The industry's power structure is visibly realigning around data control, not model capability. Anthropic's leaked Claude Code source code and the LiteLLM supply chain breach expose how fragile proprietary moats have become when humans remain in the chain, while OpenAI's $122 billion valuation and $3 billion retail raise signal that capital markets now price AI companies on deployment scale and user lock-in rather than technical superiority. The real leverage belongs to whoever controls the data flowing through systems, the infrastructure hosting it, and the switching costs that keep users captive once they arrive.
Anthropic's month of self-inflicted wounds tells a story about organizational immaturity at scale. A source map file exposure of 512,000 lines of code is not a sophisticated attack; it is an employee mistake that competitors and security researchers will spend weeks studying. Yet the same company published a 2023 study measuring AI's "theoretical capabilities" in the job market based on assumptions about "anticipated LLM-powered software" that has not materialized. The gap between what Anthropic claims to know and what it actually controls keeps widening. Meanwhile, OpenAI raised capital at $852 billion valuation while still private, with Amazon, Nvidia, and SoftBank leading the round. That valuation reflects not breakthrough reasoning or coding but rather the installed base of ChatGPT users, the switching costs embedded in enterprise deployments, and the belief that an IPO will unlock retail capital for years to come.
The infrastructure and application layers are where actual value is migrating. Runway's $10 million fund for startups building on its video models, Nomadic's $8.4 million to structure autonomous vehicle footage into datasets, and Ring's new app store betting on AI use cases beyond security all point toward a shift from generic model performance to specialized applications that own customer relationships and data. Microsoft's multi-model Critique and Council system for Copilot Researcher, Salesforce's 30 new AI features for Slack, and Amazon's Alexa+ food ordering integration are not about model breakthroughs; they are about embedding AI into workflows where switching costs are highest and data collection is continuous. The companies winning are those locking users into platforms where AI becomes infrastructure, not those publishing benchmarks.
Sloane Duvall