AI Business & Monetization
Foundation vs Application: Why Wrappers Win
OpenAI: $80B valuation, thousands of employees, the best models in the world - and losses. Midjourney: 40 employees, no VC, $200M profit. This paradox is the key to understanding where to actually build an AI business.
- Cursor: $100M ARR, trained zero models - just the best UX for developers
- Midjourney: forked open models + added community via Discord - 20M users
- Perplexity: AI on top of web search - $3B valuation without training a foundation model from scratch
Предварительные знания
Foundation Model Economics: Why They Lose Money
**OpenAI is valued at $157B** (end of 2024 valuation). Yet they're unprofitable at $3.4B ARR. How is this possible? And more importantly - what does this teach us about building AI businesses?
| Company | Type | ARR / Revenue | Profitability | Note |
|---|---|---|---|---|
| OpenAI | Foundation model | $3.4B | Unprofitable | Inference costs eat margins |
| Anthropic | Foundation model | $1B+ | Unprofitable | $7.3B raised, all going to compute |
| Mistral | Foundation model | $50M+ | Unprofitable | European competitor, VC-funded |
| Cohere | Foundation model | $50M+ | Unprofitable | B2B focus, also in the red |
| Midjourney | Application | $300M+ (2024) | PROFITABLE | Bootstrapped, 40 people |
| Cursor | Application | $100M+ | Likely profitable | Lean team, low API overhead |
**Why are foundation models unprofitable at massive revenue?** Each new user = new inference costs. Training the next model costs $100M+. Competition forces low API prices. This is a structural trap that's hard to escape without the scale of AWS/Microsoft.
Anthropic raised $7.3B in investments and is still unprofitable. What's the main reason?
Why Applications Win: Distribution Beats Technology
**Cursor hasn't trained a single neural network.** They forked VSCode, added smart integrations with Claude and GPT, built a context engine for codebases - and reached $100M ARR. This isn't "just a wrapper" - it's smart distribution of technology.
**Most importantly:** when OpenAI released ChatGPT with a code interpreter - it didn't kill Cursor. Why? Because Cursor is embedded in a developer's workflow (VSCode), understands the entire codebase, not just selected code. Distribution + UX = moat.
| Factor | Foundation Model (OpenAI) | Application (Cursor) |
|---|---|---|
| Can switch AI provider? | No (they are the provider) | Yes - easily |
| Knows their customer? | Abstractly (all people) | Very well (developers) |
| Embedded in workflow? | Via separate app | Right in the IDE |
| Growing moat? | Brand + model quality | User data + distribution |
| Gross margin | 20-40% | 60-80% |
**Distribution rule:** the most valuable asset of an AI application is not technology, but access to the user at the right moment. Cursor is valuable because it's opened every 4 hours by a developer. That's the moat.
OpenAI released Codex and GitHub Copilot competes with Cursor. Why does Cursor keep growing?
"Wrapper" Is Not an Insult - It's a Strategy
**In AI communities, "wrapper" is used as an insult** - as if GPT-4 was just wrapped and called a product. But look at the facts: Jasper, Cursor, Midjourney, Perplexity - all "wrappers". And all of them make real money.
**When a wrapper loses:** Jasper lost ~50% of their valuation when OpenAI built similar features directly into ChatGPT. This shows the limitation: if a wrapper only solves what the base model will soon solve itself - it's vulnerable.
| Wrapper Type | Defensibility | Risk | Example |
|---|---|---|---|
| Generic content tool | Low | OpenAI will add these features | Early Jasper |
| Vertical domain tool | Medium | Needs a domain-specific competitor | Harvey AI (legal) |
| Workflow integration | High | Hard to switch away from | Cursor, Notion AI |
| Data-rich vertical | Very High | Data = unique asset | Abridge (medical transcription) |
**Rule of a defensible wrapper:** the best wrapper is one embedded in a critical user workflow that accumulates data improving the product. Cursor sees the developer's code. Harvey AI learns from the firm's legal documents. That cannot be copied.
Jasper lost ARR when ChatGPT added similar functionality. Which type of wrapper is most resilient to this risk?
Key Ideas
- Foundation models are structurally unprofitable: inference costs scale linearly with users
- Applications are profitable: 60-80% margins, no CAPEX for training models
- Distribution beats technology: Cursor won not with a better model, but better integration
- "Wrapper" is not an insult - it's a business strategy with $100M+ ARR case studies
- Vulnerable wrapper: generic functions; defensible wrapper: workflow + user data
What's Next
It's clear the application layer is better for profitability. But what type of AI company to build within the application layer? We break down 4 models.
- 4 Types of AI Companies — Which model fits available resources
- AI Funding — When VC is needed, when bootstrap is better
Вопросы для размышления
- Name 3 products used daily that could theoretically be built on an AI API. What prevents someone from just copying them?
- What type of wrapper would make most business sense: domain-specific tool, workflow integration, or data-rich vertical? Why that one?
- If OpenAI added a product feature directly to ChatGPT tomorrow - would it kill the business? How would one protect against this?