China Now Has 16 Million One-Person Companies. What That Looks Like From Inside One.
The one-person company went from niche term to national economic category in two years. Real registration numbers, the new July 2026 definition, and my own two-computer setup as one data point inside the boom.
Bottom line up front: as of mid-2025, China had over 16 million registered one-person companies, about 27% of all companies in the country, with new registrations growing 47% year-over-year in the first half of 2025. In July 2026, China published its first official standard defining an “AI OPC”: one core person in control, no more than 10 employees, business built on AI. I run one of these companies from Huizhou. This is what the statistics look like from the inside.
What “OPC” means in China right now
OPC stands for one-person company. The term caught on after a Suzhou AI conference in November 2025, and it moved fast from there: by mid-2026 the Ministry of Industry and Information Technology plus six other departments had issued a joint document calling to “accelerate the cultivation of AI one-person companies,” more than 20 cities had launched dedicated OPC support policies, and the first half of 2026 left 618 OPC entrepreneur communities across 24 provinces and 75 cities.
The July 2026 group standard matters because it turned a buzzword into a definition you can cite:
An AI OPC is a company controlled and led by one core individual, generally with no more than 10 employees, whose main business involves developing, applying, or servicing AI technology. It scales through AI agents rather than headcount.
The draft said “3 shareholders or fewer, 20 employees or fewer.” The final version tightened it to one person in control and 10 people. Regulators chose decision structure over team size. What they were pointing at: a human who judges and decides, and AI that does the volume.
The numbers, with receipts
| Signal | Value | Source |
|---|---|---|
| Existing one-person companies in China | ~16 million (27.4% of all companies) | national registration data, cited by EO Intelligence/iResearch 2026 report |
| New registrations, H1 2025 | 2.86 million, +47% YoY | same |
| Solo-founder share of new startups | 23.7% (2019) → 36.3% (H1 2025) | Carta (US-global, included for comparison) |
| Sole-proprietor firms with active credit profile on Sesame Credit | ~5.7 million, +42% YoY | Zhima Enterprise Credit, 2026 |
| Startup cost of a typical AI OPC | under ¥4,000 (~$560) | 2026 Global AI OPC Business Insight Report |
| Share of highly profitable OPCs that deeply use AI tools | 92% | same report |
These numbers agree on two things. The boom is real at the registration level, not a social-media narrative. And AI tooling is the difference between the profitable slice and the rest.
The honest part: most OPCs are not glamorous
Read past the policy headlines and the typical Chinese OPC looks less like a startup studio and more like one person refusing to be limited. An example I saw reported from Yiwu: a merchant selling artificial flowers, started with a ¥10,000 loan and nearly zero English, now takes orders from 50-60 countries using an AI agent for real-time translation plus automated customer service around the clock. He doesn’t know Cursor or Claude Code. He knows flowers.
The failure mode is equally common, and Chinese tech media gives it blunt airtime. Most one-person AI companies that die, die the same way: they automate the fun part (generating stuff) and never solve the boring part (distribution, payment, finding a customer who actually needs it). Deep AI adoption correlates with profit at 92%. The rest of the 16 million registrations stay mostly quiet.
One data point from inside: my company is two computers
Since people keep asking what “my team” means, here is the whole org chart.
Machine 1, the operations box. Always on, at home:
- Ubuntu 24.04, AMD Ryzen 9 7940HS (16 threads, Radeon 780M), 32 GB RAM, 1 TB NVMe
- Runs a low-power home server hosting eight AI agent instances with separate workspaces and names: operations, a fiction production line, a second writing line, finance, a stock-analysis desk, a knowledge base, a study buddy for my son, and a coordinator that supervises the others
- Five scheduled data jobs fire every morning between 7:30 and 8:25 without me awake (app-store rankings, review mining, content stats), plus an inbox watcher every 30 minutes and an hourly trend crawler. Nine scheduled jobs active tonight. The market database it maintains held 12,996 tracked apps when I counted tonight
Machine 2, the development box. In the office:
- Windows 11, Intel Core Ultra 9 285K, 64 GB RAM, NVIDIA RTX 5080 16 GB
- Does the heavy coding and model training work (voice-clone training is what its GPU is busy with this week), and hosts two more agent instances: one for domestic product development, one for our international line
My headcount is me. The machines run on electricity, and the “employees” run on a flat subscription that costs about what an old smartphone did.
What the agents don’t do is the part most demo videos skip. They don’t decide what’s worth building. They don’t take responsibility when a number is wrong. They don’t answer your customer’s angry email. That’s still mine. The July group standard got this right: the definition hangs on one person in control, not on the AI.
Why I’m writing this down
China’s OPC moment is a live experiment in whether one person plus agents can replace what a funded team used to do. The honest answer so far: they can, if the human is the part they can’t fake. I don’t know how this ends either, but I can report what it looks like from Huizhou, receipts included.
If you’re building alone somewhere else, in Berlin or São Paulo or Lagos or Jakarta or a spare room in your parents’ house, I’d like to compare notes. Different countries, same keyboard. Write to me: what you build, what your AI bill looked like last month, what broke. I answer most letters personally, and I publish the good ones with permission.