A Seoul National University study separates the entry wave from the survivors, and 97 percent of solo AI launches never come back.
Of the solo-founded products in the study, 97 percent launched once and never launched again. One shot, then nothing. That figure comes from a paper by Hyunso Kim, Hyo Kang and Jaeyong Song of SNU Business School, submitted 11 May 2026, built on 160,143 Product Hunt launches. It is the first checkable, non-self-reported evidence the one-person-company story has ever had. Everything before it was vendor content and founder tweets: three million a year, half a million a month, numbers nobody could audit.
The paper's finding is not the one the solopreneur content wave would like. Generative AI does lower the barrier to entry, sharply. More people ship alone than before, and they ship faster. But at the top of the distribution, where products survive, iterate and get traction, teams still lead, and their lead has widened. Entry got cheap. Endurance did not.
What eight failed searches say about a niche
For eight consecutive days, our research on the one-person-company niche returned nothing but supplier marketing and SEO filler. Self-reported revenue, no methodology, no way to verify a single claim. That is not a gap in our search. That is a description of the evidence base for one of the loudest business narratives of the past two years.
Then a named institution, a dated window, an identification design and a significance level. The contrast is instructive on its own. When a story is worth hundreds of millions in course sales and tool subscriptions, and the only serious academic treatment of it arrives in 2026 and partially contradicts it, the sensible reaction is not disappointment. It is relief that someone finally measured.
Entry and endurance are different questions
The useful distinction the paper draws is between getting started and keeping going. These have always been separate problems, but the AI discourse collapsed them. A model that writes your landing page, your first product spec and your outbound email removes the friction that used to stop people at day one. It does not remove the friction at day two hundred: the customer who churns, the integration that breaks, the decision nobody else can make for you.
The 97 percent number puts a floor under this. It says most of the solo wave is experimentation, which is fine and healthy and probably good for the world. It also says that treating a first launch as evidence of a viable one-person business is a category error. A launch is an entry event. A business is a survival record.
This matters directly for how our own thesis gets read. We have never argued that solo founders win. The claim is narrower and more specific: small teams, roughly one to ten people, can now run operations in overlooked industries that would have needed forty. Construction, maritime, water, energy, food distribution. The SNU paper does not weaken that. It sharpens it, by cutting away the version of the story that says one person plus a model equals a company.
The 62 percent that has nothing to do with technology
The same week, a different data point landed that fits the pattern. A PwC study covered by Zeitung für kommunale Wirtschaft found that among German-speaking utilities, 79 percent have started with AI, 61 percent cannot get out of pilot mode, and zero percent have arrived. The stated obstacles: technological hurdles 30 percent, skills shortage 28 percent, legal and ethical concerns 18 percent, cultural and organisational resistance 16 percent. That is 62 percent people and organisation, 30 percent technology, once you sort them.
Read the two findings together and they say the same thing from opposite ends of the market. The solo founder gets to launch and cannot sustain. The utility gets to pilot and cannot ship. In both cases the model did the easy part, and the hard part turned out to be the part it was never doing.
The practical move is to stop treating a launch, a pilot or a demo as evidence of anything except that the barrier to starting has fallen, and to ask instead what specific human capacity carries the thing from month three to month thirty.
What the paper does not settle
The complication is real and worth naming. Product Hunt is a skewed sample: consumer software, English-speaking, self-selecting, heavy on the kind of tool that gets built in a weekend. It tells us little about a two-person operation running predictive maintenance for a regional water utility, which is exactly the segment we care about. The paper is a preprint, not yet through peer review. And "teams lead at the top" is a statement about a distribution, not a prohibition. Somebody is in the tail.
There is also a measurement problem that keeps recurring. Bitkom measured German AI adoption twice and got 36 percent and 41 percent, a nineteen-point spread by one reading, from the same organisation using different instruments. If the country's largest digital association cannot hold its own headline number steady, then careful measurement is not a hygiene factor. It is a service somebody has to provide.
That is the position the SNU paper leaves us in. Not a story about whether one person can build a company, but a standard against which the next unaudited revenue claim gets held. Ninety-seven percent one-shot is now the number that sits opposite every screenshot of a solo founder's Stripe dashboard. It does not disprove any of them. It just asks how many launched twice.