← Back to blog

Chegg Lost $1 Billion in a Day to ChatGPT, and Two Older Collapses Show Where the Margin Went

On May 2, 2023, Chegg's stock fell 48.41% in a single day. The music industry and newspapers ran the same collapse first. In all three, the money didn't disappear. It moved one layer up, to a bottleneck the loser didn't own.

Published July 2026 · 8 min read

On the afternoon of May 2, 2023, Chegg's stock fell 48.41 percent. One trading day, roughly one billion dollars of market capitalization, gone. The homework-help company had just become famous for the worst possible reason: it had just done what almost no public company ever does on an earnings call: admit that an AI product was eating it.

The admission was almost polite. CEO Dan Rosensweig told analysts that "since March we saw a significant spike in student interest in ChatGPT," and that the company now believed it was hurting new-customer growth. Chegg guided the next quarter's revenue to $175 to $178 million against an analyst consensus of $193.6 million. That gap, maybe eighteen million dollars of guidance, vaporized a billion in equity value before dinner. Markets do not price the quarter; they price the trend, and the trend was legible to anyone with a browser: the thing Chegg sold for a subscription, a worked answer to a homework problem, had started coming out of a free chat window.

It is tempting to file this as an AI story. It is really an economics story, and an old one. The same collapse has run at least twice before, with different products and different villains, and both prior runs ended the same way: the money did not disappear. It moved. Specifically, it moved up, to a layer of the stack the loser did not control. If you sell anything whose value lives in a copy, the three collapses read less like history and more like a weather forecast.

The law in one sentence

Here is the pattern, stated plainly so you can test it against each case: when the marginal cost of a copy collapses toward zero, the margin does not vanish; it re-forms one layer up, at the new bottleneck, owned by someone else.

The copy can be a song, an article, or a homework answer. The bottleneck can be a platform, an ad-targeting duopoly, or a foundation model. The constant is the direction of travel. Margin behaves like heat: it flows from where copies became free to wherever scarcity still lives.

Case one: the song

In 1999, US recorded-music revenue peaked at $14.6 billion. That number rested on a physical fact: a song was an object. You wanted the song, you bought the disc, and the disc had a marginal cost, a price, and a retail channel that the industry controlled end to end.

Napster and the MP3 turned the song into a file, and the file's marginal cost was zero. By 2014 and 2015, US recorded-music revenue had fallen to about $6.7 billion, less than half the peak. The industry spent that decade suing file-sharers and shutting down services, which worked about as well as suing gravity. The copy was free now. No court order could reprice it.

Then something instructive happened: revenue came back, but the industry did not. Streaming grew from 7 percent of the US market in 2010 to 83 percent in 2020, and total revenue recovered to $12.2 billion, still roughly 46 percent below the 1999 peak in nominal terms. Look at where that rebuilt money pools. Rights holders earn something like $0.003 to $0.005 per stream. The aggregation layer, first Apple with iTunes and the iPod, then Spotify, sits above every label and every artist, owns the customer relationship, and captures the margin that used to live in the disc. The song survived. The label's chokehold did not. The scarce thing stopped being the copy and became the shelf every listener walks past, and Spotify owns the shelf.

Case two: the article

Newspapers ran the same experiment with higher stakes, because their real product was never news. It was attention, sold to advertisers. In 2005, US newspaper advertising revenue peaked at $49.4 billion. A local paper was a bundle of scarcities: the only classifieds in town, the only display ads reaching that zip code, the only printing press that mattered for local reach.

The web made the article a free copy and, worse, unbundled the scarcities. By 2022, US newspaper advertising revenue had fallen to $9.8 billion, a drop of more than 80 percent from the peak. The often-told hope was that digital advertising would backfill print losses. It never came close: online newspaper ad revenue grew from zero to only about $3.1 billion, a fraction of what print surrendered.

So where did forty billion dollars of annual advertising go? Up a layer. Google and Facebook built the new bottleneck, targeting. An advertiser who once needed the Cleveland paper to reach Cleveland readers could now reach exactly those readers, individually profiled, without the paper's involvement. The two platforms together now take in on the order of $200 billion a year in advertising. The article became free. The knowledge of who is reading, and the machinery to route an ad to them, became the scarce asset. Newspapers owned neither.

