The Five Places Worth Building in AI (Everyone Else Is Toast)

Lovable just raised $330 million at a $6.6 billion valuation. They're shipping 100,000 new projects every day. Not every month—every day. And according to AI strategist Nate B Jones, they might still be doomed.

Not because they're doing anything wrong, but because they're building in what he calls "the middleware trap"—that shrinking space between the model makers (OpenAI, Anthropic, Google) and actual structural value. The trap works like this: if your product is essentially a UI wrapper around someone else's intelligence, your competitive moat lasts about as long as it takes a competitor to replicate your interface. With Claude and GPT-4 in the wild, that's roughly a week.

The conventional escape route is to train your own model. Replit's doing it. Cursor did it. Vercel trained a custom autofix model with Fireworks AI and just updated their terms to use customer code for training. But Jones argues this misses the point entirely.

"Training your own model isn't actually what separates the survivors from the casualties," he says. "The companies that are going to make it through this middleware trap share a very different trait. They own something structural that the model providers cannot replicate."

So what is structural when building software becomes free? Jones maps five layers of value that persist regardless of how good the models get. These aren't product categories—they're bigger. Think of them as the organizing principles for whatever the web becomes next.

Trust: The Verification Layer

When anyone can generate a professional-looking checkout page in seconds, "looks legitimate" stops meaning anything. The web is about to be—already is being—flooded with millions of AI-generated apps, services, and storefronts. Most will be indistinguishable from each other. Some will be actively malicious.

"The companies that become the verification layer, the ones who tell you that this app will not steal your credit card and we will back it up... those companies capture a tremendous amount of value," Jones explains.

This is why Stripe's position keeps getting stronger, not weaker. When you process over a trillion dollars in transactions, "Powered by Stripe" becomes a trust signal, not a technical feature. Same for Shopify. Same for Apple's App Store review process.

In an agentic economy—when your AI is autonomously booking flights and making purchases on your behalf—trust becomes the routing layer for the entire web. If an agent can't verify a service, it won't transact with it. Possibly won't even access it. Trust becomes a walled garden, and LLMs can't generate their way into one.

Context: The Choke Point

Notion doesn't pretend they want to train an AI model. They offer a model picker—choose Claude, ChatGPT, or Gemini, whatever you want. Their bet is completely different: "We don't care which model wins. We care that 100 million users have built the largest structured knowledge graph of organizational information on the planet and every model needs to come to us to access that information."

Context is the most valuable thing on the internet right now. Not compute. Not even your ability to prompt. It's your company's data, your customer relationships, your medical records, your meeting notes from last Tuesday. AI is a general-purpose tool—to be useful, it needs the specific data unique to your situation.

The companies that become the authoritative store for context and the permissioning layer that governs where it gets served own the choke point. Every agent, every model, every workflow has to flow through them. This is the same structural play that makes Salesforce durable, Epic durable in health, Palantir durable in security.

"An agent without context is just going to be a chatbot," Jones says, "but an agent that has your context can be a dependable junior employee. And it really is that big a difference."

Distribution: When Supply Is Infinite

You can generate an app in seconds now. Great. Who's going to see it?

Second-time founders know what first-time founders don't: the bottleneck was never building the thing. It was always distributing it. Field of Dreams lied—you build it and then you have to round people up and convince them to come.

When supply becomes infinite, curation becomes the scarcest resource in the world. The gatekeepers—Google, Apple's App Store, TikTok, YouTube—get stronger when the flood is bigger because they tell people where to go.

For the agentic web, this creates a new problem: agent discovery. If every business has AI agents, who helps those agents discover where to do business with each other and with humans? Jones sees an emerging category here—something like an agent-native app store.

"The question of what makes a business viable for an agent to transact with is one of the most interesting questions of 2026," he says. It's not just about putting up an MCP server. It's about transaction speed, how easily agents understand your offerings, how quickly they can make selections, how simply they receive the service. The entire mechanism of commerce needs rethinking with agents at the core.

Almost no one is thinking like this yet.

Taste: What You Choose to Build

This one's harder to quantify, which is exactly why it matters.

When producing software is free, what you choose to produce becomes the entire game. Your product decisions. Your design sensibility. Your editorial judgment about what's worth building. Your ability to look at what AI generated and know if it's right or wrong and be accountable for it.

"Taste is a conviction about what should exist in the world that is not easily derivable from training data," Jones says.

His analogy is music production. After GarageBand went mainstream, everyone could make a track. The flood was enormous. Now with Suno, you can generate an entire AI music track in seconds. The producers who thrive aren't the ones with the most expensive studio anymore—production is free. They're the ones with taste. An ear for what will connect with audiences.

The same thing is about to happen to software. The vibe coder who ships an app in minutes hasn't done the hard part yet. They haven't figured out how what they're building will deeply connect with their audience.

The Structural Question

Replit doesn't escape the middleware trap by outtraining Anthropic on models (they won't). They escape because Claude can't execute your code. Replit owns the runtime—the actual compute environment where your application lives. That's structural.

Vercel doesn't win because V0 uses a slightly better system prompt. They win because they're the deployment infrastructure that already hosts production applications for OpenAI, Anthropic, Nike, PayPal. They're not an AI wrapper with hosting—they're an infrastructure company that happens to have built an AI front door recently.

The pattern holds: AI commoditizes production. The companies that survive are building on layers that production can't replace.

Jones's framework raises an uncomfortable question for anyone building in AI right now: What do you own that matters if AI gets 10x better tomorrow? If the answer is "a clever prompt" or "a nice interface," you might want to find a different answer.

—Yuki Okonkwo, AI & Machine Learning Correspondent