Why a Rooftop Developer Delivers Food Every Night

Opening

Dear reader, on the rooftop of an old residential building in Chengdu, China, there’s a house that doesn’t officially exist. It’s an unauthorized two-room structure built on the roof, and the rent is 1,000 yuan — about ₩200,000 in Korean currency. There’s no lease; the tenant simply wires the money each month to the neighbor living downstairs. A 28-year-old developer couple has lived here for three years.

The husband, Jared, used to be an engineer building AI at a Big Tech company. After enduring the 996¹ work schedule, he quit, and now spends his days writing freelance code in the rooftop room. Then at 6 p.m., he puts on a delivery vest and gets on his motorbike. Riding for four hours until 10 p.m. earns him about 1,000 yuan a month — exactly the same amount as his rent.

I recently watched a documentary following a day in his life, and I couldn’t dismiss it as just “one eccentric young man’s story.” Here’s my conclusion up front: this isn’t one person’s strange choice — it’s a bill that every individual worldwide who sells technical skills for a living has been handed. And the very same bill has already arrived in Korea.

---

Code by Day, Delivery Bag by Night

Calling Jared’s household frugal would be an understatement. In a country where 98% of urban households have air conditioning, he has none — in summer, he cools the walls by spraying them with a water mister. Meals are pre-packaged rice boxes costing 5 yuan, about ₩1,000, per pack. He fixes broken appliances himself and even assembles his own speakers from parts. With income this unstable, “don’t buy what you can build” has become a survival rule. Together with his wife, who works at a hospital pharmacy, they bring in about $1,000 a month — roughly ₩1.4 million.

What’s worth noting is that this income has actually dropped sharply compared to his first two years as a freelancer. It’s not that the work disappeared — the rates collapsed. Jared knows exactly why, better than anyone, because he used to be the one building the AI. As he puts it: “The more I understand AI, the more frightened I get. Getting it to do something has become far too easy. All you need is one prompt. And anyone who figures out that prompt can build the exact same thing.”

Because coding income alone isn’t enough to live on, he delivers food every evening. During the evening peak hours, he earns 23 to 30 yuan an hour, roughly ₩5,000. His personal best was 50 yuan an hour, about ₩10,000, earned riding through a downpour so heavy that other riders had given up.

One scene from the filming stood out. Jared made a delivery to a hospital, and the person who’d ordered turned out to be a fellow rider who’d broken his leg in a delivery accident. When the documentary was filmed in May 2026, China’s occupational injury protection² program for platform workers — essentially China’s version of workers’ comp — was still only a regional pilot. Nationwide rollout didn’t begin until July 1, 2026. That rider was injured two months before the program took effect, and had to cover roughly 20% of his treatment costs himself while going months without income.

Even so, Jared doesn’t blame the platforms. “There’s an oversupply of riders, so of course platforms push rates down as far as they can. Cutting costs is just what capital does.” Instead of getting angry, he calculates. And he adds: “Once you’ve tasted freedom, you can’t go back.” The documentary appends one cold-blooded line to that: freedom without an economic foundation can always turn out to be fake freedom. The moment you have to deliver food just to make rent, part of that freedom has already been surrendered.

---

Laziness Isn’t What Cut His Rates

Reading Jared’s falling income as “personal laziness” misses the point entirely. His rates are being squeezed between two enormous forces.

The first force comes from the demand side: AI is shrinking the pool of available work itself. A recent study published in the journal Management Science measured this with striking precision. The research team tracked 1.4 million job postings on global freelance platforms over two years starting in July 2021, and found that after ChatGPT launched, postings for automation-vulnerable occupations fell by an average of 21%. Writing jobs dropped 30.4%, and software/web development — Jared’s own trade — fell 20.6%. Graphic design requests fell 17% after the arrival of generative image AI, with no sign of recovery by the end of the study period. This is data from online outsourcing platforms, so it isn’t representative of the entire market, but because prices there are disclosed in real time, it functions like a thermometer that registers the AI shock first.

