these deals is the same problem: running OpenAI's software on non-Nvidia hardware without starting from scratch.

Here is what gets interesting. OpenAI already released a tool aimed at exactly this problem. Triton is an open-source GPU programming language the company released in 2021. It lets developers write efficient GPU code without touching CUDA — and it runs on AMD chips as well as Nvidia's. The GitHub repository has 6,291 commits, 19,300 stars, and active development. A recent commit from an OpenAI-affiliated contributor added AMD-specific optimizations to the Gluon matrix-multiplication kernel, improving inner-loop efficiency by 12 to 20 percent depending on the variant. AMD buffer-offset handling got similar treatment.

So is Triton the internal tool The Information's sources are describing? Possibly. The distinction matters. Triton is already public, already maintained, and already being extended — which would make the story less about a secretive new weapon and more about a bet on what the existing weapon can become. OpenAI declined to comment for this article.

The question the wire coverage leaves open is whether OpenAI is planning to release something that builds on Triton in ways that are not yet public — a higher-level abstraction layer, a deployment tool, something that makes cross-vendor clusters as easy to manage as a homogeneous Nvidia setup — or whether the internal effort is already substantially captured by what is on GitHub.

Katti did not respond to a request for comment. Nvidia declined to comment.

The market is not yet convinced the release happens. A prediction market on whether OpenAI publicly shares the software in 2026 currently sits at 29 percent. That uncertainty is the honest answer. The tool may be vapor. The partnerships may not produce working silicon on schedule. OpenAI remains heavily dependent on Nvidia for its most critical training workloads. Souring that relationship carries real cost.

But the direction of travel is clear. OpenAI is building, buying, and partnering its way off single-vendor dependency at a scale that makes the CUDA moat question not academic. If someone cracks it — whether through Triton, through something on top of it, or through something else entirely — the rippling effect on how AI infrastructure gets priced and built is substantial.