Skip to main content

Cost of Cognition

What happens to every industry built on expensive thinking when the cost of thinking collapses?

The cost of cognition is collapsing the way distribution costs collapsed when software ate physical media. That prior shift did not shrink demand for content — it exploded it. The same dynamic is now running on knowledge work: cheaper reasoning expands how much gets attempted, re-rates sectors that were cost-disease victims, and shifts scarcity to something harder to automate than thinking.

Understanding this thesis tells an operator or strategist where surplus will accrue, where linear-labour middle layers face pressure, and what it means to build something that lasts.

Deflation Curve​

Cognitive work is following the same deflation path software ran on distribution. Inference costs for frontier AI models dropped more than 99% between 2020 and 2025 [source: benchmark data from public model pricing; heuristic — exact trajectory is model-dependent]. At some threshold, the question stops being "can we afford to apply AI here?" and becomes "why would we not?"

The pattern is recognisable. Unit cost of a capability falls far enough that cost stops being the binding constraint on usage. Volume then rises to absorb the excess capacity — and often rises faster than unit cost fell, so total spend on the capability goes up while per-unit cost goes down.

Three dynamics follow from that deflation.

Jevons Paradox​

Cheaper cognition does not shrink total demand for reasoning — it expands it. This is the Jevons paradox applied to intelligence: when the resource becomes affordable, use cases that were never economically viable come to life.

Work that was never worth a human hour at $150/hour may be entirely worth attempting at a marginal cost near zero. Creative exploration, hypothesis generation, document drafting, scenario analysis, customer communication — each has been rationed by price. Deflation un-rations them.

The implication: do not forecast AI's economic impact by counting how many current tasks get replaced. Count how many tasks were never started because they cost too much.

Baumol Re-Rating​

Baumol's cost disease describes sectors where productivity gains are structurally blocked. A live performance, a therapy session, a legal consultation, a nursing shift — these could not be made faster without losing the core value. So their costs rose in step with the rest of the economy's wages, even though output per hour barely moved.

Those sectors are the ones AI re-rates most sharply. When cognition becomes a near-zero input, the labour-priced, presence-limited ceiling lifts.

A single practitioner can serve more clients. A single consultation can run with AI assistance that was previously unavailable. The cost structure re-rates — not because output quality falls, but because the bottleneck shifts.

The sectors most protected from prior productivity gains are, paradoxically, the ones where deflated cognition has the largest effect.

Services to Products​

Labour-priced, linear services have a hard ceiling: output scales only as fast as headcount. When cognition deflates enough, that delivery shape changes. A repeatable knowledge service becomes an outcome-priced product with near-zero marginal cost per additional customer.

The surplus from that conversion does not distribute evenly. It accrues to owners of proprietary loops — the organisations that hold the decision logic, the data, the feedback mechanism, and the customer relationship that makes the loop improve over time. Those are the logic moats.

Linear-labour middle layers face the sharpest pressure. If the value a role provided was reliable cognitive throughput — reading, writing, classifying, summarising, drafting — and that throughput is now available at near-zero cost, the role's market value compresses toward the coordination cost of hiring a person instead of invoking a tool.

That pressure is not a prediction about job counts. It is a claim about where value flows when a constraint dissolves: away from the constraint and toward what remains scarce.

New Bottleneck​

When cognition is abundant, scarcity moves. The new bottleneck is judgment — the capacity to recognise which problem is worth solving, which output is trustworthy, which customer relationship earns the next contract, which experiment is worth running.

Judgment is hard to automate not because it requires esoteric reasoning but because it requires context that is tacit, situational, and accumulated through experience and accountability. An AI can draft a legal argument; it cannot carry the relationship accountability of a partner who has been wrong in public and repaired it. An AI can generate a business forecast; it cannot hold the conviction through a board meeting under pressure.

Taste and distribution follow the same logic. When output is cheap, the selection problem intensifies: which output to trust, which direction to pursue, which audience to reach. Those selection acts require human judgment applied to a sea of machine-generated material.

That is where durable value concentrates. And it is where durable work concentrates too — not in the tasks that deflated, but in the judgment acts that govern them.

Positioning​

The deflation thesis has one practical diagnostic: is your venture building a proprietary loop or renting one?

A proprietary loop accumulates data, decision logic, and customer proof with each transaction. The loop improves faster than a competitor can replicate it because the improvement inputs are locked to the loop. One signal of a proprietary loop: the loop is harder to replicate after six months than it was at launch, because the data and decisions embedded in it are not available to a competitor who buys the same underlying model.

Renting a loop means using commoditised AI infrastructure to deliver a service that any other operator could equally deliver. The deflation benefit flows to the customer as lower prices, not to the operator as margin. That is not inherently wrong — buyers of deflated cognition capture real value — but it is not a durable competitive position.

The question worth asking before committing capital: when cognition deflates further, does this venture capture more of the benefit or pass it through?

Context​

  • Market Forces — endogenous vs exogenous framing for AI and crypto as economic forces; layer-by-layer question set
  • Economy of Things — machine identity, machine payments, and the protocol layer for an economy where devices are economic actors
  • AI Economy — the spine connecting cognition costs to agent commerce and coordination
  • Performance Metrics — the scoreboard for whether a loop is compounding; deflation improves the economics only when the loop's output is measured
  • Purpose and Intentions — judgment is the scarce resource; purpose is what shapes how it is spent
  • Tight Five — the operating lens that connects purpose, performance, platform, process, and people in a compounding loop

Questions​

What in your current model depends on cognition being expensive — and what breaks when it is not?

  • If Jevons applies, which problems in your industry have gone unattempted because the cognitive cost was too high — and who reaches them first?
  • Which sectors in your market are Baumol-constrained today, and what does a re-rated cost structure unlock for the buyer?
  • What makes your loop proprietary rather than replicable — data, relationships, accountability, or something else?
  • When judgment and taste become the primary scarcity, who in your organisation holds them — and is that visible in how you allocate attention?