The frontier AI labs—OpenAI and Anthropic—have a narrow competitive moat: they're maybe one or two models ahead of open source alternatives and the rest of the world. This advantage comes from talent, compute resources, and crucially, access to user data (coding traces, conversations, work patterns) that trains the next generation of models. But this moat is fragile. China's labs are closing the gap, and the data advantage won't last forever. As cheaper models get good enough, companies are already pulling back spending and redirecting budget to dirt-cheap alternatives. The labs are trying to escape this trap by expanding into adjacent markets—OpenAI into ads and robotics, Anthropic into biotech—but building $100B revenue streams in new industries is brutally hard. Most industries don't even have businesses that size, and the broader market keeps moving. The real pressure comes from spending discipline hitting companies now. CFOs are asking hard questions: what's the actual return on this new AI line item? Employees and shareholders love the "AI-first" narrative, but when the income statement doesn't show results, the spending gets cut. This is already happening—companies restricting top-tier models to key users, pivoting to cheaper alternatives. The frontier labs have high margins on inference and are genuinely good at serving models cheaply, but competitors like Deepseek are catching up. If the labs can't maintain a meaningful performance gap, they're stuck competing on margins like any cloud provider—still profitable, but no longer "building god." The only path that keeps the labs dominant is if they can either lock people out from using competing models through regulation (treating AI like a controlled import, restricting model access), or if they achieve recursive self-improvement—automated AI researchers that can improve themselves without human intervention. The regulatory approach is nearly impossible to pull off; you can't physically control software files the way you control hardware chips. The recursive self-improvement angle is what both labs are actually chasing. That's the real prize: a system that generates its own successors and automates everything else. Without that breakthrough, the frontier advantage erodes into a commodity business.