The second force is supply. China’s flexibly employed population is estimated at roughly 200 million, of whom 84 million are classified as “new-type gig workers” — people receiving jobs through platform apps. Meituan alone had 7.45 million riders take orders in a single year, and here’s the striking part: nearly half of them worked fewer than 30 days that year. That means delivery has stopped being an “occupation” and become a nationwide standby side gig. With millions of people ready to jump in whenever they need extra cash, rates are structurally unable to rise. Youth employment pours fuel on the fire: China’s unemployment rate for ages 16 to 24 stood at 15.6% as of May this year, and recent reports suggest it could climb to 20% as AI spreads further.

Here’s the summary: AI cuts demand, and the downturn expands supply. The individual caught in between gets squeezed from both sides. Full-time riders working high volume in tier-1 megacities like Beijing and Shanghai earn an average of 7,354 yuan a month — about ₩1.4 million. The fact that this is almost exactly what a former Big Tech AI engineer couple earns together in a month shows just how far this squeeze reaches.

---

The Same Bill, Delivered to Korea

Saying “Korea is actually similar” might sound like a hunch, but the numbers are already speaking for themselves.

People who file “personal service income”³ with Korea’s National Tax Service under the 3.3% withholding rate — essentially, “freelancers” in the broad sense as captured by tax statistics — grew from 3.26 million in 2008 to 8.47 million in 2022. This happened even as the share of wage workers in employment statistics actually rose over the same period. Labor market researchers read this gap as a sign that “employment that looks like employment but isn’t” is expanding — a quiet swelling of people who appear to work inside a company but are contractually classified as sole proprietors.

People receiving work through platforms also grew 11% year-over-year to 883,000 as of 2023. What’s worth noticing is the breakdown: delivery and driving actually declined, while IT and professional services workers surged 141%. In other words, Korea’s gig economy is already spreading from motorbikes to laptops. The number of “N-jobbers” — people juggling side gigs on top of a main job — also hit an all-time high of about 680,000 in 2025.

The delivery scene here looks even more like Chengdu’s. Korea’s food delivery market hit an all-time high of about ₩40 trillion in 2025, yet riders on the ground report their per-delivery take-home pay has fallen from around ₩5,000 to as low as the low ₩3,000s. In a large-scale survey conducted in 2023, 63% of riders said their income had decreased. It’s exactly the same structure: the pie grows bigger while the slice for the person running on top of it shrinks.

There are differences, of course. Korea’s institutions moved faster. In July 2023, Korea abolished the exclusivity⁴ requirement for workers’ compensation insurance, extending coverage to most delivery riders — three years ahead of China’s nationwide rollout in July 2026. But being ahead institutionally doesn’t mean the rate structure is any different. The distance between a rooftop room in Chengdu and a studio apartment in Seoul is three hours by plane, but in labor-market time, the gap is a matter of years — maybe even less.

Finally, I want to flag one scene that runs directly against these numbers. These days, Instagram and Threads are flooded with people calling themselves “solopreneurs,” flaunting ₩10 million a month, hundreds of millions of won a year in revenue. Bios reading “mentor,” “high-net-worth individual,” or “successful CEO” are everywhere. But look at the closest official statistic — the National Tax Service’s data on registered one-person media creators — and only 34,806 creators reported income for tax year 2024. The bottom half of them earned an average of ₩24.63 million a year, or about ₩2 million a month in gross revenue before expenses. The top 1%, just 348 people, take home 18% of all that income. In a world of 8.47 million personal-service filers, “ten million won a month” is a statistically vanishingly thin sliver at the very top. So doesn’t that seem strange? Is it really pure coincidence that all these rare people happen to be gathered on Threads and Instagram?

---

Oz’s Lens

Honestly, this documentary didn’t feel like someone else’s story to me. I’ve effectively been running a one-person consulting business since 2024 myself. People often tell me, “you must be so free,” but after two years of doing this, here’s what I’ve learned freedom actually is: it isn’t a state, it’s a fixed cost. You’re free only for as long as you can keep covering what goes out every month. Jared’s fixed cost was 1,000 yuan in rent, and delivery was simply the means he used to cover it.

