Concept

Model Commoditisation

conceptaieconomics-of-aistrategymoats

Model Commoditisation

Benedict Evans‘s thesis on where value accrues in the AI stack: foundation models are heading toward being an undifferentiated, low-margin commodity utility, with the durable value sitting above them — in applications and distribution. The shape to expect is cloud, not Windows.

The argument

  1. No network effects. Unlike a social network or an operating system, a frontier model has no winner-take-all dynamic — nothing makes one run away from the others. So competition persists indefinitely (three to six labs, not one).
  2. Persistent competition → no pricing power. With several broadly-equivalent models and no radical product differentiation that a user can perceive, there is no basis for sustained pricing power. “Have you got like three or four companies all selling the same thing?”
  3. Value moves up the stack. If the chatbot isn’t the UX and it has to be “apps” the labs won’t all build themselves, then — as with AWS — the model is undifferentiated infrastructure and the leverage sits with whoever builds on top and owns distribution. (Contrast Windows/iOS, where platform lock-in let the layer below capture value.)

Evans’s tell is Sam Altman’s line about “selling intelligence on a meter like water or electricity”: utilities are the canonical low-margin commodity — when you watch television the broadcaster pays the power company nothing as a cut. His telecoms career is the cautionary case: global mobile data is ~1,500–2,000× its 2010 volume, the industry spends ~15–20% of revenue on capex, and the stocks went nowhere for 25 years because “all the cool stuff is made by you.”

Disequilibrium vs steady state

A key distinction: today’s economics are a radical price disequilibrium (the OpenClaw user burning $1.5m of tokens in a month is like someone running up a $50k mobile-data bill in 2010) — not the steady state. The real question is what margins look like once the lines on the chart settle. Evans’s bet: compressed margins for the model layer, value up-stack. Its corollary is Task vs Job (the human judgment AI doesn’t capture) and distribution as the moat when the product is a commodity.

Where mainstream views differ

This is a genuinely contested forecast, made by an analyst rather than an operator.

  • The “labs capture everything” view. Many close to the labs assume today’s situation persists — that foundation-model providers will hold strong pricing power and accrue most of the value. Evans thinks this just presumes the present continues.
  • Oligopoly with differentiation (the strongest counter). Dario Amodei, in Dario Amodei on the End of the Exponential, AI Diffusion, and the Economics of AGI, agrees frontier AI is a small-numbers market (a Cournot oligopoly, like cloud — 3–4 players, very high entry costs) but argues models are more differentiated than cloud (Claude, GPT, Gemini differ in subtle, stylistic ways), so margins won’t collapse to zero, and the API business is durable. The cruxes between them: how perceivable model differentiation is to ordinary users, and whether the chatbot or third-party apps become the dominant surface.
  • Evans’s own caveat. He stresses humility — having this conversation about the internet in 1997 you’d have got almost everything wrong — so he holds the thesis as the base-rate default, not a certainty.

See also