# The cost of copying was the whole story

Slow Variables · Edition 001 · 14 Sep 2026 · status: revised
For thirty years, one number explained the Internet. It is about to stop being the important one.

There is a number that explains almost everything strange about the last thirty years of the Internet, and it is not a number most people can name. It is the marginal cost of a copy. In 1990 a copy of a song cost a record company perhaps a dollar to press and ship. By 2000 it cost nothing, and it cost nothing for everyone, not just the record company. Every argument about piracy, paywalls, advertising, platforms, streaming, open source, the collapse of newspapers and the rise of the attention economy is, underneath, an argument about what happens to institutions built for a world where copies cost a dollar when copies suddenly cost nothing.

This is a slow variable in the strict sense. It moved once, over about a decade, and then it stayed moved. Everything fast that happened on top of it — the startups, the lawsuits, the business models, the moral panics — was downstream of that one shift. If you understood that copying had become free and would stay free, you could have predicted the broad shape of 2000 to 2020 with embarrassing accuracy. Most people who lived through it did not, because the number was invisible. Nobody publishes the marginal cost of a copy. It is not on a dashboard. It is the water.

## What the free copy actually did

The standard telling is that free copying destroyed the value of information. That is wrong, and the way in which it is wrong matters for what comes next.

Free copying did not destroy value. It relocated it. When copies cost a dollar, value sat in the copy, and the institutions that controlled copies — publishers, labels, studios, broadcasters — captured it. When copies became free, the value drained out of the copy and pooled in the two places copying could not reach: the original act of creation, and the point of attention. Shapiro and Varian saw the second half of this clearly in 1998.[1] Price to value, not cost; compete on lock-in; own the relationship, not the artefact. That is the entire playbook of the platform era, written before the platform era began.

What they did not see, because nobody did, is that the first half — the act of creation — would remain expensive for a full generation after copying became free. That asymmetry is what made the period legible. Producing a newspaper still cost a newsroom. Producing a song still cost a musician and a studio. Producing software still cost engineers. Copying any of them cost nothing. So the whole economy reorganised around a single question: how do you get paid for something expensive to make and free to copy? Advertising was one answer. Subscriptions were another. Open source, with its strange economy of reputation and adjacent services, was a third. All of them are answers to the same question, and the question only exists because production stayed dear while copying went cheap.

The value did not vanish. It pooled where copying could not reach.

## The second number

Here is the argument of this edition. The marginal cost of a copy stopped being the important number sometime in the last three years, and the number that replaced it is the marginal cost of production. And that number is now doing what the copying cost did in the 1990s: falling by an order of magnitude every few years, for everyone, with no floor in sight.

This is not a claim about whether machine-generated text or images or code are good. Set quality aside entirely. The claim is about cost structure. For the first time since the printing press, the cost of producing a plausible artefact — an article, a song, a function, a legal brief — is converging on the cost of copying one. When two costs that used to differ by a factor of a million converge, every institution that was built to arbitrage the gap loses its reason to exist. That is what happened to record labels in the 2000s.[2] It is what is starting to happen now to a much larger set of institutions, most of which do not yet think of themselves as being in the copying business.

Consider what the platform playbook actually assumed. Own attention, because production is scarce and attention is the bottleneck. But attention was only the bottleneck because production was scarce. If production is abundant, attention is not a bottleneck; it is a landfill. The feeds are already showing this. The value of a distribution channel falls as the cost of filling it approaches zero, because filling it no longer signals anything. Nobody has priced this in, because it is a slow variable and slow variables do not show up in quarterly numbers until they have already done their work.

## Three places to watch

If the argument is right, the next decade's fights will be legible in advance the same way the last two were. Three places to look.

Provenance becomes the scarce thing. When copies were the unit, the scarce thing was the original, and copyright existed to protect it. When production is free, the scarce thing is the warrant — the credible claim that a particular human, with a particular stake, made this and stands behind it. Expect the law to spend a decade trying to bolt provenance onto copyright, the way it spent the 2000s trying to bolt DRM onto copies, and expect it to fail for the same reason: the abstraction no longer matches the physics.

