Josh Cutler
← Back to writing

Supply Without Demand

Two posts I have stopped being able to scroll past. The first is a screenshot of a small web app, captioned with the “triumph” of having vibe-coded it in a weekend. The replies are mostly other people sharing their own apps or boredom/cynicism with the whole process (spend any time in Magic the Gathering subreddits and you will see the active pushback). The second is a twenty-minute AI-generated anime film, posted by someone announcing that they have discovered they were always meant to be a filmmaker. The replies are mostly other people posting their own films. I’m not going to link to individual posts because I’m not actually mad at the posters (if you’re not seeing these types of posts I envy you), but I think they are missing the point.

The point is mechanical, and the mechanics work like this: supply went vertical, demand mostly stayed horizontal and in a few markets even bent down. Per-piece attention is collapsing toward zero and that collapse is what the word democratization actually delivered. Refeudalization is the more honest word. This essay is the long version of that paragraph.

What was promised

Every AI lab/product/platform now leads with the word democratization. Anthropic, OpenAI, GitHub, Cursor, Lovable, Sora: anyone with a laptop and a credit card now has access to tools that previously belonged to specialists. Lovable’s 2026 mobile launch lets users vibe-code apps from a phone by voice or text. OpenAI’s Sora 2, launched in September 2025, lets anyone generate cinematic video and drop their own face into the result.

The pitch is mostly true on its own terms in that production really has gotten cheaper. The unstated half of the pitch is the part I want to chew on, because it is that part that fails. The implicit promise was bidirectional growth: if anyone can ship an app, more apps will get used. If anyone can make a movie, more movies will get watched. The supply curve and the demand curve, the labs strongly implied, would move together. They didn’t because they couldn’t.

What didn’t grow

In a 1971 essay called “Designing Organizations for an Information-Rich World,” Herbert Simon wrote the line the next fifty years should have been built around: a wealth of information creates a poverty of attention. Attention is biological. The cap is the day and the day, as far as anyone has been able to verify, is still twenty-four hours long.

eMarketer’s 2025 time-with-media report puts the average US adult at more than eight hours per day with digital media alone, with Nielsen reporting an additional three-plus hours of daily audio listening on top of that. The shape inside the totals shifts year to year (less broadcast, more streaming, more TikTok) but the total volume of attention an average human is willing to commit has been approximately stable for years because it has to be. There is no version of a human in which the number doubles. There is no app on the way that creates more hours.

Apps tell the same story before AI got involved. Sensor Tower’s 2026 State of Mobile report finds mobile users spending roughly 3.6 hours per day in apps. Industry compilations put the average installed catalog north of eighty apps once you include the pre-installed ones, and the average user opens only about ten of those in a day. More than half of what is sitting on the average home screen does not get opened in a given month. Industry data on app retention is even bleaker: analytics firm Localytics found that two months after install only a third of installed apps are still being opened, and by three months that has fallen to 29 percent. People are not under-served by the existing app catalog. In reality, they are over-served, and have been for a decade. The marginal app is already worth nothing.

Now drop a logarithmic increase in artifact supply into that market. I wrote about this in 2011 in a different context: when something growing exponentially meets a fixed resource ceiling, the gap between “comfortable” and “consumed” is one doubling. Per-piece engagement collapses faster than the producers are equipped to notice, which we can measure directly. eMarketer found enthusiasm for AI-generated creator content collapsed from sixty percent in 2023 to twenty-six percent in 2025, while supply went vertical over the same window. Sprout Social’s Q3 2025 Pulse Survey of more than two thousand consumers found that fifty-two percent are now concerned about brands posting AI content without disclosure, tied with mishandling personal data as their top social-media worry. CivicScience’s May 2025 tracking found that thirty-six percent of US adults say AI in ads makes them less likely to purchase from the brand, up from thirty-two percent six months earlier. The eMarketer trend line is the one that matters most, because it sits in the exact category the labs are most aggressively expanding. It is not the demand curve flattening. It is the demand curve actively bending down in the market this essay is about.

The vanishing audience, in software

The strongest counterexample to my argument is a Dutch programmer named Pieter Levels, whose MMO flight simulator fly.pieter.com was vibe-coded in roughly thirty minutes and reportedly grosses more than $50,000 a month. This is real but worth dwelling on to understand why it is likely different.

