AI Business & Monetization
Build vs Buy vs Partner: A Practical Framework
A startup spent $800K and a year building their own AI model. A competitor shipped on the OpenAI API in 3 weeks. Both closed Series A - but the second one did it 12 months earlier.
- Notion AI: integrated OpenAI API in a few months → immediate $10M+ ARR from the AI tier
- Jasper AI: $75M ARR built on GPT-3/4, not a single line of custom ML code
- Harvey AI: $100M ARR - fine-tuned models on legal data (a rare case where Build is justified)
Предварительные знания
The Framework: Make the Decision in 10 Minutes
**Mistake #1 for AI startups:** building what could be bought for $200/month. Engineers spend 6 months on a custom solution while a competitor ships on a ready-made API in 2 weeks. The right framework is three questions that eliminate 90% of the uncertainty.
| Scenario | Decision | Rationale | Real Example |
|---|---|---|---|
| NLP for legal documents | Build / Fine-tune | Unique data + core advantage | Harvey AI - $100M ARR |
| Marketing copy generation | Buy (API) | Commodity, GPT-4 handles it | Copy.ai pays OpenAI $5–10M/year |
| AI features in a CRM | Partner | Need Salesforce distribution | Einstein GPT - Salesforce + OpenAI |
| Foundation model from scratch | Buy (API) | Costs $50M+, takes years | 99% of companies use APIs |
| AI for medical imaging | Build | Regulatory + unique data | Nuance DAX - $7B acquisition |
**Notion AI vs Cohere** - a perfect contrast. Notion integrated the OpenAI API in a few months and shipped AI Writing. Cohere spent 3 years and $270M building their own models. Both were right - for their respective goals. Notion doesn't need its own model. Cohere sells to enterprise clients who need customization and data privacy.
A company must build its own AI model to create a real business
99% of successful AI companies use third-party models via API
OpenAI/Anthropic/Google have invested $10–100B in model development. The real advantage lies in data, UX, distribution, and vertical expertise - not the model itself.
A team is building B2B HR SaaS with an automatic resume screening feature. Which approach should they choose?
Real Math: What Everything Actually Costs
Let's look at concrete numbers. Often the choice between Build and Buy is the difference between $300/month and $600,000 in year one. Here's a real-world comparison for a single feature.
**Hidden Build costs:** the figures above are optimistic. In reality add: recruitment ($30–50K per ML engineer), project delays (×1.5–2x timeline), infrastructure ops, monitoring, and model degradation. The true cost of a custom model is $1–5M in year one for a serious product.
| Option | Year 1 (all-in) | Monthly (post-launch) | Time-to-market | Control |
|---|---|---|---|---|
| Build custom model | $500K–5M | $8–30K | 6–18 months | Maximum |
| Buy API (OpenAI/Anthropic) | $5–20K | $200–5K | 1–4 weeks | Medium |
| Buy SaaS platform | $10–50K/year | $1–5K | 1–2 weeks | Low |
| Partner (white-label) | $0–50K setup | % of revenue | 2–8 weeks | Low |
At 100K requests/month with avg 700 input + 200 output tokens - how much will GPT-4o API cost?
Partnerships as Leverage: Real Cases
Partnerships are the most underrated option. The right partnership provides distribution that couldn't be bought for any amount of money. Three models, each with their own economic logic.
| Model | How It Works | Economics | When to Choose | Risk |
|---|---|---|---|---|
| OEM / White-label | Embed the company's AI in someone else's product | Upfront fee + 10–20% royalty | Have tech, lack distribution | Partner dependency |
| Reseller / Marketplace | Sell someone else's AI under a private label | 20–40% margin | Fast start, no R&D | Price pressure |
| Co-build / JV | Joint development and shared IP | Shared IP + revenue split | Complementary assets | Long, complex to manage |
| Platform ecosystem | Build on someone else's platform | 15–30% revenue share | Need their audience | Rules can change anytime |
**Microsoft + OpenAI** - the most significant AI partnership. Microsoft invested $13B and got priority integration across Azure. Result: Azure AI services contributed $30B+ in additional ARR by 2025. OpenAI got $10B+ in GPU compute and enterprise distribution without building their own sales team. Win-win: each does what they do best.
**Partnership selection rule:** the right partner must have what the team could never build alone - distribution, unique data, or regulatory access. If those assets are acquirable independently at reasonable cost, partnering means giving up too much control.
A startup built a great AI model for medical imaging analysis. Technology is ready, no sales yet. What should the team do?
Key Takeaways
- Build only when it's the core competitive advantage or there is unique training data
- The cost difference between Buy and Build is often $300/month vs $600K in year one
- Partner when the needed distribution can't be bought: healthcare, enterprise, regulated industries
- Microsoft + OpenAI: $13B investment → $30B Azure AI revenue - the best partnership example
- 99% of successful AI products run on third-party models
What's Next
After choosing Build/Buy/Partner, the next decision is the monetization model. Next lesson: SaaS + AI - how to build a subscription AI business.
- SaaS + AI Models — next lesson
- AI Company Types — prior context
Вопросы для размышления
- In an AI product - what constitutes the core competitive advantage worth building vs. buying?
- Name 3 features in a product that could be bought for $100–500/month instead of built in-house
- What types of partnerships provide distribution that cannot be obtained any other way?