Subversive, est. 2013Talk to Dave
DEVON, AND WHEREVER YOU AREFRIDAY 6 NOVEMBER 2026
← Uncommon Sense

Field Notes

AI Is Doing to Software What Napster Did to Music

The record industry lost its prices and recovered on different terms. Anyone renewing a software contract this year can learn from how it did so.

In October 2007 Madonna left Warner Music, her label for twenty-five years, for Live Nation, a concert promoter. The ten-year agreement, reported at $120m, gave Live Nation a share of her albums, tours, merchandise and films. A concert business had decided that the records were now the cheapest thing she had to sell.

It had reason to. Napster arrived in the summer of 1999, the record industry's best year by revenue, and by 2014 that revenue had roughly halved. It took more than two decades to regain its 1999 total in nominal terms, and in real terms it remains well below it. Live Nation reported revenue of $25.2bn for 2025, against $31.7bn for the whole of recorded music. The comparison is imperfect, since much of a promoter's revenue passes straight to performers and Live Nation's adjusted operating income was $2.4bn, but over a quarter of a century spending has moved towards live performance.

Businesses that sell services online are now going through something similar. Chegg, which sold homework help to students, cut 45 per cent of its staff in October 2025, citing AI and falling traffic from Google. Stack Overflow, where programmers have gone for answers since 2008, was on course by May 2025 for fewer questions a month than in its third month of operation, although its decline began several years before ChatGPT. In February nearly $1tn came off software and services shares in a week, on fears that AI agents would reduce demand for paid licences. That reflects what investors expect, and what customers do will become clear as their contracts come up for renewal.

The music industry recovered in two ways, and both are being tried again. The labels licensed their catalogues to the new distributors, first Apple and then Spotify. Reddit has done likewise with Google, reportedly for $60m a year, and in 2024 Stack Overflow signed data agreements with Google and OpenAI, selling access to answers its users had contributed free of charge. Licensing pays only where the catalogue cannot be obtained elsewhere. The alternative was to sell something that could not be copied at all, which in music meant live performance. Chegg was poorly placed for either.

Uncommon sense (for busy people)

  • When copying costs nothing, the money moves to what cannot be copied.
  • A catalogue is worth licensing only if it cannot be had elsewhere.
  • Chegg sold answers, and answers are now the cheapest thing on offer.
  • In a service business, the scarce asset is expertise someone will put their name to.
  • Contract renewal is the buyer's best chance to test the price.

The analogy fails in one respect. Napster copied a finished recording, and listeners still wanted the artist, so the labels lost the physical format and the retail margin but kept the music. A model reproduces the service itself, and a student whose homework is finished has no further use for Chegg. Stack Overflow's archive helped train its rivals years before the company charged for it, and the courts have yet to settle how much of that scraping was lawful.

I suspect this transition will be quicker, though the evidence for that is thin. Listeners abandoned CDs gradually, over a decade. Companies review software contracts annually, and a licence count can be cut at the first renewal after agents start doing the tasks those licences paid for.

What a service firm can still charge for is expertise that a named person stands behind: someone who knows the client's business, approves what is delivered and is answerable when it goes wrong. A platform selling the reproducible part of a service has no automatic claim on that person, any more than Warner had a claim on Madonna's touring income.

Buyers can put this to use at renewal. Ask the supplier which items on the invoice pay for identifiable people and which pay for output a model could now produce, and at what cost. Then negotiate the second category down and leave the first alone.

If you would like a second opinion on which parts of your software bill a model could now supply, you'll find me at subversive.ventures.

Further Reading