Let me answer the question I raised earlier. No, it’s not a coincidence. When I teach data courses, the very first question I make students ask is: “How was this sample drawn?” A social feed is not a random sample. It’s a self-selected sample where only people with a reason to display success show up at all. And that “reason” is the whole point. A great deal of these income-proof posts aren’t performance reports — they’re ad creative. The structure isn’t “I earned ₩10 million a month by selling a product”; it’s “I earned that by selling an e-book or course titled how to earn ₩10 million a month.” It’s the old gold rush story — the ones who got rich weren’t the miners, they were the people selling pickaxes — with the stage simply moved to Reels and Threads. Sure, some of it is real. But the real ratio lives in the statistics, not in the feed.

That’s why I think the buzzword “solopreneur” badly needs a dividing line. Someone who sells their time and skills, and someone who owns a product, a customer base, and a distribution channel, are running completely different businesses even if both are technically “one person.” The former ends up in a price war with AI, because the price of their skill converges toward the price of “one line of a prompt.” The latter uses AI as leverage. Sam Altman’s talk of a “one-person unicorn” is a story about the latter, not a future available to the former. The dream-selling industry sells precisely this confusion — showing people who are, in effect, freelancers, a future that belongs to owners, and charging course fees to bridge that gap.

Jared instinctively understands this distinction too. He says he plans to shift away from taking freelance jobs and toward building his own product, or a startup’s product. That’s the right direction. But there’s one cruel fact: crossing that line takes time, and the rent eats up that time first.

---

Closing

To sum up: freelancers in the AI era aren’t going broke because they’re lazy. Their prices are being squeezed by a structure where AI cuts demand and a downturn swells supply. This structure isn’t confined to a rooftop room in Chengdu — it has already arrived in Korea, a country of 8.47 million personal-service filers. The success stories flooding the feed aren’t a refutation of those statistics; they’re closer to evidence that the success story itself has become a product. And because freedom is a fixed cost, not a feeling, it can turn into fake freedom at any moment you lack the system to cover that cost.

Here’s something to try this week: split your income into exactly two lines — “money earned by selling my time” and “money earned by an asset I built (a product, content, a system).” If that second line is zero, chances are the freedom you have right now is freedom you’re buying by the hour.

What about you? If you’ve stepped outside an organization, or you’re weighing whether to go independent, tell me in the comments what unexpected “fixed cost of freedom” you ran into. Your two-line income breakdown works too. I’ll continue this story in the next issue with readers’ own stories.

---

💬 If you’ve experienced the “fixed cost of freedom,” tell me about it in the comments — I’ll fold it into the next issue. 📨 If you have a colleague weighing independence, pass this along to them.

---

References & Further Reading

Primary sources

• No Boss, No Money: The Raw Reality of China’s Gen-Z Freelancers··· This is the video following a day in Jared’s life. Filmed in Chengdu in May 2026.
• Demirci, O., Hannane, J. & Zhu, X., “Who Is AI Replacing? The Impact of Generative AI on Online Freelancing Platforms”, Management Science, 2025. Link ··· This is the study that forms the backbone of today’s piece — it measured, using 1.4 million job postings, exactly which kinds of work shrank after ChatGPT and by how much.
• 每日经济新闻 (via China Economic Net), “Meituan’s 7.45 Million Riders: Their Income Revealed”, 2024.9. Link ··· A news report on a Meituan study detailing the real income picture for 7.45 million riders. It also explains why the rumor that “80,000 delivery riders hold master’s degrees” is an exaggeration.
• Ministry of Employment and Labor & Korea Employment Information Service, “Platform Workers Reach 883,000, Up 11.1% Year-on-Year”, 2024.8. Link ··· This is the official scale of Korea’s platform labor force. The surge in the IT sector is worth reading alongside this Etnews report.
• Park Young-sam, “The Rapidly Growing 3.3% Personal-Service Workforce: The Widening Gap Between Employment and Tax Statistics”, Maeil Labor News. Link ··· This is the source of the “8.47 million personal-service filers” figure — a column that reads the invisible freelancing trend into the gap between these two statistics.

