Your AI Feature Runs on a Handshake Deal

The Slack channel was full of rocket emojis. A small team at a well-funded startup had just pushed a new feature, a clever AI assistant that summarized complex legal documents. It worked flawlessly. It felt like magic. Then the email landed in the CTO’s inbox, subject line: “Updates to our API Usage Policy.” The magic vanished. The new rate limits and a 40% price hike for their preferred model meant the feature, their core differentiator, was now unprofitable.

This is the quiet terror stalking thousands of companies in the AI boom. We used to talk about platform risk—the danger of building a business on Facebook’s social graph or Apple’s App Store. That was a risk of access and distribution. This is different. This is model risk. It’s the existential gamble of building your company’s core intelligence on a foundation you don’t own, can’t inspect, and can’t control.

A handful of labs—OpenAI, Anthropic, Google—have become the de facto landlords for the next generation of software. Their large language models are the ground upon which everyone else is building. But this ground is not solid. It shifts. An update pushed for "safety" reasons can silently lobotomize the specific capability your product relied on. A model that was brilliant at creative writing yesterday might be neutered into a corporate drone tomorrow. There is no changelog for its soul.

This makes the relationship fundamentally different from using AWS or any other cloud provider. Amazon doesn’t change the performance of an m5.large instance based on its corporate mood. The laws of computation are stable. The "laws" of a proprietary AI model are not. They are a black box, subject to unannounced tweaks, philosophical realignments, and sudden price changes that feel less like enterprise software billing and more like a utility company discovering you’re running a server farm in your basement.

The venture capital is flowing to companies that are, in essence, sophisticated tenants. They build beautiful apartments on rented land. They design clever workflows, intuitive user interfaces, and smart integrations. But the core asset, the reasoning engine itself, is leased. And the lease terms are written in pencil.

What happens when a dominant model provider decides to compete directly with its most successful customers? We saw this story play out in the mobile ecosystem. But it’s more acute now. When your entire product is a thin, albeit clever, layer over a single API, you aren’t just facing competition. You are facing obsolescence. The landlord can simply build a prettier apartment in your building and offer it for free.

The stakes here are not just about a few startups failing. This is about a dangerous consolidation of cognitive power. The ability to reason, to summarize, to create—these capabilities are being centralized in the hands of a few unaccountable organizations. The thousands of flowers blooming are all potted plants drawing from the same few spigots of water. If those spigots run dry or the water is poisoned, the entire garden withers. The real innovation, the hard-won moat, may not be the clever application, but the painful, capital-intensive work of owning the intelligence itself. For everyone else, they’re not building a business. They’re just waiting for the next email.