Building on Borrowed Intelligence

The scene repeats itself in thousands of startups every night. A founder, lit by the blue glow of a monitor, isn’t checking sales figures or user engagement. They’re staring at the OpenAI or Anthropic status page, watching a little green dot. Their company’s entire value proposition—the clever summarization tool, the automated customer service bot, the code generator that actually works—depends entirely on that dot staying green. Their product isn't just hosted in the cloud; its very brain is rented, query by query.

This is the precarious architecture of the current AI boom. We celebrate the explosion of new applications, but we fail to see the dangerously concentrated foundation they all share. A handful of companies are building the foundational models, the core engines of reasoning that everyone else is plugging into. This creates a new kind of dependency, far deeper than relying on Amazon Web Services for storage. Startups are not just renting infrastructure. They are renting intelligence.

In the last tech cycle, building on a platform meant using someone else’s servers or payment processing. The core logic, the unique intellectual property that made a business defensible, was your own code. You controlled the repository. You could, with great effort, move it somewhere else. The dependency was real, but it was on commodity services.

Today, the core logic is often an API call. The unique intellectual property is a meticulously crafted series of prompts sent to a third-party model you have no control over. The "secret sauce" is no longer an algorithm but a recipe for coaxing predictable behavior from a black box owned by someone else. This is not a stable foundation for an ecosystem; it is a tower built on a trapdoor.

The trapdoor springs open when the model provider changes its terms. Imagine the billing alert that announces a 50% price hike on the specific model version your entire product is tuned for. Your unit economics, painstakingly calculated and presented to investors, evaporate in an instant. Or consider the email announcing that a model is being deprecated. The feature your customers love will simply cease to function. The fix isn't a simple code patch. It is a desperate, months-long scramble to re-engineer your whole system around a new model with different quirks, biases, and failure modes.

The common retort is that companies can just switch providers. This betrays a fundamental misunderstanding of the technology. Moving from one foundational model to another is not like swapping a Ford engine for a Chevy. The prompts, the fine-tuning data, the guardrails—everything is bespoke. The subtle differences in how each model interprets a command can cascade into catastrophic product failures. The switching cost is immense, creating a vendor lock-in that makes the old database wars look quaint.

This is the real story, happening one API call at a time. While venture capitalists fund a thousand flowers to bloom, they are all being planted in a handful of proprietary gardens. The gardeners can change the soil, raise the price of water, or decide to stop letting people in at any moment. The most exciting applications of this new era may turn out to be little more than feature requests for the companies that own the underlying intelligence. The true power isn't in building the cleverest chatbot; it's in owning the brain it borrows.