Archives get repriced. For thirty years the past was free. Newspaper morgues, forums, mailing lists, the whole sediment of the open web — all of it copyable, all of it worth roughly nothing per unit. Training changed that overnight. A corpus of verified human writing from before 2022 is now a non-renewable resource — low-background steel for language[3] — and the institutions that hold such corpora are only beginning to understand what they are sitting on. Watch archive access become a priced good, and watch that reverse the open-web norm faster than anyone expects.

Institutions built on the production gap hollow out from the middle. The 2000s destroyed the middle of the recording industry — the mid-list act, the regional label — while the top and the bottom survived. The same shape is coming for every field where a credential certified the ability to produce a plausible artefact. The stars survive because provenance attaches to them. The amateurs survive because they were never paid. The professional middle, whose entire value was the production gap, does not.

## What this edition is not saying

It is not saying the machines are good at the work. That is a fast variable, it changes monthly, and arguments about it are mostly people talking past each other with different examples. It is not saying jobs are ending; the record industry is larger today than it was in 1999. It is saying that a specific number, which quietly governed everything for a generation, has been replaced by a different number, and that most of the institutions and arguments we inherited were built on the old one. The fights of the next decade will make more sense if you keep the new number in view.

The claims that follow from this are on the ledger below, with confidence attached. Come back in 2028 and see how it went.

[1] Carl Shapiro and Hal Varian, Information Rules: A Strategic Guide to the Network Economy (Harvard Business School Press, 1998). Varian went on to become Google's chief economist, which is either a vindication of the book or a footnote to it. [2] Worth being precise: recorded-music revenue fell by roughly half in real terms between 1999 and 2014, then recovered on streaming. The industry survived. The structure did not — the mid-list and the regional label were the casualties. [3] Steel smelted before 1945 carries no fallout from atmospheric nuclear testing and is prized for sensitive instruments; it is salvaged from pre-war shipwrecks. The analogy to pre-2022 text is imperfect but hard to shake.


## Fast variables

* A major label licenses its catalogue to a model trainer (Music industry)
The interesting part is not the deal but the shape of it: a flat fee, not a royalty. Royalties are a per-copy instrument. Flat fees are what you charge when you have stopped believing copies are the unit.

* A frontier lab publishes per-token prices that undercut last year's by an order of magnitude (AI pricing)
Cost of production falling at roughly the rate cost of copying fell in the 1990s. Watch what happened to the institutions that assumed the old number was stable.

* A newspaper's archive goes behind a scraping wall, then reopens for a fee (Publishing)
Access to the past is becoming a priced good again. That is a reversal of the last three decades, and almost nobody is treating it as one.

* A court declines to rule on whether model weights are a 'copy' (Law)
The legal system's core abstraction is the copy. When the courts cannot say whether something is one, the abstraction has stopped describing the world. The law is going to be slow to notice.

## Long memory

* 1945 — Vannevar Bush's As We May Think imagines the memex: a desk that stores everything you have read and lets you trail from one document to the next. Bush assumed the bottleneck was retrieval. He never considered that copies would become free, because in 1945 they very much were not.

* 1984 — Stewart Brand at the first Hackers Conference: information wants to be free and information wants to be expensive, and those two fight each other. The first half became a slogan. The second half was the more important prediction, and it took forty years to come due.

* 1998 — Shapiro and Varian's Information Rules formalises the economics: high fixed cost, near-zero marginal cost, so price to value rather than cost, and compete on lock-in. It was the most accurate business book of its era, and every one of its strategies presumes that producing information stays expensive.

* 1999 — Napster. The moment the copying cost hit zero for ordinary people, and the recording industry spent a decade litigating against arithmetic. The lesson everyone drew was about piracy. The actual lesson was about which variable had moved.

## Ledger

* [55%, by 2028-12-31, open] By the end of 2028 at least one of the five largest English-language news publishers will charge model trainers more per year for archive access than it earns from print subscriptions.

* [40%, by 2027-06-30, open] By mid-2027 a major jurisdiction will have created a legal category for trained model weights that is explicitly neither a 'copy' nor a 'derivative work'.

* [70%, by 2030-12-31, open] Per-copy royalty income for recorded music will be lower in real terms in 2030 than in 2025, despite higher total industry revenue.

## Colophon
Written by a person, under a pseudonym.
=> https://slowvariables.net/colophon/ Why
=> https://slowvariables.net/issues/001/ Web version
=> mailto:editors@slowvariables.net Corrections