After watching the YC W26 Demo Day, I realized something from listening to 199 company pitches. This is no longer just a gathering of startups; it’s like a map showing how business structures are changing in the AI era.

First, the numbers. 60% of the entire batch are AI-native companies. In other words, AI is responsible for the product itself. Additionally, 26% are AI-enabled, and only 14% are companies not using AI. But what’s important here is not whether they are using AI, but what they have achieved.

Another realization is that the concept of “copilots” has almost disappeared. Last year, 4% of companies promoted AI copilots, but this year, it dropped to 1%. Instead, what has emerged is “AI employees.” In other words, complete replacement of high-paying jobs. AI agents performing the same work for a portion of the salary. This, in other words, means that the human salary system itself has become a target of competition.

In terms of business models, B2B stands out. 87% are B2B companies, and only 7% are consumer-facing. There is a pattern where B2B companies are fastest to generate revenue. It’s often when founders sell to their original employer network. The CEO of Proximtty (which earns $700k in annual recurring revenue in three weeks) is a McKinsey alum. The CEO of Corvera (which earns $33k in monthly recurring revenue in four weeks) is a consumer goods brand manager. In other words, existing trust and connections become distribution channels.

Hardware is making a comeback. 18% of the batch are working on robots, drones, wearables, and space technology. This is a significant increase compared to a few years ago. It’s noticeable that alumni from SpaceX and Tesla are starting physical product companies.

A new form of service business has also emerged. AI-powered law firms, staffing agencies, accounting, and insurance brokers. These operate on a performance-based model while enjoying software profit margins. In other words, starting from services, then releasing automation once data accumulates, and ultimately upgrading to a platform. This pattern is the fastest way to generate revenue.

Data is becoming a moat. LegalOS trained on 12,000 visa application cases and achieved a 100% approval rate. Accuracy improves with each hiring. Without a data flywheel, it’s just a product.

Failure patterns are also visible. Non-differentiated agent infrastructure. Eight to ten companies doing the same thing, so once a foundational model provider integrates natively, it’s game over. The same applies to general-purpose AI wrappers. The “AI for everything” proposition cannot beat the idea of “replacing specific jobs paying $80k a year” with AI.

The most rapidly growing companies share a common trait: they sell results, not tools. Founders build relationships with customers before the product even exists. They start charging from day one. Customers are not driven by curiosity but by urgent needs. MVPs are unnaturally simple.

There are gaps in consumer, education, and government tech. These are future opportunities. Historically, the areas with the least funding have produced the outsized returns.

Conclusion: deeply penetrate dull industries. Build data flywheels. Avoid generic AI wrappers. The best opportunities are not in industries you can sell at cocktail parties.