A P&L With Zero Revenue
I have built a lot of apps. Not one of them made money.
Servers, domains, CDN, tokens—a few thousand yuan all in, and zero coming back. For a while I could not work out why, and I assumed the product was bad. Then I spread that P&L out and followed every single payment to wherever it landed, and it became obvious: the server money went to the cloud vendor, the domain money to the registrar, the CDN money to the network operator, the token money to the model vendor. Every one of them made out well. I was the only party in the chain who lost money.
This is not a startup-failure story. It is a sample of how this industry actually runs: every layer I depend on is somebody else’s gross margin, and the one layer I own has no pricing power. The costs are certain—rent gets paid every month. The revenue is close to zero, because distribution was never in my hands. The expected value of this business is structurally negative. Not because the product is not good enough, but because this particular seat does not collect money.
That same week I watched two speeches. At Tencent’s anniversary event, Pony Ma’s internal address was unusually blunt: flatter organizations, smaller teams, “half the people hold the base, half go on the offensive,” and 50% of newly written code already AI-assisted. At Baidu World, Robin Li laid out three principles of organizational evolution—break down job boundaries, accumulate tacit knowledge, design closed-loop processes—and said the next act of AI is the intelligent organization. One was an internal speech about crisis, the other an external speech about methodology, and both pointed at the same thing: the unit of competition in the internet industry is shifting from “role/headcount” to “task/loop.”
I was not surprised. Years ago a former boss of mine tried a more radical version: abolish the org chart, go find your own projects. The company was turned into an internal market. The experiment was eventually rolled back, but I remember that dress rehearsal—skip straight to “task-ification” without a platform or a floor and you get chaos, not efficiency. Task-ification is not the abolition of organization; it is the redesign of it. Now the big companies are running that redesign with real money and tens of thousands of people.
The Longer I Work, the More I Depreciate
In traditional industries, experience is a compounding asset. Doctors, lawyers, craftsmen—their judgment accumulates case by case, and the marginal value of seniority goes up. In the internet industry, what I accumulate is three kinds of depreciating asset. One is stack experience: frameworks, tools and languages turn over every few years, so what I am fluent in is knowledge of one specific stack—and that stack is being retired. One is platform experience: the so-called “expertise” is really knowledge of WeChat’s, Douyin’s or a cloud vendor’s rules, and the moment the platform rewrites its rules, my experience resets to zero. One is process experience: knowing how to push a project through the company, coordinate resources, work the internal process—worthless the day I leave, and AI and agents are learning to work that process too.
So seniority does not equal value. The internet pays for output, not for years served. The age-35 crisis is not age discrimination, it is a balance-sheet problem: assets in this industry depreciate so fast that the book value of a 35-year-old with “ten years of experience” can be lower than that of a 25-year-old with “three years”—because the older worker’s assets are split evenly across stack, platform and process, while the younger worker’s are fresher.
AI is an accelerator, not the cause. Code generation voids the entry ticket: the main value of a junior role—writing code, executing—is now AI’s. Penetration rates of 50% or 90% of code are not slogans, they are the recovery rate on those entry tickets. Meanwhile senior “experience” becomes platform lock-in that AI can navigate: the knowledge of “how to do it” now lives in the platform and the agent, not in my head. But the depreciation structure predates AI—stacks were already turning over, platforms were already rewriting rules, processes were already rotting. AI just marks that depreciation to market faster. It turns balance-sheet depreciation into income-statement layoffs overnight.
The Bottleneck Never Moved
Take short dramas. Even if productivity rises—even if one person can be the screenwriter, the designer and the producer—who sells it? Still the platform. Once AI drives the marginal cost of content production down to near zero, supply explodes and perfect competition bids away all producer surplus: if I can make it, so can everyone else, and zero economic profit is the normal state. The scarce factor shifts from “can be made” to “can be seen,” and attention and distribution become the new bottleneck. Ownership of that bottleneck sits with the platform, and the platform’s objective function is revenue, not quality, so marginal slots get allocated by auction. Paid ranking persists stubbornly not because platforms are evil, but because the payment signal is verifiable and monetizable while the quality signal is noisy at scale and gameable. “Recommend by quality” degenerates into recommend by engagement, and engagement can be bought.
So that is how the surplus flows: from individual producers, to platforms and capital. Productivity rises, income does not. Which is why I say AI may still be a growth engine in aggregate, but for the individual it is a transfer engine—it moves surplus out of the hands of labor and producers and into the hands of whoever owns the bottleneck.
My app P&L is the most complete evidence of this: I own not one layer. The means of production are rented—the model is an API, tokens are metered, intelligence is somebody else’s asset; and even if the product took off, token costs scale linearly with users, which is a permanent royalty that the model vendor charges on my product’s core function. The means of operation are rented—servers, domains, CDN, bill every month. The means of distribution are rented—store placements, search rankings, feed traffic, tolls everywhere. All I own is the wrapper in the middle, and that is precisely the layer AI commoditizes most thoroughly.
Who made the money? Three tiers, each one worse than the last. The shovel sellers: servers, domains, CDN, tokens—the merchants of the gold rush, where the prospector carries all the variance and the merchant books risk-free revenue. The gatekeepers: app stores, ad networks, feeds—the toll booths on distribution, collecting whether or not I survive. The dream sellers: the “AI empowers the super-individual, the one-person company” narrative functions as demand-side marketing for token and cloud vendors. Every dream of a one-person unicorn translates into one metered-API customer; the more the narrative convinces me that I could do it too, the steadier their revenue. The old saying about being sold out and then counting the money for the seller is not rhetoric. My self-flattery is not a character flaw—it is their sales funnel.
