The Multiplication

Why do two companies renting the same model, at the same price, on the same day, get different results from it? That is the only question this article answers. The answer is a formula, and everything else I publish in this series is a footnote to it.

Realized capability = Capacity × Knowledge quality × Relevance precision.

Definitions first, then the arithmetic, then what the arithmetic does to a budget.

1. The three factors. Capacity is what the labs sell: the model’s ability to reason, write, and act, priced per token. Every business can rent the frontier version of it on the same day for the same money. I made the case in the layers article that this makes capacity a commodity in the strict sense, a good on which no buyer can differentiate.

Knowledge quality is the condition of what the company has decided is true: how much of it exists as claims rather than as documents, how much of it is current, how much of it contradicts itself. It is produced inside the company. Nobody sells it.

Relevance precision is the share of delivered knowledge that is right for the task, right for its scope, on time, and still in force, against the share that is noise. I redefined the word in the relevance article. It is enforced by the layer between the store and the agent. Nobody sells it either, though several vendors sell the first of its four factors.

2. The operator is the argument. If the formula were additive, a weak factor could be compensated by a strong one. A company with a poor knowledge base could buy its way out by renting a better model. Under multiplication there is no such compensation. A factor near zero pulls the product toward zero regardless of the other two, and raising the other two does not restore it.

3. Illustrative arithmetic, marked as such. These are not measurements. They exist to show the shape. Score each factor from 0 to 1.

A company rents the frontier model: capacity 0.9, the same as its competitor. Its knowledge base is half current and partly self-contradictory: 0.5. Its retrieval is similarity-only and returns superseded claims about as often as live ones: 0.4. Product: 0.18.

The competitor has the same 0.9, a governed canon at 0.8, and status-aware delivery at 0.8. Product: 0.58. Same vendor invoice. Roughly three times the realized capability.

Now give the first company the theoretical maximum capacity, 1.0. Product: 0.20.

4. The collapse property. The point of 0.20 is not that the upgrade did little. It is that the failure mode under multiplication is not degradation. A weak knowledge factor does not make the agent slower or duller. It makes a fully capable agent execute a wrong belief at full capacity, with full confidence, and the higher the capacity, the more finished the wrong work looks. The relevance article contains one instance: a login flow written correctly against a vendor the company had dropped seven months earlier. The model did nothing wrong. The product was near zero because one factor was.

5. Where competition lives. Since every competitor can buy the same first factor, the first factor cancels out of any comparison between two businesses. What remains is the ratio of their knowledge quality and relevance precision. That ratio is the whole of the difference in what two companies get from AI. It is set inside each company, by work no lab does and no lab can do, because the inputs, what this company decided and what applies to this task, never leave the company.

6. The objection, as it is usually made. “Capacity is still improving fast. Next year’s model will read our documents well enough to work out what is current on its own.”

The first sentence is true, and I want it to be true. Every jump in capacity raises the product for everyone, which is why nothing in this series argues against the labs. The second sentence confuses reading with deciding. A model reading two contradictory decision records can tell you they contradict. It cannot tell you which one the company holds now, because that fact was never in either document. It was in the act of a person choosing, the distinction I drew in the memory article. Better capacity multiplies whatever it is given. It does not manufacture the second and third factors out of the first.

7. Falsifiability. The formula is wrong if a company with stale, contradictory knowledge and similarity-only delivery can be shown to close the gap with a governed competitor by renting a stronger model alone. That would mean the factors add. I have not seen it. The pattern I see is the reverse: the stronger the model, the more expensive the wrong belief.

8. What this does to a budget. Most AI spend today buys the factor on which differentiation is impossible: tokens and seats. The two factors that decide the product are usually unfunded, or funded as a tooling line rather than as work. The formula says the allocation is inverted. For any company already renting the frontier, the cheapest way to raise realized capability is not a better model. It is to move one of the other two factors from 0.4 to 0.8, which doubles the product, while the labs keep raising the first factor for you at no marginal cost.