I run two businesses and I am the only person in the operating layer of both.
JAK Labs is genuinely solo — no employees, no contractors, no one else. Haslett Handyman has crews, because somebody has to be on the roof and it cannot be me in two places at once. But everything above the work itself is one person: scheduling, quoting, invoicing, payouts, collections, the software all of it runs on, and the decision about what gets built next. There is no operations manager. There is no bookkeeper. There is no second engineer.
That is the part people mean when they say "one-person company" — not the absence of hands, the absence of a layer. Until this year it made me a slightly odd freelancer.
Now there are a few million of me in China, and their government is paying for the compute.
What is actually happening
In the first half of 2025, China registered 2.86 million new one-person companies — a 47% jump year on year. Not freelancers: registered businesses, run by a single person using AI to cover everything a team used to.
Then the state got behind it. In September the Ministry of Industry and Information Technology folded one-person companies into the national entrepreneurship support programme for 2026–2028. Provinces moved faster than that. Guangdong is building ten "ecological communities" for them this year and a hundred by 2028. Suzhou has pledged thirty communities and a thousand one-person enterprises. Pudong covers computing costs up to 300,000 yuan — about $44,000 a head. Wuhan offers loans where the state absorbs the default. Cities are converting idle data centres and empty offices into free incubators.
Rest of World quoted a researcher, Lin Zhang, on why this looks so unlike Silicon Valley: when a new technology appears, China mobilises "the entire bureaucratic system." There is no venture capitalist in this story. There is a policy.
The number that should stop you
A think tank that tracked more than 500 of these companies found roughly 80% fail within a year.
Hold those two facts next to each other. Millions of registrations, real subsidies, free compute, and the overwhelming majority are gone inside twelve months.
That is not an argument that AI does not work. I use these tools every day and they do exactly what they claim. It is an argument about what they work on.
The thing AI did not change
AI collapsed the cost of producing almost anything. Code, copy, images, video, analysis — the marginal cost of making a competent version of nearly any artefact fell through the floor, and it fell for everyone simultaneously.
It did nothing at all to the cost of finding somebody who will pay you.
Those are two different businesses. The first is production. The second is demand. Every one-person company that fails inside a year failed at the second one, and no amount of the first fixes it — a solo operator with world-class production capability and no customers is a solo operator with no customers, now with better tooling and a subsidised GPU.
Read the reporting and you can see it plainly. Rest of World profiled Ma Ruipeng, 41, who left a twenty-year programming career three months earlier to build AI software for making mobile apps. He works in Claude Code, Figma and OpenClaw — genuinely capable tools, the same ones I use. His earnings so far: none. He is living off savings.
He is not incompetent. He is a competent engineer who has solved production and has not yet solved demand, and the state has handed him a subsidy for the half he had already solved.
Where this came from, which matters
Much of the enthusiastic coverage is state media. Coverage of a policy, by outlets attached to the government running it, is not proof the policy works.
The context those pieces mention less: youth unemployment in China is running at roughly one in six for 16- to 24-year-olds. Domestic consumer demand is weak. Wei Xin, a 34-year-old document reviewer at a consulting firm, told reporters she knew AI would take her job before it did — so she took a Gemini course, built an AI version of herself, and moved into content.
Sponsoring one-person companies is a cheap answer to that. Cheaper than jobs programmes, and it reframes a labour-market problem as an entrepreneurship opportunity. Baiguan put the question better than I can: is this an entrepreneurial revolution, or an elaborate coping mechanism for a labour market under stress?
I do not think it is only one of those. Both can be true at once, and the 80% number is what it looks like when they are.
What I did differently, and it was not skill
I want to be careful here, because the easy version of this post is a solo operator congratulating himself while a few million people fail.
I did not start a one-person company. I had a business first, and then became the only person running it.
The handyman company existed. It had customers, invoices, crews and a scheduling problem, and I wrote software because the business was drowning in admin. To date $216K has been invoiced through software I wrote and still operate, and $192K collected, across 253 jobs carried end to end, with 51 scheduled automations running the parts nobody should be doing by hand.
Every one of those numbers is downstream of demand that already existed. The crews exist because customers do. The software exists because I could not hire an operations team and would not have wanted to — it is what lets one person carry scheduling, billing and payouts for a business with people in the field.
None of it conjured anybody who wanted their gutters cleaned.
That is the whole difference, and it is a difference of sequence, not talent. I solved demand first — badly, slowly, by doing the work — and applied leverage second. The structure being subsidised across China inverts that: leverage first, demand later, hopefully.
Sequence is not a small detail. It is most of the outcome.
What I would take from this if I were starting
Whatever AI makes cheap is not your moat. If a tool let you do it this morning, it let several million other people do it this morning. Production capability was scarce, so it felt like an advantage. It is now the floor.
Find the demand before you build the leverage. The least glamorous version of this works: sell the thing by hand, badly, to one real customer who pays. Then automate what hurt. That order is slower to start and far harder to kill.
Subsidy is not demand. Free compute, a free desk and a state-backed loan reduce the cost of existing. None of them is a customer. A business that is alive because the cost of being alive was lowered has not been tested yet.
The bottleneck was never making the thing. This is the part I would most want somebody starting out to hear, because it is the opposite of how the tools are sold. What nearly killed my business was not an inability to produce — plenty of people can swing a hammer or write a function. It was scheduling, quoting, chasing invoices, splitting crew pay, and the fact that all of it ran through one head. Admin does not scale by working harder, it scales by being carried by something that does not get tired.
That is the actual gift here, and it is quieter than the demos: AI did not make me able to build things. It made me able to be the entire operating layer of a business that has people in the field, without hiring a layer of managers to do it. That is real, and it is worth more than another content generator.
Being small is a real advantage, just not the one being sold. One person can carry an entire operation now, which used to be impossible. That is genuinely new and it is why my business works. But it is an advantage in serving demand, not in finding it.
What I will actually be watching
Not the registration counts. Those measure how many people took the offer.
The number that would change my mind is revenue per surviving company after the subsidies end, two or three years out. If a meaningful share of these are still running on customer money once the free compute stops, then something genuinely new has happened to the shape of a business and I will say so.
If the survivors are mostly the ones who had customers before the subsidy, then this was a very large and very expensive demonstration of something already well understood — and the people who learned it are a few million young Chinese founders who deserved better than to learn it this way.
I do not know which. Nobody does yet. It is the most interesting live experiment in my industry, and I am, in a small way in Michigan, one of its data points.
If you are in the middle of this
I build the operations software that lets a small business run without its owner — the scheduling, quoting and invoicing that otherwise lives in a group chat and somebody's head. That is leverage applied to demand that already exists, which is the only version of this I have seen work.
One concrete, free thing first: the website check takes about a minute, needs no account, and tells you what your own site is doing to the people who find it. Running a company alone means nobody else is going to look.
If you have customers and the admin is eating you, tell me what you are running. And if what you actually have is a good tool and no customers yet, I will tell you that plainly, because it is a different problem and software is not the fix for it.
Sources
The figures here come from: Rest of World on the provincial incentives and Ma Ruipeng; Sixth Tone and People's Daily on the registration numbers; The Epoch Times reporting the Zhuanzhi Think Tank OPC Research Institute's survival data; and Baiguan for the framing question. My own figures are in the register behind this blog.
