The dividend window is open; programmers, don't stockpile courses from the sidelines

You've probably seen these kinds of programmers.

Someone posts an app written with AI, and the top comment is: “Too much AI flavor.” It gets the most likes.

Someone was just laid off, reposts an article titled “AI Is Just Another Capitalist Scam,” with the caption: “Saw through it a long time ago.”

Then there's the most common type, and the one that hurts most: on the subway after work, they scroll past “How ordinary people can use AI for a side hustle” and save it; on the weekend they buy an Agent course, then a prompt engineering course; their collection folder holds dozens of tutorials, but almost none have actually been opened. Every day they feel like they're falling further behind, and every day they're “getting ready to start.”

Let me state my position first, so you don't think I'm here to sell courses: I myself have a pile of projects that are “alive, usable, and have made no money” - they run on free tiers, and nobody has ever paid. So this is not success coaching; it's a judgment made after checking every number directly: at this moment, the best move for an ordinary developer is to use AI to get a commercial result and complete your own initial capital accumulation. Not to rate other people's work, not to be pushed by anxiety from lesson a to lesson n, only to end up empty-handed in your bookmarks.

I. The dividend is not a trend, it's a time gap

Many people understand “dividend” as a “trend,” as if the moment you step on it you can fly. This understanding is harmful: it makes people either charge in blindly, or conclude the trend is fake because they saw someone fall.

In business terms, a dividend is a time gap: the cost of production falls faster than the market price falls. During this period, customers are still pricing results according to old costs, while you are already producing at new costs. The difference in between is profit.

Gross margin window = the price customers anchor to old costs - your new production cost

How fast has this round of cost decline been? a16z measured it: for models at the same capability level, inference costs fall about 10x per year. In November 2021, GPT-3-level capability (MMLU score 42) cost $60 per million tokens; by the end of 2024, the same level cost $0.06, a 1000x drop in three years. Epoch AI's later measurement was even more aggressive: looking only at data after January 2024, the median annual decline jumped from 50x to 200x.

And price? The price anchor in customers' minds is still stuck in the era of “this requires paying someone.” Price is set by perception, and perception iterates far more slowly than token prices.

For people who write code, this time gap has an even more intuitive version: five years ago, making the same product required a frontend developer, a backend developer, and a designer, and months of salaries; now one person plus an AI toolset can squeeze annual costs down to a few thousand yuan. The cost of a single attempt has dropped by several orders of magnitude - this is the most developer-friendly part of this dividend: you don't need to bet everything; you just need to try a few more times.

The original text of a16z's LLMflation says one thing: same capability, 10x cheaper every year.

II. The window is real: don't look at sentiment, look at who is paying

The simplest way to judge whether something is a scam is to see whether anyone is paying real money. Sentiment can be faked; payment records are hard to fake.

Stripe analyzed the 100 highest-revenue AI companies on its platform: the median time to $1 million annualized revenue was 11.5 months, 4 months faster than the fastest SaaS cohort of the golden age; reaching $5 million annualized revenue took 24 months, while top SaaS companies in 2018 took 37 months. Individual cases are even more extreme: Lovable hit $17 million ARR in 3 months, Bolt hit $20 million ARR in 2 months, Cursor reached $100 million ARR in three years.

One-person companies also put up strong numbers. Danny Postma runs HeadshotPro, an AI professional headshot product, alone; annual recurring revenue is about $3.6 million. Maor Shlomo's Base44 was acquired by Wix for $80 million in cash six months after launch. Pieter Levels has no employees and no funding, earning over $3 million a year.

There are domestic examples too. Manus launched in March 2025 and passed $100 million annualized revenue by end of year; in December 2025, Meta announced an acquisition for around $2 billion, and on April 27, 2026, the National Development and Reform Commission called it off. Kling's rhythm is even more dramatic: on March 25, 2026, OpenAI announced it was shutting down Sora; on the same day, Kuaishou released its earnings - Kling's Q4 2025 revenue was 340 million yuan, December alone passed $20 million in monthly revenue, annualized revenue run rate was $240 million, and full-year revenue was about 1.04 billion yuan. Same track, same day: one is exiting, the other is taking money.

