Tokenized Cognition Model (TCM)

The Tokenized Cognition Model (TCM) is a framework for measuring the economic value of AI cognition by treating each token of model inference as a unit of priced reasoning. It makes compute consumption, margin structure, and product viability legible at the unit-economics layer rather than the application layer.

The model exists because AI economics do not work the way SaaS economics work. SaaS revenue is largely decoupled from per-user compute cost; a marginal user is nearly free to serve. AI revenue is coupled to per-interaction compute cost; every query, agent step, and retrieval pass consumes tokens that have a direct cost basis. The same product can show strong top-line growth and structurally negative unit economics simultaneously, and the difference is invisible until cognition is priced explicitly.

TCM resolves this by making three quantities measurable:

The first is cognition throughput — the number of tokens an entity processes to deliver a unit of value to the user. A search query may cost a few hundred tokens. An agent completing a multi-step task may cost tens of thousands. A long-context reasoning workflow may cost hundreds of thousands. Throughput is the consumption side of the equation.

The second is cognition pricing — the revenue an entity captures per unit of value delivered. Some pricing models capture this directly (per-call APIs, per-token billing). Most capture it indirectly through subscription, advertising, or transaction fees, and require imputation to make comparable.

The third is the cognition-compute spread — the difference between cognition pricing and the underlying compute cost basis. This spread is the actual unit economics of the AI product. Positive spreads compound; negative spreads burn capital regardless of revenue growth.

Three derived metrics extend the framework:

Agent ARPU measures revenue per agent-served interaction rather than per human user, recognizing that agentic workflows displace human sessions rather than adding to them. As agentic deployment scales, traditional ARPU loses meaning and Agent ARPU becomes the operative number.

Valuation-implied cognition load works backward from market capitalization to the compute throughput an entity must sustain to justify its multiple. When implied throughput exceeds available compute supply, the valuation is dependent on infrastructure that does not yet exist.

Compute Yield Index measures how much economic output an entity produces per unit of compute consumed. Entities with rising compute yield are extracting more value per token; entities with falling compute yield are subsidizing usage and require capital injection to continue.

TCM sits inside the Four Forces of AI Power as the unit-economics framework underneath the Compute pillar. The Four Forces describe what is contested at the structural layer; TCM describes what the contest costs at the per-transaction layer. Entities can win the structural contest and still operate at negative cognition margins; entities with positive cognition margins can compound even from a smaller compute footprint.

The practical use of TCM is to evaluate AI businesses on the same unit-economics discipline applied to any other capital-intensive industry. The questions it surfaces — what does each token cost, what does each token earn, what is the spread, and is the spread widening — are the same questions a manufacturer asks about each unit produced. AI products that cannot answer these questions in production are not businesses yet. They are usage subsidies waiting to be repriced.