Building on Borrowed Ground

Anthropic just shipped Claude Design. Prototypes, pitch decks, marketing collateral, interactive mockups… describe what you need and it builds. It’s closer to Figma than Lovable today, but the distance from there to building full apps is small, and the direction of travel is obvious. Lovable is sitting at a $6.6B valuation doing exactly that, on top of Claude’s API.

Some people are calling this a betrayal. That framing is emotionally satisfying and analytically useless. It implies surprise. It lets founders recast a structural problem as an ethical one, which is a comfortable story that doesn’t hold up. Anthropic didn’t violate some unspoken agreement. They followed the incentive that every infrastructure provider eventually follows: extract value from your highest-margin use cases once you can see exactly where the margin lives.

And they can always see where the margin lives. That’s the part founders underweight.

You’re Building Their Roadmap

Every startup building on an AI API is generating a continuous stream of product intelligence for the model provider underneath them. Which workflows retain. Which prompts convert. Which use cases drive the most tokens. The founder thinks they’re building a company. The infrastructure provider is watching a detailed product roadmap take shape in their usage logs. It’s not adversarial. It’s just the natural consequence of building on someone else’s ground.

This isn’t unique to AI. It’s the oldest dynamic in platform economics. Amazon watched which third-party categories were converting and launched Amazon Basics. Apple watched which apps were sticky and built the functionality into iOS. The pattern is consistent. Platforms expand toward value, and applications built on top of them are always operating with an implicit clock running.

The Real Weapon Is Cost Structure

The Lovable story isn’t really about Anthropic being predatory. It’s about what Lovable’s moat actually was. Strip away the valuation and the growth curve, and you have a product with clean UX and a distribution advantage… all sitting on top of infrastructure it doesn’t control.

Here’s the part that usually gets glossed over. Platforms don’t beat applications by building better products. They beat them by pricing the application’s core offering at marginal cost, because they earn their margin one layer down. Lovable has to charge enough to cover Anthropic’s API fees plus a margin of its own. Anthropic can charge the API fees and nothing else. That’s not a product fight. That’s a cost-structure fight the application literally cannot win.

This is why infrastructure and platforms age better than applications. The value in infrastructure accrues to whoever owns the layer everything else depends on. Applications sit above that layer, exposed. Infrastructure sits below it, collecting rent. AWS doesn’t worry about which startups it’s hosting. It charges them all. Stripe doesn’t care which e-commerce model wins. Every transaction runs through the rails regardless. The infrastructure provider’s business gets stronger as the ecosystem above it grows, even when individual applications inside that ecosystem fail or get absorbed.

What Actually Survives

So what does defensibility actually look like?

UX gets copied. Speed-to-market is table stakes. The companies that are hard to displace share a few characteristics. They’re sitting on proprietary data that compounds the longer you use the product. They’re embedded in workflows where switching costs are real, not theoretical. Or they’re operating in domains where regulatory complexity, trust relationships, or distribution networks create friction that can’t just be shipped around.

The companies that are genuinely hard to compete with aren’t the ones with the best product at a moment in time. They’re the ones where the underlying layer is doing work that’s invisible until you try to remove it… existing data relationships, compliance infrastructure, deep integration into how an industry actually operates. The enabling technology makes them faster. It doesn’t make them fragile.

If you’re allocating capital into the AI application layer right now, run every deal through one filter. If Anthropic, or OpenAI, or whoever sits underneath shipped the adjacent product tomorrow at cost, what survives? If the answer is the brand, the UX, or the growth curve, you’re underwriting a time-limited arbitrage dressed as a company. The price should reflect that. The companies worth paying up for are the ones where the answer is proprietary data, embedded workflows, regulated distribution, or trust relationships the model provider structurally can’t replicate by shipping a feature.

A lot of what gets funded as category-defining turns out to be a feature of the layer below it. Some of them have real revenue on the way to figuring that out. But if the honest answer to “could the infrastructure provider absorb this if they wanted to” is yes… that’s borrowed time.

The more useful question for anyone building right now isn’t “how do I avoid getting copied.” It’s “what am I building that compounds in a way they can’t replicate.”

Most founders haven’t answered that yet.