AI is starting to rewrite the software that made Nvidia untouchable

Nvidia's CUDA software, which has been a crucial part of the company's success, is now facing potential challenges as artificial intelligence (AI) is beginning to write its own code. The software, shorthand for Compute Unified Device Architecture, was developed by Nvidia executive Ian Buck and has been instrumental in turning the company's chips into the building blocks of AI.

It boasts ready-made code for common AI tasks, tools to find bugs, and software that allows thousands of chips to work together to train models. However, AI is now starting to rewrite this software, which could weaken CUDA's long-standing dominance. One founder of an AI software startup claimed that AI coding agents could recreate CUDA-like software in just 10 hours, suggesting that Nvidia's moat could be crossing.

Cloud giants like Google, Amazon, and Microsoft have also been building software around their own AI chips, while AI models capable of generating system software have emerged from companies like OpenAI and Anthropic. DeepSeek founder Liang Wenfeng recently stated that AI coding agents and his startup's programming language, TileLang, have made AI software substantially easier to build.

Although the shift from software development to AI inference could change the equation, Nvidia's tightly integrated hardware and software offerings are becoming more valuable as AI models are put to work. However, some analysts question Nvidia's CUDA advantage, citing the stagnant stock price of the company over the past year. While coding agents may make it easier to generate software, experts believe that AI-generated code still needs to be verified and optimized, potentially becoming CUDA's next moat.