Every country with a technology ministry is now funding AI apps by the hundred. Every founder with API access can ship something that looks like a real product in a weekend. Almost none of it will still exist in three years. Two companies started from the exact same playbook — a clean interface on top of someone else's model — and one is worth $29.3 billion while the other's revenue cratered by more than half. The difference wasn't the model underneath. It's a framework that applies as much to a national AI strategy as it does to a single founder's weekend project.
Building an app on top of a foundation model has never been easier — and that's exactly the problem. A clean interface, a system prompt, a few weeks of work, and anyone can ship something that looks like a real AI product. The data on what happens next is blunt: only 3–5% of these products ever cross $10,000 in monthly revenue, and roughly 90% are expected to be dead by the end of 2026. An undifferentiated wrapper gets commoditized in about 18 months, often faster. This isn't a founder problem anymore. It's a national one — governments are now funding AI app development by the hundred, and most of what gets built will follow the exact same curve.
Cursor and Jasper are the cleanest case study available, because they started from the identical playbook. Both were, at their core, a polished interface sitting on top of someone else's foundation model — GPT-4 and Claude for Cursor, OpenAI's API for Jasper. Cursor has since crossed roughly $2 billion in annualized revenue and is valued at $29.3 billion. Jasper reached a $1.5 billion valuation within two years of launch — then its revenue cratered by more than half and its internal valuation was slashed. Same category, same underlying models, opposite ending. The difference wasn't which model either company used. It was what got built on top of it, and whether that layer became harder to leave the longer someone used it.
The market has a name for the split now: thin wrapper versus thick wrapper. A thin wrapper adds a system prompt and an interface — a competitor can rebuild it in a weekend, because anyone can call the same API. A thick wrapper is, functionally, a normal software company that happens to use an LLM underneath: it owns something the model provider doesn't and a competitor can't quickly copy. Jasper stayed thin. When ChatGPT itself got good enough to do most of what Jasper did, users had no real reason to stay. Cursor went thick — it became the place developers actually work, not a tool they open occasionally, and every session inside it generates workflow data that makes the product better in ways a rebuilt weekend clone simply doesn't have.
The frontier model was never the moat. It was the entry ticket. What happens after someone walks through that door — whether the product becomes part of how they actually work, or stays a tool they could swap out in an afternoon — is the only question that decides which side of the 90% a given app ends up on.
Strip away the branding and there are only three genuine sources of defensibility once the model itself stops being one. First, a proprietary data flywheel: the product collects usage, corrections, and outcomes that no competitor can access or scrape, and that data measurably improves the product over time — not data that just sits in storage. Second, workflow lock-in: the product becomes a system of record — the approval flow, the compliance layer, the operational pipeline a team can't simply route around. Third, distribution into a niche the large model providers themselves have no reason or ability to reach directly. The pattern that separates survivors from the 90%: durable products rarely rely on just one of these. They stack at least two, so that even if a foundation-model update erodes the technical edge, the workflow and the data the product has already accumulated don't disappear with it.
This is also why the "every third country building AI apps" instinct is, on its own, strategically incomplete rather than wrong. Capital and ambition solve the first problem — getting something built and shipped. They don't solve the second, harder one: whether what gets built has anywhere to accumulate a moat once shipped. A national AI strategy that funds a hundred thin wrappers into existence is, mathematically, funding the same 90% failure curve a hundred times over, just spread across more founders and more government grant lines. The honest strategic question for any serious builder or funder isn't "can we build an AI app" — that bar is now trivial. It's "what does this specific product own by month eighteen that a fast follower with the same API access can't simply rebuild."
The AI app gold rush is real, and so is the shakeout already underway inside it — 90% dead by year-end isn't a forecast, it's the trajectory the data already shows. That's not a reason to avoid building; it's a reason to be precise about what's actually being built. A thin wrapper is a legitimate way to test demand fast and cheaply, but it was never meant to be the finished product — it's the fastest way to find out whether people will pay for the outcome before investing in the harder layer underneath. The letter's core framework, worth carrying into any serious AI build, whether it's one founder or one government: identify which two of the three real moats — proprietary data, workflow lock-in, or a niche the model provider can't reach — this specific product can plausibly build within eighteen months, and build the thick layer around that from day one rather than after the thin version proves the market exists. The risk to this thesis: even a genuinely thick wrapper isn't permanently safe — Cursor's own moat holds only as long as the workflow data it accumulates stays harder to replicate than the model underneath it stays easy to swap out, and the exact same forces commoditizing thin wrappers today are what any thick wrapper is racing against over a longer horizon.
Founder, NextGen Economics · Bangalore, India · August 2026
Sources: DEV Community, "AI Wrappers Are Dying: Why Most AI Products Fail" (May 2026) · Preuve.ai, "Are AI Wrapper Startups Worth Building in 2026? Moat Test" (Jun 2026) — Cursor ARR/valuation data · HatchWorks, "AI Wrapper Product Strategy: Most Founders Get the Moat Wrong" (Mar 2026) — Cursor/Jasper case comparison · Joe Reis, "WTF is a Software Moat in 2026?" (Apr 2026) · BuildMVPFast, "AI Wrapper Startup? Build a Defensible Business in 2026" (Mar 2026) — startup survival statistics · Forbes, "Every Company Is Now An AI Wrapper So GTM Is The New Moat" (Jun 2026). This letter presents a strategic framework using real market data; it is not a recommendation to build or fund any specific company.
A deeper research note applying this moat framework to specific sectors — including the healthcare platform concept in Letter 161 — is available to NGE clients beyond what's covered in this free letter.
Not investment, tax, or retirement advice. This letter describes public retirement-income systems for informational purposes; consult a qualified financial or tax advisor in your own country before making any retirement income decision.