The AI industry has consolidated around a new approach: build one massive, general-purpose model and then adapt it for specific tasks. These "foundation models" - GPT, Claude, Gemini, Llama, and others - are trained on enormous datasets at enormous cost, then fine-tuned, prompted, or otherwise adapted for particular applications. This is a fundamentally different paradigm from a decade ago, when you'd build a separate model for each task. The economics are striking: training a frontier foundation model can cost hundreds of millions of pounds, which means only a handful of organisations can afford to do it. Everyone else builds on top of their work. This creates a supply chain that didn't exist before - foundation model providers at the top, application developers in the middle, and end users at the bottom. The implications are significant: if you're building an AI-powered product, you're almost certainly dependent on someone else's foundation model, which means their decisions about pricing, access, and capabilities directly affect your business. Understanding this dependency is essential for anyone making strategic decisions about AI.