Background

• Xinhua, “Expanding Nationwide in 2026! Occupational Injury Protection Pilot Program Launches Expansion”, 2025.7. Link ··· This lays out how China’s platform workers’ comp program evolved from a 2022 pilot to nationwide implementation in 2026.
• Korea Institute for Health and Social Affairs, “Policy Tasks for Protecting China’s Platform Workers”, Global Social Policy Brief. Link ··· This shows China’s institutional struggles in an era of 200 million flexibly employed workers.
• Korea Business Review, “The Paradox of a ₩40 Trillion Delivery Market: Why Are Rider Incomes Moving in Reverse?”, 2026.6. Link ··· A briefing laying out how Korea’s delivery market size and rider pay per delivery are moving in opposite directions.
• Seoul Shinmun, “YouTubers’ Average Annual Income Tops ₩70 Million — And the Top 1% Take Home How Much?”, 2026.2. Link ··· This is the National Tax Service’s data on one-person media creators (tax year 2024). The gap between the top 1% and the bottom half puts numbers to the gap between the feed and the statistics.
• The Scoop, “The Cry of a Generation for Whom Side Gigs Became the Default: What 2030 N-Jobbers Mean”, 2025.12. Link ··· An article covering the generational landscape of Korea’s 680,000 N-jobbers, per National Data Agency figures.
• Kyunghyang Shinmun, “Designated Drivers and Delivery Riders to Get Workers’ Comp Starting in July”, 2023.6. Link ··· This shows exactly what changed when the exclusivity requirement was abolished.
• Newsis, “China’s Youth Unemployment Hits 15.6% in May”, 2026.6. Link · Ajunews, “China, World’s No. 1 in AI Adoption, Could See Youth Unemployment Reach 20%”, 2026.7. Link ··· These are the Chinese youth employment indicators behind the oversupply story.
• AI Times, “Is the ‘AI One-Person Unicorn’ Becoming Reality?”, 2025.4. Link ··· This sits at the opposite end of the solopreneur discourse — the story of the “$1 billion one-person company.” Reading it alongside today’s piece sharpens the gap between the two.

Related Past Issues Worth Reading

• People Switching Off Their 20-Watt Brains ··· This piece covered how AI is reshaping individual judgment. If today’s piece is about your price tag, that one is about your capacity to think.
• Lockdowns Built China — Now China Is Locking Its Doors ··· This is an analysis of the same China, viewed at the national level. Read alongside today’s rooftop-level view, it adds real depth.

---

The author, Kwangseob Ahn, is a professor of business administration at Sejong University and lead consultant at OBF (Oswarld Boutique Consulting Firm). He teaches statistics and data analysis, including business data management and business analytics, while leading GTM and AI strategy consulting in the field, designing the seam between technology and business. He has published academic research on a memory architecture for AI dialogue systems (HEMA) and runs Daily Arxiv, a daily curation of global AI papers. He holds a master's from Korea University's Graduate School of Technology Management and a KMBA. He is the author of Homo Brainless: The People Who Outsource Their Thinking.

Footnotes

1. 996: A long-hours work practice common in China’s tech industry — 9 a.m. to 9 p.m., six days a week. Courts have ruled it illegal, but the practice persists regardless. ↩

2. Occupational injury protection (职业伤害保障): China’s version of workers’ compensation insurance for delivery, ride-hailing, and freight platform workers. It was piloted in 7 provinces starting in July 2022, had enrolled a cumulative 12.34 million people by June 2025, and expanded nationwide on July 1, 2026. ↩

3. Personal service business income (3.3%): A Korean income-filing method in which an individual performs work without an employment contract and has 3.3% of the payment withheld as tax. People captured in this statistic are, in effect, “freelancers” in the broad sense. ↩

4. Exclusivity requirement: A condition that limited workers’ compensation coverage to those working primarily for a single company. Riders who worked across multiple apps were excluded from coverage because of this rule, which was abolished in July 2023. ↩