AI Agents Could Build CUDA Rivals, But Nvidia Isn't Standing Still

Nvidia's advantage lies not only in its GPU accelerators but also in its CUDA software platform, which includes ready-made code for common AI tasks, debugging tools, and applications that enable thousands of chips to work together during model training. The deep entrenchment of CUDA in millions of lines of code and workflows creates a strong lock-in effect, making alternatives costly and difficult to adopt. For instance, internal Amazon documents cite CUDA as the main obstacle to adopting its AI chips Trainium and Inferentia. However, some experts argue that AI can automate the creation of system software, a historically hard task. Jeremy Nixon, former Google Brain researcher and founder of AI software startup Infinity, said his company used AI agents to build CUDA-like software for D-Matrix in just 10 hours, demonstrating that this barrier to leaving Nvidia is surmountable. OpenAI and Anthropic have also shown AI models capable of generating system software, and DeepSeek founder Liang Wenfeng said AI agents, along with his startup's own programming language TileLang, have significantly simplified AI software creation. Nvidia counters that developers increasingly use CUDA code libraries for AI applications and that it also uses software AI agents to speed up CUDA development and verification. The company notes that as AI shifts to inference and agentic workloads, the need for deep, full-stack optimization grows, increasing the value of its integrated hardware and software. However, Bing Xu, founder of AI software startup INT21, said CUDA's mature ecosystem of verification tools and other features helps programming agents work more efficiently, making it the next defensive barrier. Agents can generate lots of code quickly, but verification is the bottleneck, he said. While AI may help competitors catch up, Nvidia benefits from the same technological advances, Xu added, noting the chipmaker is not sleeping. Modular CEO Chris Lattner noted that writing code is only a small part of software development compared to harder tasks like optimization for industrial use.