Case three: the answer

Which brings us back to Chegg, the compressed, high-speed rerun. Chegg's subscription product was, at bottom, a library of worked answers and the service of producing new ones. A homework answer is a knowledge good: expensive to produce once, nearly free to copy. Chegg's moat was that producing a new answer to a new question still cost human effort, so students paid $14.95 a month for the archive plus the production line.

ChatGPT collapsed exactly that. A large language model is a machine for producing plausible worked answers on demand, at a marginal cost near zero, to questions that have never been asked before. The moat was not the archive; it was the cost of the next answer, and that cost went to approximately nothing in a single product cycle. Students noticed by March. Rosensweig said so in May. The market repriced Chegg in an afternoon.

The most telling detail is not the crash. It is Chegg's response. The company announced CheggMate, its own AI study assistant, built in collaboration with OpenAI. Read that with the pattern in mind: the incumbent whose margin had just fled to the model layer could answer only by renting capability from the model layer. That is the same move as a record label distributing through Spotify or a newspaper buying traffic from Google. It may be the right move; it is also a confession of position. You rent the chokepoint you failed to own, on terms set by the layer above you.

Same law, faster clock

Set the three cases side by side and the structure is identical. The copy that went free: a song, an article, an answer. The loser: whoever ran the copy-and-sell business. The winner: whoever owned the layer the margin fled to, Spotify at the platform layer, Google and Facebook at the targeting layer, OpenAI at the model layer.

Two things changed between 1999 and 2023, and both should concern anyone whose revenue depends on a knowledge good.

First, the clock. Music took roughly a decade to halve. Newspapers took about fifteen years to lose 80 percent of their ad base. Chegg lost half its market value in one day, on the strength of two months of user-behavior data. Markets have learned the pattern, which means the repricing now happens at the speed of recognition rather than the speed of the underlying decline. You do not get a decade of denial anymore. You get an earnings call. (We pulled that acceleration apart on its own in The Commoditization Clock.)

Second, the breadth. Napster freed one kind of copy. The web freed a few more. A general-purpose model collapses the copy cost of many knowledge goods simultaneously: answers, boilerplate code, product descriptions, stock illustration, translation, first-draft legal prose, tutoring scripts. Every one of those is somebody's subscription business, and each of those businesses now faces the question Chegg answered in public. Categories that assumed they were different because their copy was specialized are discovering that "specialized" was a statement about production cost, and production cost is exactly what collapsed.

To keep the ledger honest: none of the three stories is theft in the shape that morality-tale versions suggest. Streaming genuinely rebuilt music revenue and put every song ever recorded in your pocket for ten dollars a month. Google and Facebook built targeting that advertisers demonstrably prefer. ChatGPT gives a student at midnight something no $14.95 subscription offered. Consumers won every round. The law is not "technology steals." The law is "value re-forms at the new scarcity," and it is indifferent to whether the old scarcity holder was virtuous.

The question to ask about your own product

The transferable lesson is a single question, asked before the earnings call instead of on it: if the marginal cost of your copy goes to zero, which layer does the margin flee to, and are you on it or under it?

Notice what the question is not. It is not "will AI disrupt my industry," which invites a comfortable maybe. Assume the copy cost collapses; for most knowledge goods that is now the base case, not the tail risk. The strategic content is entirely in the second half. Identify the layer above you: the aggregator, the distribution chokepoint, the model, the compute underneath the model. Then be honest about your relationship to it.

The three cases suggest only a few durable postures. Own part of the layer above, which is rare and expensive. Be genuinely indispensable to it, the way a marquee catalog still matters to Spotify. Or hold something the model layer cannot generate: a live relationship, a proprietary data feed, a certification that carries legal weight, a physical footprint. What does not survive contact, on the evidence of three industries and forty years, is the plan of selling a better copy. Chegg had better answers than ChatGPT for quite a while, by accuracy. It did not matter, because the fight was never about the quality of the copy. It was about who owns the bottleneck once the copy is free.

The margin always goes somewhere. The only choice you get is whether you figure out where before the market does.

← Back to blog