Who	Number	Nature
Stripe platform Top 100 AI companies	$1M annualized revenue median 11.5 months, 4 months faster than fastest SaaS	Stripe annual report
Lovable / Bolt	3 months $17M ARR / 2 months $20M ARR	Same as above
HeadshotPro	One-person company, about $3.6M annual revenue	Founder self-reported
Base44	Six months after launch, $80M cash sale to Wix	Wix official announcement
Manus	8 months after launch, annualized revenue passed $100M; Meta acquisition halted	Multiple reports
Kling	2025 Q4 revenue 340 million yuan, ARR $240M; Sora exited same day	Kuaishou earnings

These are all companies, far from an individual. But note the common thread: they all earn the gap between the “old price anchor” and the “new production cost.” HeadshotPro does the same thing as Mi Ya Camera, which went viral in 2023. The difference is that it sells to working professionals as LinkedIn headshots, a narrower, more essential use case, and customers are willing to pay more for “looking professional.”

Stripe's official blog says it plainly: AI companies are rewriting the rules of growth.

The headline of the Securities Times report was the whole story: on the same day Sora exited, Kling set a pledge to double revenue.

III. First, some cold water: more than half of products haven't made money

If you're getting fired up now, it's time for cold water. Everything above is survivors, and survivorship bias is the biggest scam in this industry, even bigger than the sentence “AI is a scam.”

The more honest numbers come from TrustMRR. The platform connects directly to Stripe and other payment channels, showing verified real revenue. In September 2026, a researcher scraped all 10,150 products on the platform:

Metric	Number
Never had any revenue	26.1%
Revenue in the last 30 days	49.1% (median $169)
Annualized revenue reaching $100K	2.4%
Revenue share captured by top 1% of products	76.1%

Two details deserve attention. First, AI products and non-AI products have almost the same probability of having revenue (49.3% vs 49.0%); median monthly revenue is $161 vs $175 - AI changes the cost of making products, not the basic question of whether anyone buys. Second, another dataset of 8,817 products scraped in August found that developer tools are the worst zero-revenue category: 69% received nothing. The people making tools are the most numerous, and the people getting paid are the fewest.

In China, it's an enlarged version of the same table. The AI comic-drama market was 16.8 billion yuan in 2025, which sounds like a trend; but in the first half of 2026, 221,900 new titles were launched on Douyin, and 98.7% didn't reach break-even. Costs fell 90%, barriers fell, and everyone flooded in.

Mi Ya Camera is a specimen of the window closing. In July 2023, it exploded overnight on the strength of “9.9 yuan for photo-studio-style photos”; peak queues had more than 4,000 people waiting ten-plus hours, and daily active users passed 600,000. Then similar products swarmed in, and users' price anchor shifted from “a photo studio costs a few hundred yuan” to “AI photo, isn't it just a few yuan.” The team disbanded at the end of September 2025; from viral hit to disbanding took less than two years. The window is real, and it closes faster than you think. The people who mocked “AI photos all look alike” in their social feeds that summer didn't get a single yuan.

Tech Startups' analysis of the full TrustMRR dataset; the headline is the answer.

CNR: a 16.8 billion yuan market and a cost of 500 yuan per minute, written in the same report.

Chongqing Daily republished Red Star Capital Bureau: the team disbanded at the end of September 2025 and maintained only minimum operations.

IV. Why we would rather hoard courses than get on the field

The window is real, and so is the failure rate. Rational behavior would be to increase the number of attempts and reduce per-attempt cost. But most people choose another path: commenting, and anxious studying.

Commenting costs zero. Saying “too much AI flavor” requires no skill, and instantly gives you an aesthetic sense of superiority, with likes delivered in real time. Getting on the field costs something real: time, exposing your work, facing a high chance of failure and other people's mockery. Between “zero cost, instant reward” and “real cost, delayed reward, high probability of failure,” inertia chooses for you.

“AI flavor” as an aesthetic judgment is completely valid; I also dislike that one-shot generated tone. But as a business judgment, it's almost meaningless: customers pay for the result - was the problem solved, was time saved, did the money come back. People who criticize “AI flavor” are evaluating the work; people who pay are evaluating the result; the two groups are not even in the same market. What's more, “AI flavor” is an engineering problem; models improve every generation, and the moat you defend with it gets shallower every year.

