199 companies participated in YC W26's Demo Day. After participating, looking at the data reveals the current state of AI startups.

First, the most impressive aspect is the proportion of AI-related companies. 60% of participating companies are AI-native, 26% are AI-enabled, and only 14% do not use AI. In other words, AI is no longer just a category but has become infrastructure. However, an important nuance here is that we need to look at this from the perspective of "how they are achieving what was impossible with traditional models," rather than simply "using AI."

Looking at the breakdown of business models, 87% are B2B, and only 7% are B2C. This overwhelming skew indicates that current AI technology is optimized for replacing knowledge workers' tasks. Interestingly, most companies aim for full job replacement, not just copilot-style assistance. AI agents are moving toward completely replacing human jobs, not just acting as copilots.

Data on revenue is also intriguing. The median estimated ARR is around $50k to $100k, with growth rates of 30-50% per month. However, only 5% of companies have annual recurring revenue exceeding $1 million, and about 50% still have no income. Many are still in the early stages.

What do the fastest-to-monetize companies have in common? It’s that the founders are "selling to their previous employers." Proximity achieved $700k in annual recurring revenue in less than three weeks. Corvera reached $33,000 in monthly recurring revenue in four weeks. They didn’t need to find customers because they already had customer networks.

This also relates to distribution channels. Of the top 15 rapidly growing B2B companies, 60% acquired initial customers through founder networks or YC networks. Companies struggling to develop markets almost always first build a product and then think about "how to sell." Successful companies start from "who can I reach, and what do they urgently need?"

The proportion of hardware companies is also noteworthy. 18% of participating companies include hardware, a significant increase compared to previous batches. Robots, drones, wearables, space tech—physical products are showing signs of revival.

The composition of founders is also distinctive. 46% are teams of two. The most common setup isn’t "hacker + salesperson," but two technical co-founders with different expertise. The most successful founders are those with deep industry experience—dentists creating AI surgical planning, aircraft maintenance managers developing document automation tools. They’re not selling at cocktail parties but penetrating industries that are quiet but deep.

Whether they can build a data flywheel is another critical factor. LegalOS trained on 12,000 visa application cases and achieved a 100% approval rate. The difference between companies that turn customer interactions into continuous product improvements and those that remain just tools lies here.

There are also notable gaps. No companies related to education. No consumer social media companies. Almost no government tech firms. Paradoxically, the fields with the least funding may hold the greatest potential for future returns.

Finally, clear patterns of failure are also evident. Uniquely un-differentiated agent infrastructure is risky. Eight to ten companies are building the same monitoring, testing, and compression features, but these are typically integrated natively by basic model providers. AI-native services without proprietary data are also vulnerable. General workflow wrappers will be replaced by native implementations of similar functions in GPT-5 within a few months.

The five common traits of the fastest-growing companies are: they sell results, not tools; the founders have already built relationships with customers; they start charging from day one; their customers are in urgent situations; and their MVPs are unusually simple.

What the data from these 199 companies shows is that in the AI era, the most important factors are not technological innovation but a deep understanding of problems, having a customer network, and accumulating data. The era where just "AI wrappers" can survive is coming to an end.