AI made building cheap. What decides if my startup works now?

Supabase's State of Startups 2026 asked about two thousand startup builders what changed between 2025 and 2026 (Supabase, State of Startups 2026, read 12 October 2026). Three findings matter for an idea-stage founder.

1. Building is mostly done by AI now. 61 percent of respondents have more than half of their code written by AI. 40 percent are at 76 to 100 percent. Only 2 percent use none. 2. The build problem shrank. Technical complexity as the biggest business challenge fell from 24 percent in 2025 to 11 percent in 2026, the largest single shift in the survey. 3. The selling problem did not. Customer acquisition is now the top named challenge at 32 percent, ahead of product-market fit at 14 percent. Product-market fit means enough people want what you built to pay for it or keep using it.

Supabase calls this pattern "the cost of vibes", after vibe coding, which means building an app by describing it to AI. It sums the pattern up in one line: "Building got easier. Selling did not."

Not on its own. In Supabase's State of Startups 2026, 31 percent of heavy AI users (76 to 100 percent AI-written code) were earning money from their product, against 56 percent of founders who use no AI code. Supabase flags the catch: the youngest startups are the most likely to be heavy AI users, and young startups are also the furthest from earning money.

So read it as a warning light, not a verdict on the tools. The tools made shipping fast. They did not make finding buyers any faster.

Y Combinator made the same point in The State of Startups in 2026, an episode of its Lightcone podcast (episode page on TechFounder, posted 17 September 2026). The episode says nearly one in five YC companies now has a solo founder, AI lets smaller teams take on more ambitious problems, and knowing what to build is becoming more important than knowing how to build it.

A survey of builders and the hosts of YC's own podcast point the same way: customer acquisition now outranks technical complexity, and knowing what to build matters more than knowing how.

Because it removes the natural pause. When a product took months to build, you were forced to think about whether anyone wanted it. Now a weekend is enough, so it is easy to skip the question entirely. You end up with a working product and no evidence anyone needs it.

AI chat assistants can add a second trap. They are built to be helpful, and in our view helpful often slides into agreeable. Agreement from a tool is not a signal from a market. Neither is praise from friends.

Five steps, none of which need code:

1. Name the evidence you do not have yet. Compliments and waitlist sign ups are weak. A deposit, a preorder or a booked pilot is strong. 2. Write your riskiest assumption in one sentence. This is the single belief the whole idea depends on. If you cannot write it, you have a feature wish, not a business. 3. Decide what would make you stop. If nothing could, you are protecting the build, not testing the idea. 4. Separate "I can build this" from "someone will pay for this." The monetizing gap above is one sign that those two answers can drift apart. 5. Treat AI as a drafting partner, not a customer. Ask real people with the problem.

If you still need to build something to test the idea, build the smallest version that tests one belief. To rank the evidence you already have, see is my idea good or am I just excited.

Some ideas can only be judged by using them, like a game or a new kind of social app. For those, building a small version fast is the honest test, and cheap building is a real advantage. The point is not to stop building. It is to know which belief each build is testing.

You describe the idea, who it is for, and where you are. Idea Validator returns one of three verdicts, Build it, Sharpen it or Kill it, with its reasoning, how confident it is, your riskiest assumption and a test for it. Sharpen it is not a soft yes, and Kill it is a real outcome.

It does not replace talking to the people you think will pay. It tells you which question those conversations need to answer.

Write down the one belief your idea depends on. Under it, write the cheapest test that could prove it wrong within seven days. Run that test before the next build sprint.

• Supabase, State of Startups 2026, supabase.com/state-of-startups, read 12 October 2026.
• Y Combinator, The Lightcone, The State of Startups in 2026, 17 September 2026, episode page on techfounder.tv/the-state-of-startups-in-2026.