"It's just a wrapper around GPT." That's the tech world's favorite insult. And on March 21, Fortune aimed it at a company seeking a $50 billion valuation.
Fortune's "Cursor's crossroads" piece questioned whether the world's hottest AI coding tool is a sustainable platform or a glorified API skin headed for a squeeze. The timing stings: that same month, CostLayer tracked 114 out of 483 AI models changing prices — the biggest repricing wave in the industry's history. OpenAI's head of ChatGPT called their current pricing "accidental" and promised it will "significantly evolve." Both OpenAI and Anthropic are preparing for IPOs. The subsidized-API party is ending.
This matters because most AI products are thin interfaces over someone else's API (Application Programming Interface — a way for programs to talk to each other, like a waiter between kitchen and table), charging $20/month for something you could build in a weekend. When API prices stayed artificially low, bad unit economics hid in the margins. Now the bill arrives.
But here's what wrapper-haters miss: Cursor is "just a wrapper" — sitting at $2 billion in annualized revenue. Jasper hit $80M ARR in late 2022 as "just a wrapper" — then crashed to $35M revenue in 2024 when ChatGPT ate its lunch. Perplexity is "just a wrapper." Some wrappers chase $50 billion valuations. Most are worth nothing.
I've been reverse-engineering both kinds. Here's what separates them. 🔍
The wrapper spectrum
Not all wrappers are equal. There's a spectrum from "pointless" to "platform," and where you land determines whether you exist in 12 months — especially now that 2023-era VC runway is running dry.
Level 1: The Prompt Wrapper (dead on arrival). Takes user input, adds a system prompt, sends it to the API, returns the output. A Chrome extension that "writes emails with AI." A website that "generates marketing copy." These are demos, not products. The API provider can replicate everything by updating their own UI. Zero moat — no competitive advantage that's hard to copy.
I see 50 of these launch on Product Hunt every week. They get 200 upvotes, 30 signups, and 0 paying customers by month 3. The founder posts "lessons from my failed startup" on Indie Hackers. Cycle repeats.
Level 2: The UX Wrapper (6–18 month lifespan). Same API under the hood, but the interface is genuinely better than what the provider offers. Early Cursor in 2023 was this — VS Code plus AI, packaged better than GitHub Copilot. These products have a window. It closes when the provider catches up. GitHub Copilot went multi-model at Universe in October 2024 and shipped Agent Mode in 2025. OpenAI launched Code Interpreter in July 2023. Anthropic shipped Artifacts in June 2024. Every UX advantage has an expiration date.
Level 3: The Workflow Wrapper (2–5 year lifespan). These products don't just call the API — they build an entire workflow around it. The AI is one component in a larger system that includes data pipelines (automated flows that collect, process, and route information), user-specific context, integrations, and domain-specific logic.
Perplexity lives here. It's not "GPT with search" — it's a research pipeline that crawls, indexes, synthesizes, and cites sources. Rebuilding that pipeline takes months, not a weekend. The AI model accounts for maybe 30% of the product value. The other 70% is infrastructure.
Level 4: The Data Wrapper (5+ year moat). The unicorn tier. These products use the AI model, but the real value sits in proprietary data that makes the model's output uniquely useful. Every user interaction improves the product. The AI model is replaceable — swap GPT for Claude, nothing changes — but the data layer isn't.
Cursor sits here now. Every codebase it indexes, every code completion accepted or rejected — that's behavioral data that drove the jump from early wrapper to $2 billion ARR. A new competitor starting today would have the same AI model but none of that data. The gap compounds daily.
Five signs your wrapper is doomed
I can spot a dying wrapper within 5 minutes of using it:
1. The AI model IS the product. Ask "what breaks if the AI model disappeared?" If the answer is "everything" — you're a prompt wrapper. The model should be a component, not the entire product.
2. No flywheel. A flywheel is a self-reinforcing cycle: more users → more data → better product → more users. If your 1,000th user gets the same experience as your first, you're running in place.
3. Monthly churn above 8%. Churn — the percentage of customers who cancel each month. At 8% monthly churn, you lose half your user base every 8 months. You need a waterfall of new signups just to stay flat.
4. "Me too" positioning. "Like [successful product] but cheaper" works for exactly one purchase cycle. Then the customer realizes the original is better and switches back. Being 20% cheaper is a temporary discount, not a moat.
5. No switching costs. If I can export all my data and move to a competitor in 15 minutes, you're a commodity. The best products create lock-in through accumulated customization, learned preferences, and team-wide adoption. Leaving means losing all of that.
What the 10% do differently
The wrappers that reach unicorn valuations share three traits:
They own the workflow, not just the model call. Cursor doesn't just autocomplete code — it understands your entire codebase, your file structure, your coding patterns. The value isn't "AI generates code." It's "AI generates code that fits YOUR project."
They build proprietary data loops. Every interaction makes the product measurably better. Perplexity's search index gets richer with every query. Jasper's brand voice profiles got sharper with every edit — until ChatGPT launched in November 2022 and offered 80% of the same capability for free. Jasper's revenue plummeted from a $120M peak in 2023 to $35M in 2024 before recovering to $88M in 2025 by pivoting hard into enterprise workflow. The data loop saved them. Without it, they'd be dead.
They make the AI invisible. The best wrappers don't feel like AI products. Grammarly doesn't market itself as "GPT for writing." It markets itself as a writing assistant. When the underlying model improves, the product gets better silently. When the model has a bad day, the product still works because it's not entirely dependent on a single API call.
The one question that matters
March 2026 is the inflection point. With 114 AI models repricing in a single month, OpenAI and Anthropic racing toward IPOs, and subsidized API costs climbing to sustainable levels, every wrapper built on cheap tokens faces a margin squeeze it hasn't experienced before. The startups that raised 18–36 months of runway in 2023 are hitting zero. The wrappers with real unit economics survive. The rest don't.
If you're building an AI product right now, ask yourself: "If OpenAI and Anthropic both doubled their API prices tomorrow, would my product survive?"
If yes — you're building something real. Your value sits above the API layer.
If no — you're subsidized, not sustainable.
Build the data. Build the workflow. Build the switching costs. Let the AI model be the replaceable part. 💰
The 90% build on top of AI. The 10% build AROUND it. That one preposition is a $50 billion difference — just ask the wrapper everyone loves to hate. 🦝
