Don’t Build on Rented Land
In the AI gold rush, everyone is building. Wrappers around GPT. Chrome extensions powered by Claude. Agents that stitch together three APIs and call it a product.
The energy is real. The risk is also real: most of these builders are constructing on rented land.
The Landlord Problem
When you build a product on top of someone else’s model, API, or platform, you inherit a dependency you can never fully control. The landlord can:
- Raise the rent. OpenAI has changed pricing multiple times. What was viable at $0.01 per 1K tokens may not survive at $0.03.
- Move in next door. The platform sees your traction, ships the same feature natively, and you’re competing against your own foundation.
- Change the locks. API deprecations, policy shifts, rate limit changes — any of these can break your product overnight.
This isn’t hypothetical. It’s the pattern we’ve seen with every platform shift: Facebook apps, Twitter bots, Shopify plugins, Chrome extensions. The ones who built on the platform thrived — until the platform decided they shouldn’t.
What Counts as “Your Own Land”?
If the model is rented, what do you actually own? The answer matters more than most founders think.
Domain knowledge that’s hard to replicate. Not “we fine-tuned a model” — anyone can do that. But deeply encoded understanding of a specific industry’s workflows, regulations, edge cases, and pain points. The kind of knowledge that takes years to accumulate and can’t be scraped from the internet.
Proprietary data loops. Every user interaction that makes your product smarter in a way the underlying model can’t replicate on its own. If your product gets better with use and that improvement stays with you — not the model provider — you have something.
Workflow lock-in that isn’t artificial. Not dark patterns. Real integration into how people work — the kind where switching costs are high because your product genuinely understands the user’s context, not because you made it hard to export data.
Distribution that doesn’t depend on the platform. Direct relationships with users. Brand. Community. Channels the landlord can’t shut off.
The Barbell Approach
You don’t have to avoid AI platforms entirely. That would be foolish — they’re the most powerful tools available. The key is how you allocate your investment:
Put the bulk of your effort — say 85% — into building the product that works today, using whatever tools make sense. Use the APIs. Use the models. Ship fast.
But put 15% into the thing the platform can’t take from you. That 15% is your land. “Build domain expertise” is easy to say. Here’s what it actually means:
1. Plant yourself in the gaps of regulation. No matter how smart the model gets, law and regulation are interpreted by humans. Tax, healthcare, finance, legal — the more complex the regulatory landscape, the less a model alone can solve. “Can I file it this way in this situation?” — the answer isn’t in the training data. It’s in the judgment built from handling hundreds of cases in the field. Encode that judgment into your product, and it survives any model swap.
2. Turn user-generated data into a compounding asset. The model has general knowledge. But “this company’s revenue spikes every March in this pattern, and the CEO prefers this tax optimization strategy” — the model provider doesn’t have that. Context data that accumulates through usage is the asset platforms can’t replicate. The key isn’t storage — it’s designing a loop where accumulated context makes the next decision more accurate.
3. Translate industry language into code. Every industry has tacit knowledge that doesn’t exist in any official manual. The tax accountant who says “just process it this way.” The legal clerk who knows “this registration pattern usually causes this problem.” Codify that into rule engines, checklists, workflows — that’s your layer on top of the model. The model is swappable. This layer isn’t.
4. Convert offline trust into online lock-in. Professional networks, client relationships, industry reputation — you can’t build these through an API. When word-of-mouth says “the accountant who uses that tool is different,” the platform can ship the same feature and users still won’t switch. Because the trust isn’t in the tool. It’s in the person using it.
5. Make switching costs structural, not artificial. The deeper a user’s history, settings, and workflows live inside your product, the harder it is to leave. Not because you’ve trapped them — because leaving means abandoning accumulated context. Would you throw away a tool that holds a year of your tax decisions and learned your preferences? That history is the asset.
The 15% isn’t “the part where you don’t use AI.” It’s the layer you build on top of AI that the AI provider cannot replicate: regulatory judgment, accumulated context, codified tacit knowledge, offline trust, and structural switching costs. That’s your land.
When the landlord changes the terms — and they will — the 85% adapts. You swap one model for another, adjust to new pricing, migrate to a different API. It’s painful but survivable. The 15% is what makes it worth rebuilding.
The Test
Ask yourself three questions:
- If your model provider doubled their prices tomorrow, would your business survive? If the answer is no, your margins belong to someone else.
- If the platform shipped your exact feature as a native capability, what would you still have that they don’t? If the answer is nothing, you’re a feature, not a product.
- How long would it take to switch to a different underlying model? If it’s more than a week, you’ve accidentally welded yourself to rented infrastructure.
Build the Thing They Can’t Copy
The AI era rewards builders. But it punishes those who confuse access to powerful tools with ownership of something durable.
The API is a tool. The model is a tool. The platform is a tool. Tools are rented.
Your land is the problem you understand better than anyone, the data that accumulates only through your product, and the trust you’ve built with the people who use it.
Build there.