AI labs profit on tokens now, but fat margins hold only in the top tier — resolves 2027-07-20

The debate: model companies sell their service by the token, the unit models are billed in, and what used to be a money-losing business turned profitable in the first half of 2026, so the fight now is over whether those margins will last.

Why it matters: if the margins are structural, model companies have pricing power and buyers should plan to keep paying high prices, but if they are a temporary windfall from scarcity, competition will push prices down sooner or later.

Who is on each side: SemiAnalysis argues that scarcity plus a quality gap gives labs the power to price by value, Evans and Narayanan argue that selling tokens is ultimately a commodity business, and Gurley and Chamath argue that today's demand is padded by subsidies and has not yet faced a return-on-investment test.

SecondSource’s reading: whether model companies have pricing power has to be judged in three layers, not with a single yes or no.

The first layer is confirmed by multiple sources: the unit economics of selling tokens have turned positive, from losing money to making it.

The second layer is conditional: margins last only where three things overlap, flagship products for AI agents that complete whole tasks on their own, a compute shortage that has not eased, and buyers who have not yet tallied their return on investment.

The third layer is where the price war sits: it is already burning in consumer and low-cost tiers, while list prices for the most expensive flagships are moving up.

SemiAnalysis
“Anthropic's ARR has exploded from $9B to over $44B…”

SemiAnalysis
“AWS margins inflected this past quarter driven…”

AI Snake Oil / Normal Technology (Arvind Narayanan & Akash Kapur)
“If we assume a 5% net margin — much higher than…”

arXiv (Satoshi Matsuoka)
“a depreciation conveyor delivers newly amortized…”

Reviewed at 12 months, in July 2027: (1) a real cut to flagship list prices, meaning an OpenAI drastic cut that actually lands on the flagship tier, voids the durability claim; (2) Sonnet 5's promotional price failing to return to standard after August 31 is a weakening signal, checked early in a short window in September 2026; (3) the delivery gap closing together with rental rates breaking below their October 2025 low breaks the scarcity leg; (4) open weights matching on hard tasks under the same harness, with the token-volume difference converging, breaks the quality-gap leg; (5) an audited Anthropic gross margin in the S-1 far below the mid-60s reopens the turned-positive narrative; (6) annual token demand growth falling below 2x raises the weight of the commoditization price-collapse tail (the Matsuoka corridor, arXiv 2607.07207).

What would overturn it

The tests are listed above. If none of them trigger and flagship pricing keeps moving up, the durability call gains weight.

Correction to the settlement rule

Rewritten on 2026-08-03: the "what would overturn it" line used to say only "see the six checks in the receipt" and did not say how many have to trigger before the whole call falls. It now states the combining rule: any one of the six weakens or breaks its own leg, and the call is upgraded only if none triggers while flagship list prices keep rising. The call itself, its due date and the six checks are unchanged, and no result had been seen. The original wording stays in our ledger and in the public mirror's version history.