There Are Only 5 Safe Places to Build in AI Right Now. Are You in One?
TL;DR
The AI app-builder gold rush is collapsing into a middleware trap — Nate says Lovable, Replit, Bolt, Shipper, and others are increasingly “thin wrappers” around Claude, ChatGPT, or Gemini, which means their moat can vanish in a week once a model maker or competitor copies the UI.
The durable businesses aren’t the ones with slightly better prompts — they own structural layers models can’t replicate — his key examples are Replit owning runtime/compute, Vercel owning deployment infrastructure for customers like OpenAI, Anthropic, Nike, and PayPal, and Notion owning a massive organizational knowledge graph across 100 million users.
There are only five safe places to build: trust, context, distribution, taste, and liability — those are the layers Nate argues become more valuable as AI makes production nearly free and floods the web with millions of low-quality apps, storefronts, and content streams.
Trust and context become choke points for the agentic web — Stripe processing over $1 trillion in payments, Shopify, Apple’s App Store review, Notion, Salesforce, Epic, Palantir, Snowflake, and Databricks all sit in positions where agents will need verification signals or proprietary data to act usefully.
Distribution gets more important, not less, when software supply goes infinite — Nate’s line is basically the opposite of “build it and they will come”: in a world with 10x–100x more software, curation and discovery become scarce, which strengthens Google, Apple, Amazon, TikTok, YouTube, and any future “agent-native app store.”
The real strategic test is simple: if AI gets 10x better, does your product get stronger or obsolete? — if better models erase your value, he says reposition now; if better models increase the value of your trust layer, liability layer, context store, or agent infrastructure, you’re in a safer zone.
The Breakdown
The app-builder boom is starting to look fragile
Nate opens by arguing the real story isn’t just that players like Lovable and Replit are scrambling to add “OpenClaw”-style capabilities — it’s that the whole build layer is collapsing into sameness. Lovable has raised $330 million at a $6.6 billion valuation and says it sees 100,000 new projects per day, but underneath the hype, many builders are still wrapping the same few base models.
Why “just train your own model” isn’t the answer
He pushes back on the obvious escape hatch: owning a model. Cursor is training its own coding model, Replit has code completion models built with Databricks and released on Hugging Face, and Vercel has a custom autofix model with Fireworks AI — but Nate says that still isn’t the real dividing line between survivors and casualties.
The companies with a shot own something structural
His examples are concrete and sharp. Replit matters because Claude can’t execute your code, while Replit owns the runtime where the app actually lives; Vercel matters because it already owns production deployment infrastructure for companies like OpenAI, Anthropic, Nike, and PayPal. Notion’s move is even cleaner: it doesn’t care whether Claude, ChatGPT, or Gemini wins, because the real asset is the structured organizational context of 100 million users.
Safe zone #1 and #2: trust and context
The first durable layer is trust: in a web flooded with AI-generated apps and scams, “powered by Stripe” stops being a technical detail and becomes a signal that someone will back the transaction. Then comes context — your meeting notes, medical records, company data, customer relationships — which is why Notion, Salesforce, Epic, Palantir, Snowflake, Databricks, and even Google Maps sit in such powerful positions.
Safe zone #3: distribution still rules the game
This is where Nate gets especially blunt: first-time founders obsess over building, second-time founders know the bottleneck was always distribution. He uses the “Field of Dreams” contrast — software does not work like “build it and they will come” — and says infinite software supply makes curation the scarce resource, which only strengthens gatekeepers like Google, Apple, TikTok, YouTube, Substack, and Amazon.
Safe zone #4: taste is what matters when production is free
He treats taste not as a soft personal trait but as a real web layer. His analogy is music: after GarageBand, and now with Suno, anyone can make a track, but the people who win are the ones with the ear and conviction to create something audiences actually want. In software, and especially in agent systems, that shows up as orchestration quality — a human choosing the workflows, prompts, tools, and definition of what “good” looks like.
Safe zone #5: liability is boring — and incredibly valuable
The least glamorous layer is accountability, but he thinks it may become one of the biggest businesses. If an AI financial plan loses money, a medical app gives bad advice, or an AI contract gets you sued, “the AI did it” won’t survive court; that’s why players like Deloitte, McKinsey, 11Labs, Veeva, and Elation are all inching toward AI assurance, insurance, and governance.
His closing test for builders
Nate ends with a simple strategic question: what do you own that still matters if AI gets 10x better? If a better Claude model kills your product, you’re exposed; if it makes your trust, context, distribution, taste, or liability layer more valuable, you’re building on something durable. And he lands on one old lesson that now matters more than ever: shipping an MVP in a day is nice, but getting real distribution and customer validation is still the hard part.