Nvidia’s grip on AI hardware is well documented. The company commands roughly 86% of data-center GPU revenue. But its real moat has never been the chips themselves. It’s been CUDA, the software ecosystem that makes developers so dependent on Nvidia hardware that switching feels like learning a new language while skydiving.
OpenAI is betting it can hand everyone a parachute. The company’s open-source tool called Triton, first released in July 2021, is being positioned as the key to running AI models on non-Nvidia hardware with minimal code changes.
From research project to strategic weapon
Triton started life as a relatively modest project. Its original purpose was letting developers write high-performance GPU code in Python instead of wrestling with low-level CUDA programming.
The tool has evolved considerably since version 1.0 dropped in mid-2021. Analysis from early 2026 points to Triton reaching something of an inflection point, where it now enables porting AI models across different hardware platforms with minimal or even zero code rewrites.
OpenAI isn’t just building tools in a vacuum, either. The company entered a multi-year agreement with AMD in October 2025 to deploy up to 6 gigawatts of Instinct GPUs. The first wave, 1 gigawatt of MI450 series chips, is expected to arrive in the second half of 2026.
Following the hiring trail
OpenAI is actively hiring inference engineers specifically focused on AMD GPU enablement. Reports from 2026 also indicate that OpenAI has expressed dissatisfaction with certain Nvidia chips.
What this means for investors
Nvidia’s 86% share of data-center GPU revenue isn’t going to evaporate overnight. CUDA has decades of accumulated optimization and a developer ecosystem that runs deep.
AMD stands to benefit most directly. The company already has competitive silicon, and the OpenAI partnership validates its AI hardware ambitions in a way that no benchmark ever could. When the biggest AI company in the world commits to deploying gigawatts of your GPUs, it sends a clear message to the rest of the industry that there’s a viable alternative to Nvidia.
The risk to watch is execution. Building a tool that theoretically runs on any hardware is one thing. Making it perform at parity with CUDA-optimized code on Nvidia’s own chips is another challenge entirely.