The moat everyone thought was 20 years deep just got crossed in 10 hours by an AI coding agent.
The Summary
- AI coding agents are now writing CUDA-equivalent software — startup Infinity recreated CUDA-like code for chip startup D-Matrix in 10 hours using AI agents
- CUDA took Nvidia decades to build and has been the true competitive advantage, not just the chips themselves
- Cloud giants (Google, Amazon, Microsoft) plus OpenAI, Anthropic, and DeepSeek are all building alternatives or using agents to close the software gap
- The shift from AI training to inference work weakens CUDA's lock on the market
The Signal
For 20 years, CUDA has been Nvidia's real moat. Not the chips. The software. CUDA is what made Nvidia's GPUs programmable for AI work. It included pre-built code for common tasks, debugging tools, and orchestration layers that let thousands of chips train models together. Competitors could build faster chips. They couldn't build CUDA.
That advantage is compressing fast. Jeremy Nixon, former Google Brain researcher and founder of AI software startup Infinity, says his team used AI coding agents to build CUDA-equivalent software for chip startup D-Matrix in 10 hours. Not 10 months. Not 10 weeks. 10 hours.
"One of Nvidia's biggest moats is being crossed."
The pressure comes from multiple directions:
- Cloud giants building custom chip software (Google TPUs, Amazon Trainium, Microsoft Maia)
- Foundation model labs (OpenAI, Anthropic) shipping models that generate system-level code
- DeepSeek combining coding agents with TileLang, their programming language built specifically to make AI software easier
This isn't just about Nvidia losing exclusivity. It's about the nature of moats in the agent era. Software advantages that took decades to build can now be replicated in hours if you have the right agents and the right training data. The code libraries, the optimization tricks, the debugging tools — all of it becomes reproducible once agents can write, test, and iterate at machine speed.
The timing matters because the industry is shifting from training to inference. Training massive models required CUDA's full stack. Inference work is less complex, runs on different hardware, and doesn't need the same depth of tooling. As more compute shifts to inference, CUDA's lock weakens further.
Key dynamics at play:
- AI coding agents commoditize what used to be durable advantages
- The shift to inference reduces technical complexity and CUDA's necessity
- First-mover advantage in software shrinks when agents can catch up in hours
Nvidia knows this. The company says developers now use CUDA's code libraries primarily to build applications, and that Nvidia itself uses AI coding agents to develop and maintain CUDA. Translation: even the moat-builder is using agents to defend the moat.
The Implication
If your competitive advantage is complex software, watch this closely. The agent economy doesn't just automate grunt work. It compresses timelines for building technical infrastructure that used to take years. Nvidia will be fine because they're adapting and because hardware advantages still matter. But the lesson is broader: moats built on accumulated code, process knowledge, and technical complexity are all getting shallower.
For founders building in AI infrastructure, this is both threat and opportunity. You can move faster than ever. So can everyone else. The question isn't whether agents can replicate your work. It's whether you're using agents faster than your competitors are.