# The Secret Behind 5 Small Companies That Made Big Money (Little Known in Korea)

[Watch on tuber](https://tuber.guzus.xyz/video/qROfKLNhmiw/summary) · [YouTube](https://youtu.be/qROfKLNhmiw)

**Score: 66/100 · skim**

Five tiny companies made large sums with AI, but building the product wasn't what made them money: AI made building cheap. What separated them was something rivals couldn't copy, such as an audience, a problem the founder lived firsthand, or someone else's license. Of the five, only two would work in Korea exactly as they are, two would work with conditions, and one would be illegal.

*(The transcript is a machine translation from Korean, and several won figures are garbled. Where amounts conflict, the dollar figures and the logic are more reliable.)*

- **0:00-0:41** — **Intro and premise.** The video ranks five companies with fewer than 10 employees from #5 to #1, by how much they earned. Each either sold AI or built its product with AI. Examples include a solo developer earning about 200M won a month, a company with zero employees, and a two-person company with about 540B won in revenue. For each one, the host asks whether it would work in Korea and says up front that only two would work unchanged.

- **0:41-2:44** — **#5: Pieter Levels' flight game, and why his photo app outlasted it.** A Dutch developer built a browser flight game in about three hours by giving an AI coding tool plain-language instructions. Seventeen days later he posted an annualized revenue figure of about $1M (≈1.35B won). The money came from in-game ad billboards sold to companies, not from players; jet sales totaled only $360. The real driver was his 500,000+ social media followers, who saw the launch immediately. A tech outlet's headline put it bluntly: your game probably won't earn this much. A year later his profile reportedly listed the game's earnings as $0. His AI photo subscription service still earns about 200M won in its best month and about 130M won in mid-2025. The difference was recurring payment: subscribers keep paying, while advertisers simply stopped renewing.

- **2:44-3:29** — **The core thesis, and Korea's verdict for #5.** A freelance marketplace report found that AI-driven creative contracts rose 90% while pay per contract fell 13%, because anyone can now produce the work. The host's recurring line is "making things is free; money goes to what others can't take away." The game took three hours to build, but 500,000 followers can't be built that fast. In Korea the technology is equally available, so what limits you is distribution, meaning people who will spread the word.

- **3:29-5:38** — **#4: Polsia, a company with no employees.** Polsia has only its founder; legal work and computer equipment are outsourced. It sells an "AI co-founder" subscription that builds products, fixes errors, handles customers, places ads and writes a nightly summary of the day. It has about 10,000 paying customers, forecasts about 13.5B won in sales this year (more than double the annualized figure from March), and reportedly raised about 40B won. Supporting data:
  - Payment-processor data shows the number of solo businesses passing $1M has multiplied.
  - A business-school professor studied 50,000 startups and found AI-focused ones had about 25% fewer staff.

  The catch is cost. Paid AI models made some customers unprofitable, so the founder switched to a free Chinese model. Salaries are a fixed cost, but AI inference costs grow with every customer. **In Korea it works with conditions:** first check whether the monthly AI cost per customer is lower than what that customer pays.

- **5:38-7:20** — **#3: Cal AI, built by high schoolers.** A 17-year-old in the US was counting calories to gain weight, not lose it. With another high school student he built an app that estimates calories from a food photo, launched in May 2024. According to the acquirer, it passed 15M downloads and more than $30M in annual revenue in under two years, with seven employees. A diet-app company bought it last December for an undisclosed price. Cal AI still runs independently, with the same team and the student still as CEO. The founder said growth came from paying about 250 influencers a fixed monthly fee to promote it, not commission on sales. The AI only counted calories; the money came from a problem the founder lived personally, plus paid distribution. Anyone can copy a calorie app, but not that founder's experience.

- **7:20-7:56** — **Korea check for #3.** A Korean developer with no game-industry background used AI to ship a first game in two months. It has about 30,000 downloads, with Japan accounting for 20% of daily users; revenue is undisclosed. App-store fees in Korea can still reach 26% even with external payments, but the path of building and publishing is open. **It works unchanged.** The host's takeaway: if something annoys you every day, it probably annoys other people too.

- **7:56-9:28** — **#2: Base44, the solo-founded company that was sold.** An Israeli developer started this "speak a command, get an app" tool alone as a side project, with no outside investment. Six months later a website builder (Wix) bought it entirely in cash. The figures here are garbled, but the video frames it as a roughly 100B-won-scale deal. In May it kept about 250M won in profit after AI costs, roughly seven times the *annual* average (36.2M won) of a Korean solo small business. "One-person company" is only half true, though. The buyer confirmed eight employees at the time of the deal, many reportedly hired just before the announcement. About 33.6B won of the price went to employees as retention payments, plus earn-outs for future results.

