The Next Fifteen Years

A forecast built from first principles
Section future / 02-games / 3-firms.md

Game 3 - Firms: the Red Queen, and why adoption doesn't mean profit#


Contents

The most misunderstood dynamic in the entire subject.

Consider a competitive industry where all firms can adopt AI at cost c and reduce marginal cost by Δ. In Bertrand-ish competition, prices fall to the new marginal cost. Every firm must adopt to survive. Nobody captures the surplus - consumers do.

Why this resolves the current argument#

This is why "AI will make companies hugely profitable" is mostly wrong.

And it is why "AI is a bubble because firms report no earnings lift" is also wrong.

The gains are real. They land in consumer surplus and in whoever holds the genuinely scarce input. The absence of a large earnings lift at the median adopter is not evidence of a bubble - it is exactly what competitive markets predict when a general-purpose input gets cheaper. Expecting otherwise is expecting firms to capture rents in markets where entry is not blocked.

The corollary is uncomfortable for adopters: the investment is mandatory and the return is defensive. You spend to not die, not to win. That is the Red Queen, and it is a perfectly normal outcome in the history of general-purpose technologies - electricity, containerization, ERP, the internet all ran the same way.

So where does value actually accrue?#

By elimination - to whatever AI does not make abundant:

This is Baumol's cost disease inverted and weaponized. As cognition approaches free, all value concentrates in the inelastic complements.

If one sentence carries through the next decade, that's it.

Where the Bertrand result does not apply#

The scope condition does real work, so state it precisely. Surplus escapes to producers wherever entry is blocked or price competition is muted:

The two-economy split in Prices is this section restated as macro: the competitive tradables deflate, the licensed and land-linked sectors keep their prices, and measured inflation becomes a ratio between the two.

The transient rent, and who eats the adoption cost#

Bertrand is the end state, not the path. During diffusion, an early adopter holds a real cost advantage with a half-life set by rivals' adoption lag - quarters to a few years depending on sector reorganization friction (Game 4 covers why the friction is organizational). So the honest pitch to an adopter is not "no gains" but "a temporary spread that decays to zero, which you must nonetheless buy." Two second-order effects follow:

  1. Over-adoption is its own dissipation. Because the investment is defensive and the counterfactual is unobservable, boards mandate adoption under competitive fear rather than measured return. Spending above the optimizing level - consultant-led programs, shelfware seats, tool churn - is the firm-level analogue of Game 1's over-dissipation, and it is why survey-measured "AI spend" will persistently exceed anything visible in productivity statistics.
  2. The adoption cost is recurring, not one-time. Model generations turn over in months; integrations, evals, and retraining recur with them. The c in the model is closer to a subscription than a capex, which lowers the equilibrium net benefit to adopters further than the static story suggests.

How to score the Red Queen without cheating#

Three observables, not one. Adoption (surveyed use, seat counts) can rise while margins (B3) stay flat and prices (B7 cognitive basket) fall - that is the base case and it is not a bubble. Adoption rising with margins expanding in competitive sectors is the Red Queen broken. Adoption rising with margins expanding only in concentrated/licensed sectors is the scope condition working as written - Game 3 still holds where it claimed to hold. The steelman §2 objection (markets aren't competitive) is scored by this cross-section, not by the average: uniform margin expansion everywhere falsifies; dispersion confirms the split.

The consumer-surplus claim is harder to observe directly. The proxy is relative prices: if competitive cognitive services deflate while complements inflate, surplus is leaving the adopter layer somewhere. Whether it lands with consumers or with complement-owners is the assets / Uncertainty 7 question, not a Game 3 failure.

Failure modes#

SaaS is the worked example. Seat-priced software with no data/distribution/compliance moat is Red Queen pure: adopt or die, margins flat, prices soft. Outcome pricing (B5) is the escape hatch when reliability is real - not a different game, a different pricing of the same one. → Software


Related: Inference economics on the two-year moat · Software for the SaaS case · Startups for Game 3 applied to firm birth · Prices · Compressed version

View markdown source

select · Enter open · Esc close