Where the money for answers went

In 2022 Chegg took in $766.9 million, most of it from students paying a subscription for help with their coursework, and 8.2 million people subscribed during the year.¹ What they bought was answers, backed by what the company calls "a vast network of subject matter experts".² By 2025 the revenue was $376.9 million, down 39% on the year before.³ In October 2025 it cut 388 jobs, about 45% of its staff, after revenue losses it blames in part on the AI summaries Google now shows above its search results.⁴

About $390 million a year that once reached Chegg no longer does. Not all of that need be AI's doing. But among the risks to its business the company itself lists "student adoption of generative AI products" and "the speed, scale and potential impact of Google's AIOs", its AI Overviews.³ Its complaint against Google says Chegg depends on Google's referrals "for a large portion of the revenue that it devotes to producing original online content".⁴ So some of the money now goes to chatbots, some of them paid, and some to the search page that answers before anyone clicks. And some of it goes nowhere at all. A student who gets the answer free simply keeps the subscription fee.

That is the general shape. When a model does a piece of intellectual work for almost nothing, the money that paid for the work does not move to the model's maker. Most of it stays with the person who used to pay. Of what is still spent, the surest profit goes to whoever sells the scarce thing underneath, which today is chips.

The reason is ordinary price theory. A homework answer was expensive while it took an expert's time. Once a machine can write one in seconds, and several companies sell such machines, competition pushes the price of an answer toward the cost of computing it. The money leaves the stage of the chain that became abundant and settles on the stages that did not: the hardware that does the computing, and the doors through which people ask.

The sellers of the work feel it first. On a large online freelancing platform, after ChatGPT's release, freelancers in writing-related work saw their monthly jobs fall 2% and their monthly earnings 5.2% relative to freelancers in less affected occupations, and the most established freelancers lost more.⁵ The study covers only the first months after the release.⁵ Five percent is small, but it points the way the mechanism predicts.

The obvious candidates to collect the money are the companies that build the models. Their prices say otherwise. Epoch AI tracked what it costs to reach a fixed level of performance on six benchmarks and found the price falling between 9 and 900 times a year, with a median of 50.⁶ A business whose product gets fifty times cheaper every year is handing almost all of its gains to its customers. OpenAI reportedly took in about $4.3 billion in the first half of 2025 and burned $2.5 billion in cash.⁷

The stage that keeps its margin is the hardware. In the year to January 2026, Nvidia's revenue was $215.9 billion, $193.7 billion of it from data centres, and its net income was $120.1 billion.⁸ That is the old pattern of a gold rush, in which the reliable money is made selling equipment to the prospectors. The margin lasts as long as the chips stay scarce.

The largest share is the hardest to see, because nobody books it as revenue. Erik Brynjolfsson and colleagues asked representative samples of American adults how much they would need to be paid to give up AI chatbots for a month.⁹ In March 2026 the mean answer was $124.50 and the median $11.40, which implies a value to American users of about $172 billion a year, against an estimated $14.2 billion that consumers worldwide paid OpenAI, Anthropic, Google and Microsoft for such tools in 2025, a figure that leaves out what businesses pay.⁹ The authors note that what people would accept to give something up differs from what they would pay, and that William Nordhaus found the same lopsided split, with producers keeping a small share of the value, for the major technological advances of the late twentieth century.⁹ A minority of heavy users carries the mean, so the true figure may well be lower; even a fraction of it is several times what consumers pay.

So the question of who captures the value of intellectual work after AI has an awkward answer: mostly the people who used to pay for it. The clear losers are the sellers of the work, the companies and the experts and freelancers inside them. The clearest winner on paper is the chipmaker. The model makers are, for now, a channel the money passes through on its way back to the customer.

It may not stay there. Prices cannot keep falling fifty times a year, and a door that every question passes through can start charging rent. Chegg's quarrel with Google is an early argument about exactly that door.