There has always been an exit, it is just not on the production side. The indie developers who do make money share one trait: not better code, but owning at least one layer—a vertical community and audience (owning distribution), deep knowledge of and trust within one industry (owning differentiation), an existing web of client relationships (owning the channel), or building an asset to sell rather than a cash flow to run (owning the exit). The losing posture is the opposite: rent everything, own nothing.
And one more thing: the bottleneck does not stay still, it gets re-auctioned at a new layer. The day people start telling their AI assistant “recommend me a short drama,” the distribution entry point migrates from the feed to the assistant layer, and paid ranking migrates from the search engine to the model. What changes is the rent collector, not the toll booth.
The Spiral Nobody Calls Off
AI is driving mass layoffs and cost-cutting, and wages are being compressed. It used to be twenty thousand yuan, now it is fifteen, and later maybe only ten. Wage compression is not a side effect of AI—it is the main channel of the transfer. Wages equal marginal output times bargaining power, and AI’s real blow is not taking the job, it is destroying my outside option: the company can replace me with an agent at any time, so my bargaining power goes to zero. Going from twenty thousand to ten does not require an actual layoff; the threat is enough. That makes it more efficient than layoffs—layoffs require severance, wage cuts do not.
With no money in hand, people instinctively stop consuming, or downgrade the quality of what they consume. So companies do not make money, and ordinary people have no money either. This is the fallacy of composition: cost-cutting looks rational for each firm in isolation and is suicide in aggregate—my cost is somebody else’s revenue, and one company’s “efficiency gain” is another company’s vanished order book. And the gap cannot be filled by the rich: their marginal propensity to consume is low, and eating six meals a day does not absorb the capacity of an industrial chain. Keynes called it the paradox of thrift; the folk version is “we all rot together, we all die together.” Same thing.
Quality decay has two engines. One is the race to the bottom: when demand shrinks, competition shifts from “make it better” to “look cheaper,” and quality gets cut to protect margin. One is bifurcation: companies abandon the collapsing mass market and attack the affluent top instead—so the top gets luxurified and the bottom gets worse. What feels like “everything is rotting together” is really the experience of the squeezed middle.
I stared at that thirty-thousand-yuan phone for a long time, and the numbers never added up. With industrialization this advanced and informatization this advanced, how does a “bad” phone cost thirty thousand? Then I figured it out: I am not paying for manufacturing. Manufacturing really has been driven down to the price of cabbage. What is expensive is the stack of rent layers on top of it—chip monopoly, patent licensing, brand premium, channel margin, installment finance. The price of a phone equals a little manufacturing cost plus a pile of tolls. And it was never priced for people who cannot afford it: the anchor is the willingness to pay of the top tier, decoupled from mass wages. Once demand bifurcates, prices stop listening to the majority. Is that reasonable? Rational for every rent collector, irrational for the system as a whole. Fallacy of composition again.
But I leave this spiral one exit: it is political, not physical. The 1930s ran the same script, and what ended it was not technology but a reset of distribution—the New Deal, the postwar contract: wages linked to productivity growth, progressive taxation, unions. The spiral stalled at the moment distribution was re-linked to production. So when I saw that highly-upvoted answer on Zhihu—“tie taxation to employment; companies that do not solve the employment problem pay more tax”—it sounded radical, but the direction is right. Rotting together and dying together is not destiny. It is the default outcome of policy inaction.
I also force myself to push back two steps. First, the spiral is a transition-period risk, not a permanent law: historically technology eventually created new jobs and new demand. But the “new jobs” of this round may be capital-owned agents, and the automatic correction mechanism is weaker than in any previous round, so the window is narrower and more political. Second, “everything getting worse” is concentrated in the middle: the price war at the bottom has actually made cheap goods better—a thousand-yuan phone today is not bad—and the top is luxurifying. The feeling of rot is the exclusive experience of the squeezed middle, not a physical fact about all goods.
My Three Questions
So now, before I write the first line of code, before I accept the next offer, I answer three questions first: Which layer do I own? Which loop can I accumulate? Whose trust am I compounding? If all three answers are “none,” then the project is not a startup—it is consumption. Paying rent to buy an illusion.
At the operational level there is only one line: accumulate loops, not years. Output that is measurable, repeatable and amplifiable by AI is the asset under the new unit of account; my résumé should be a list of tasks and loops, not a list of titles. There are only two directions with leverage: move upstream and own the bottleneck—channel, trust, community—or move to the capital side, so that AI works for me instead of replacing me. The layer in the middle, where workers and small producers trade capacity for money, is exactly the layer the surplus gets transferred out of.
The internet is not about to disappear; it is about to be repriced. What changes is not the industry but the unit of account: from “years/role” to “task/loop.” What depreciates the longer I work is experience-as-asset; what appreciates the longer I work is judgment-as-asset. The great upheaval is not changing the sky. It is changing the scale. And standing behind that scale is the question nobody can dodge: when the productive forces advance and the relations of production do not, who re-links distribution to production? That is the script for the next act.
评论
0 条评论