More worth discussing than commenting is anxious studying. Behavioral economics has a classic sample: DellaVigna and Malmendier tracked 7,752 members at three American gyms for three years. For members choosing monthly fees above $70, the average was 4.3 visits per month, which worked out to more than $17 per visit - when they could have bought a 10-visit pass at $10 per visit. On average each person overpaid by about $600. They overestimated their future selves.

Buying courses is the same psychology: at the moment of payment, what you buy is not capability, it's the reassuring feeling that “I'm already trying.” MOOC data is even blunter: a 2019 Science study by Reich and Ruipérez-Valiente analyzed 565 MIT and Harvard courses on edX with 12.67 million registrations, and only 3.13% earned certificates; among paying students, completion rates were 44% to 56%. The difference is not course quality; it's whether there is a goal and a stake.

Now count the time. The “China Online Audiovisual Development Research Report (2025)” shows average per-capita daily short-video use of 156 minutes, about 949 hours a year, roughly half the annual hours of a full-time job. You say you have no time for a side hustle; the time is there, it's just spent watching videos of “ordinary people earning a monthly income of 10,000+ with AI.”

For programmers there is an even more uncomfortable point: the tools of this era change every few months. A framework you spend three months learning may be obsolete the day you finish. Learning without customer demand is like running backward on an accelerating conveyor belt. The truly useful way to learn is on demand: when a customer needs a feature, learn how to build it; when a product gets stuck on distribution, learn how to run ads. Every bit of learning is tied to a specific problem that someone pays for.

While we're at it, take apart the two most common sideline remarks.

“AI took my job.” This is real and should not be dismissed lightly. In a Stanford study updated in August 2026 by Brynjolfsson's team: in AI-high-exposure occupations, employment for 22-25-year-olds is 19% lower than for peers in low-exposure occupations; a year ago the gap was 13%. But the same study says: the overall economy has not seen mass unemployment, and experienced people do not show a similar gap - the impact is concentrated in entry-level jobs that depend on “codifiable knowledge,” and it happens mainly through hiring fewer people, not firing more. Translated into business logic: AI lowers the price of “selling standardized labor by position”; judgment and the ability to be responsible for results have not had their price lowered. The rational response is to move from “being priced labor” to “someone who prices results.”

“AI is a scam.” As of December 2025, China had 602 million generative AI users, a penetration rate of 42.8%, growing 141.7% in one year. Valuations can have bubbles, many companies will die, many narratives will collapse; I agree with all that. But a bubble and a scam are two different things: when the internet bubble burst in 2000, a large number of companies died, and Amazon did not disappear. Calling a bubble a scam is logically invalid, and its only effect on action is to give yourself a respectable reason not to get on the field.

Pew's June survey wraps up this section: 52% of American adults say they are “more concerned than excited” about AI, up from 37% in 2021; among AI experts, 47% are “more excited than concerned,” while only 11% of the general public are. The closer you are to the technology and the more you use it to do work, the more optimistic you are; the farther away you are, the more anxious. Anxiety correlates with distance, not much with truth.

Stanford Digital Economy Lab's paper updated in August: the fact sits before the question mark in the subtitle.

Pew June survey age-group data: the 18-29 group, most optimistic five years ago, now has the highest share of concern.

V. People who don't get on the field have a success rate of exactly zero

Put the data above into a simple model. Suppose each product attempt has probability p of reaching $100,000 annualized revenue. According to TrustMRR, p is about 2.4% - already optimistic, because people willing to publish revenue are themselves optimistic. After n attempts, the probability of at least one success is:

P(at least one success) = 1 - (1 - p)^n, p ~= 2.4%

Number of attempts n	Probability of at least one success
0	0%
5	11.5%
10	21.6%
20	38.6%
30	52.1%
This model is crude: it assumes attempts are independent and the success rate stays constant, which is not reality. But most deviations from reality favor the person trying: the 1,045 TrustMRR founders who had made more than two products had a clearly higher success rate than those who made only one - the number of attempts itself raises p.

The most important number on this table is the 0% in the first row. No matter how many courses you study, how many tutorials you bookmark, how many times you say “too much AI flavor,” in this formula n equals 0. People who do face a low but nonzero probability that rises with experience; people who don't face a certain number: zero.

How do you add n? Look at the per-attempt cost. My own money-losing projects all run infrastructure on free tiers; fixed cost is just a few hundred yuan a month in tool subscriptions. Five years ago, the same product would have needed at least a frontend, a backend, and a designer, months of salaries.