- **9:28-10:37** — **Base44 after the sale, and Korea's barrier.** By August, annualized revenue was about 269B won, about 2.5 times the purchase price, and that revenue belongs to the buyer. Like Polsia, Base44 had an AI-cost problem: gross profit was near zero early in the year. Israeli media expected it to reach about half of sales in the second half. Growth came mostly from word of mouth while the founder built in public on social media. **In Korea, building it is fine; collecting money is the obstacle.** Korea isn't on the payment processor's list of supported business countries, so you'd need to set up a US entity to bill overseas customers.

- **10:37-12:29** — **#1: Medvi, two employees and about 540B won.** Before any press coverage, the FDA sent a warning letter on February 20. Medvi is the company that gathers customers for remote prescriptions of GLP-1 weight-loss drugs. The letter raised two issues:
  - Branding a pharmacy's compounded drug under Medvi's own name, implying Medvi made it.
  - Ads saying it had the same ingredient as a famous obesity injection, implying an approved drug.

  About six weeks later the New York Times profiled Medvi without mentioning the letter. The company is two brothers who started with about 26.9M won. In its first full year it had about 539.4B won in sales, about 87.4B won in net profit (roughly 16% margin) and 250,000 customers, figures the newspaper checked against financial data. The brothers built the website and ads with AI text and video tools. Ads carried a notice that the people shown might be actors or AI-generated. Medvi calls itself a marketing company: doctors, pharmacies and delivery are other companies' licensed operations. The founder projects $1.8B this year, but that is only his own forecast, even though one outlet put it in a headline.

- **12:29-13:39** — **Why Medvi fails in Korea.** It is illegal. Korea's Pharmaceutical Affairs Act (Article 68, Paragraph 6) bans advertising prescription-only drugs, which blocks the model at the customer-acquisition step. Korea's telemedicine system, starting in December, focuses on returning patients. First consultations are limited by region and prescription, and patient and clinic must be in the same area. What Medvi had that others couldn't take wasn't its own asset: it was other people's medical and pharmacy licenses, which the government controls. That is why the FDA letter matters.

- **13:39-15:32** — **Where Korea stands.** In the first half of the year there were more than 560,000 new startups, down 1.5% year over year. Information and communication startups still grew; the exact percentage is garbled but described later as nearly 60%. The government said tech-based startups were at their highest share on record, led by AI-enabled ICT and science/technology services. Korea has over 1.16 million solo businesses with no employees. The largest Korean example the host found is ConnectFit, an app for part-time wage and attendance management:
  - 5 employees and about 1B won in revenue, about 30% of it profit (≈200M won per person).
  - 40,000 partner businesses and 718 app updates.

  Its CEO thinks AI could take teams under 10 people to 10B or even 100B won, but that is his own opinion. Its edge is a real workplace pain point plus years of accumulated revisions. The host thinks 100B won solo isn't realistic in Korea yet, because the revenue gap with the US #1 is more than two orders of magnitude, but the trend points that way.

- **15:32-17:08** — **Scorecard, a three-point checklist, and the close.** Unchanged in Korea: Levels' photo service (if you have an audience) and Cal AI. With conditions: Polsia (AI costs) and Base44 (overseas payment setup). Blocked: Medvi. The checklist:
  1. Is there an annoying problem you know better than others?
  2. Do you have people you can show your product to? If not, you need at least a budget for promotion.
  3. If everything depends on someone else's license or someone else's AI, reconsider.

  The close repeats the thesis: AI made the game, the website and the ads fast, but money went to what couldn't be taken away (Levels' followers, Cal AI's lived problem). Medvi's borrowed licenses made it the most precarious. When you see "earn 10M won a month alone with AI" ads, ask what that person has that others can't take. The video ends by asking viewers what annoying thing they know better than anyone else.

**Practical takeaways:**
- **Distribution matters more than building.** AI makes building nearly free, so you need an audience, word of mouth or a paid promotion budget.
- **Prefer recurring payments to one-off ones.** Levels' subscriptions survived; his one-off ad sales didn't.
- **Do the per-customer AI cost math before launch.** Polsia and Base44 both had near-zero or negative margins from inference costs.
- **Start from a pain point you know personally** (Cal AI, ConnectFit). A problem you've lived is harder to copy than code.
- **For Korean founders:** check overseas payment access (a US entity may be needed), app-store fees, and the regulatory limits on healthcare and drug advertising before copying a US model.

## Audience signal

- No high-signal audience themes available.