Levels’ moat is not that he can vibe-code. His moat is fifteen years of audience-building on top of which the vibe-coding is a production-layer optimization. The audience came first. The artifact came second. This is exactly the sense in which 404 Media’s headline is doing more work than its writer probably intended: “This Game Created by AI ‘Vibe Coding’ Makes $50,000 a Month. Yours Probably Won’t.” It turns out that while the audience for new entrants collapses, vibe-coding is a real productivity gain for creators who already have audiences because the audience has always been the hard part.

Yours probably won’t and the reason it won’t is that customers were the actual scarcity all along and AI did not produce any. The vibe-coding tool platforms are real businesses generating real revenue, but it is revenue from makers paying for tools, not consumers paying for products. The picks-and-shovels boom does not tell us how much gold has been found. What it tells us is that the median product of vibe-coding is another todo app, in a market where the marginal todo app has been worth nothing for ten years.

The same dynamic is visible higher up the stack. I’ve written separately about how open source maintainers reacted to the AI-generated PR flood: curl killed a six-year bug bounty program, Ghostty banned AI contributions outright, tldraw started auto-closing external PRs. The maintainers were the audience for the new supply and they voted with their inboxes.

There is a defense of this whole genre, popularized by Maggie Appleton and Robin Sloan, that calls vibe-coded apps home-cooked software: personal, charming, not trying to be a restaurant. The analogy doesn’t survive the obvious test. Home cooks do not tweet “you should try my chili.” Vibe-coders compulsively do. The defining gesture of the genre is the announcement, not the consumption, and that gesture gives away the game. Their behavior is restaurant-shaped, and the restaurant has no customers.

The vanishing audience, in art

The art side of the argument is cleaner than the software side because art presupposes an audience in a way software doesn’t. A film with an audience of one usually means it failed; a tool with a user of one can mean it succeeded. Setting that asymmetry aside, the equation is the same. Supply is now effectively infinite and growing fast. Demand is still capped at the hours in a day. In the limit this drives the quotient to zero.

Sora 2 launched on September 30, 2025 with a feature called cameos: users opt in to make their likeness available, and anyone they have authorized can drop them into a generated scene. The feature went viral the way these features always do. Sam Altman appears in his own product’s promotional clips. Mark Cuban, Shaq, and a long list of streamers and influencers have made their faces into raw material. The streamer iShowSpeed found himself the subject of Sora deepfakes depicting him kissing a fan, racing a cheetah, appearing in a country he had not yet visited, and coming out as gay. Casual users have done well too: one creator posted a Sora video of squirrels jumping on a trampoline and watched it pull in seven million views and nearly four hundred thousand likes within a few days. It’s not clear to me who these are for. Of those seven million views I wonder how many thought the squirrels were real? How many knew this was AI and liked it anyway?

The argument for these tools is that they let anyone make the movie they want to watch. Make is doing a lot of work in that sentence. So is want. The implicit promise is that what you wanted from movies all along was a film custom-fit to your taste. I do not think that is true. I do not think the people typing prompts into Sora think it is true either. What we wanted from a film was to be moved by something other people were also moved by. The bespoke film makes that impossible. It is a TV dinner for one: sufficient calories, no shared meal.

There is a sharper version of this objection waiting inside the steelman and it is worth stating crisply. Suppose generation actually works as advertised in that anyone with a Sora subscription can produce, on demand, a beautiful film in their precise personal taste. Take the proponents at their word for one paragraph and follow the implication. Why would anyone watch yours? They can make their own. The better the model gets at custom generation, the more thoroughly it dissolves the audience for any particular custom-generated artifact, because the audience for custom artifacts is, by definition, the population that wants custom artifacts, and that is exactly the population the technology has just equipped to produce its own. The capability is anti-correlated with the demand for the things the capability makes. This is the same dynamic the software side runs into, restated in cinematic terms. The moment vibe-coding gets good enough that anyone can ship the app they want, no one needs yours. The moment generation gets good enough that anyone can render the film they want, no one needs yours. The premise of the technology is the abolition of the audience for the artifacts the technology produces.

This argument is sometimes confused with monoculture nostalgia and it isn’t. I am not pining for the night in 1983 when roughly 106 million Americans watched the M*A*S*H finale, a record that stood for twenty-seven years. The shared canon of mass television was constructed by gatekeepers and excluded plenty of people. But the function of shared cultural experience (the kind sociologists since Durkheim have talked about: collective effervescence, the way shared rituals build the substrate of social cohesion) was real, and is not replaced by everyone watching different things alone. People can taste the difference between art-made-for-them and art-made-for-no-one, and the data is the taste.