This is why I say “initial capital accumulation” is the best move right now. Initial capital is not just money; it's the first batch of customers, the first presentable case study, true understanding of a niche market, and the feel of “I've made something work once.” More importantly, money buys time: with the first chunk of initial capital, your next choice of what to do no longer has to be pushed by anxiety and rent. In a period of such rapid change, being able to choose calmly is itself the scarcest asset.

VI. The smallest action for a developer to get on the field

Compress all the data above into one executable sentence: pick a specific problem someone is willing to pay for, build it at the lowest cost, and start charging on day one.

Why charge on day one? Because charging is the only honest validation. Likes, bookmarks, “that's great” don't count; only payment counts. Almost all the improvement in p comes from getting more accurate at judging what people buy, and that judgment can only be learned from real payments and real rejections.

Why pick a specific problem? Because a broad “AI tool” is not something people buy, while a specific “saving me one photo-studio session” is. HeadshotPro only does professional headshots, not portraits, not photo retouching, not ID photos.

Why use the lowest cost? Because you will most likely fail, more than once. The lower the cost, the larger the n you can afford.

Before starting, three checks to help you roll fewer doomed dice:

1. Price anchor check

Is the price in the customer's mind still anchored to old costs? Look for things that “used to require hiring someone and AI can now do 80% of it”: studio portrait photography, outsourced animation, product detail pages, industry reports, small internal corporate tools. If customers already assume “AI can do this for a few yuan,” the window is closed; don't enter.

1. Repeat purchase check

Will the customer pay a second time? Mi Ya Camera failed because of a one-time payment: after paying 9.9 yuan for photos, users had no reason to come back, and growth depended entirely on new users. Prefer what can be billed monthly over what is billed per use.

1. Crowdedness check

Do you have something others can't get into quickly? The cost of AI comic-drama fell by 90%, and 220,000 titles appeared in half a year. Tooling lowers barriers for everyone. What actually lets you survive in a crowd is often not AI, but the things outside AI: five years of experience in an industry, the twenty customers in your hand, the niche population you understand better than others. AI is a lever; the lever has to rest on your own fulcrum.

As for the from-a-to-n study list, delete it. Keep only the one item your current product actually needs, learn it, use it. The rest can wait until customers force you to learn it.

VII. Three rebuttals, answered in advance

“You're cherry-picking survivors.” Correct, which is why Section III included that table: 26.1% of products have zero revenue, and only 2.4% reach $100,000 a year. I never said that getting on the field means success; I said that not getting on the field definitely means no success. The difference between those two sentences is the whole point of this piece.

“At the end of the day you're also selling anxiety.” I've thought about this seriously. The difference is this: people who sell anxiety end with a payment link and ask you to buy their antidote; this piece has no payment link at the end. The only advice I give is to find someone willing to pay you. If reading this made you more anxious, close it and go do a small thing that can earn 9.9 yuan. Anxiety disappears faster by doing than by reading ten articles.

“Not everyone is suited to doing business.” This one I admit; I can't answer it. Some people have mortgages, have children, and have no room to make mistakes; some people are naturally bad with people, and collecting their first payment is worse than overtime. For them, the cost of “getting on the field” is far higher than my numbers suggest. I'm leaving this question at the end, for you.

Finally

Every era has its dividend. The PC era had one, the internet era had one, the mobile internet era had one, and now it's AI's turn. Every time, a dividend is not universal wealth; it is left only for those who were ready and who actually started. While the window is open, most people are commenting, anxious, studying; after it closes, most people start saying “I knew it all along.”

You can laugh at the people who act, laugh at their AI-flavored work, laugh at their small money. But remember the formula in Section V: every time they try, n increments by one; every time you mock, n is still zero. What you miss is not a few get-rich moments, but a road that could have been tried again and again and slowly walked through.

I've organized the primary sources for every number in this piece (a16z's original, Stripe's report, the TrustMRR analysis, the Stanford paper, Kuaishou earnings coverage). Reply “hongli” in the backend and that's it.

Are you on the field or on the sideline right now? When was the last time you got paid by your own product? Leave a comment and tell me. Next issue, I'll pick the scenario mentioned most often and break down how to do it from start to finish.