Even the supply gains were oversold

The supply/demand argument so far concedes that AI did, at minimum, deliver on the supply side. It is worth pressing on that concession though because even there the numbers are softer than the marketing.

In 2002 Joel Spolsky’s “Iceberg Secret, Revealed” argued that the visible part of software (the UI, the demo, the screenshot) is roughly ten percent of the work and that non-programmers cannot tell. Twenty-four years later, AI is built on his punchline. The vibe-coded app screenshot is a one-percent artifact priced as if it were the whole iceberg.

I run an engineering organization for a living and I use Claude Code daily, and have written separately about which doors AI made cheap to walk back through and which didn’t move. The short version, relevant here, is that the velocity gains are real and they are almost entirely consumed by other work. Productivity numbers across studies are mixed (METR’s July 2025 study found experienced open-source developers were actually slower with AI assistance, while other surveys find meaningful speedups), but the consistent finding is relocation, not net gain. SonarSource’s February 2026 State of Code Developer Survey gave the dynamic a name: the great toil shift. Stack Overflow’s 2025 Developer Survey measured the same thing in developer experience: trust in AI accuracy fell to twenty-nine percent, sixty-six percent of developers report spending more time fixing AI-generated code, and forty-five percent cite “almost right, but not quite” AI output as their primary frustration. AI removes some of the old developer toil and immediately creates new toil in its place: code that looks correct but is not, vulnerabilities to audit, debugging by inference rather than by craft, the verification work that any honest team does after the model declares it is done. We do not write code faster. We review more, audit more, debug more, document more, secure more. Productivity is conserved. It is just relocated.

To simplify any business to the software it builds/uses misunderstands the true bottom of the iceberg. At Yardstik, I am very proud of the software that we build but we would not have a viable business without our massive investments in compliance, understanding a constantly changing regulatory regime, real human customer support, a great sales team that can help customers navigate a complicated purchase decision, etc. More software doesn’t make these problems go away.

That gap, between AI getting most of the way there and AI actually delivering is not a small thing and is in fact the line where the iceberg starts. Below it, AI does not help, because what is required is taste, judgment, accumulated context, and the willingness to be on call when the thing breaks at 3 AM. None of that has been democratized. None of it is on any roadmap. Which means the supply-side gain, even on its own terms, is concentrated in the cheap and visible ten percent. The expensive ninety percent is unchanged.

Refeudalization, not democratization

When supply is infinite and demand is fixed, the only valuable position in the market is sitting on the demand bottleneck. Whoever rations attention to the audience captures the surplus that the artifacts themselves can no longer carry. This is the structural reason why the platforms (Lovable, Sora’s feed, YouTube’s recommender, the App Store) have become more powerful, not less, in the AI era.

Berkeley Haas has the right word for this. AI is not democratizing creative or technical labor. It is refeudalizing it. When generation becomes free, the scarce thing is no longer the artifact. It is the path between the artifact and an audience: discovery, taste, distribution, trust. None of those got cheaper, in fact they are likely becoming more concentrated as people look for help navigating the slopocalypse. The recommender systems that decide what gets seen are now the entire game. Their feedstock is unlimited; their incentives are not aligned with what is good for people and their power to ration attention is, by the day, more valuable. We are watching machines make videos for other machines to rank for an audience that has already left.

There is precedent. Chris Anderson’s The Long Tail in 2006 was the previous generation’s version of the democratization promise: the internet would empower niches, dethrone the hits, give every artist their audience. Anita Elberse did the empirical work in 2008, and the tail came back from the data longer but flatter, not fatter. The hits got bigger. The web democratized production while consumption stayed concentrated. What is different this time is the floor. The Long Tail still required human effort to fill, and that effort filtered for intent. AI completely removes that filter. We have seen this movie before and we are watching it again at higher resolution.

What was actually being asked for

Read the democratization promise back with the attention math in mind. Anyone can make a movie! What people wanted was a movie that moved them, ideally one others were moved by too. Anyone can ship an app! What people wanted was a working tool somebody else was on the hook for. The promise solved the cheap problem and made the expensive problems worse. We doubled the writers and held the readers constant. Then we doubled the writers again. The web-era slogan was if it’s free, you are the product. The AI version is sadder, because the product was always the audience, and there was never a way to manufacture more of it. What we lost is the substrate of shared experience, the part of culture that lets a stranger in an airport bar quote a movie and find someone to drink with. What we got is the privatization of substrate, rationed back to us by recommender systems the artifact-makers do not control.

Things I read while thinking about this