Who Prices Cognitive Labor in the Age of Agents? Compute-Anchored Wages

Abstract: A natural intuition about the economics of AI agents is that, because agents can be replicated at very low marginal cost, agent labor may be supplied highly elastically, placing downward pressure on cognitive-labor wages when it closely substitutes for human labor. We argue this framing is wrong in mechanism but partially correct in conclusion, and that the correction matters for both theory and policy. \textbf{Agents are not labor; they are a production technology that converts compute capital $K_c$ into effective units of cognitive labor $L_A$.} Once this is recognized, the elastic-supply margin that anchors the equilibrium wage migrates from the labor market to the compute capital market. Building on the classic factor-pricing framework \citep{mankiw2020}, we derive a \emph{Compute-Anchored Wage} (CAW) bound stating that, on tasks where human and agent-produced cognitive labor are substitutes, the competitive human wage is bounded above by $λ\cdot k \cdot r_c$, where $r_c$ is the rental rate of compute capital, $k$ is the compute intensity of one effective agent-produced cognitive labor unit, and $λ$ is the relative human-to-agent productivity. We generalize the result through constant elasticity of substitution (CES) aggregation, separate substitutable from complementary tasks, and discuss factor-share consequences. The conclusion is concise: \emph{the price-setter for cognitive labor is no longer the labor market.}

• The paper demonstrates that compute market parameters, not traditional labor supply, determine cognitive wage ceilings in AI-driven production.
• It introduces a formal framework recoding AI agents as technologies that convert compute capital into effective cognitive labor.
• Empirical calibration using GPU rental data shows that wage compression is driven by compute intensity, cost trajectories, and substitution elasticities.

Compute-Anchored Wages: Reframing the Pricing of Cognitive Labor in the Era of AI Agents

Introduction and Motivation

The paper "Who Prices Cognitive Labor in the Age of Agents? Compute-Anchored Wages" (2605.05558) provides a formal framework correcting widespread misconceptions regarding the impact of AI agents on the equilibrium price of cognitive labor. The dominant intuition in much policy and economic discourse holds that agents—due to their replicability at near-zero marginal cost—constitute an infinitely elastic supply of labor, collapsing wages for cognitive tasks to zero. This paper disputes that mechanistic account, arguing instead that AI agents are best understood not as labor, but as a production technology that transforms compute capital into effective cognitive labor. Consequently, the margin of elasticity and wage determination migrates from the labor market to the compute capital market, embedding cognitive wage ceilings in the factor pricing of compute.

Theoretical Foundations and Analytical Framework

The analysis is grounded in the classical marginal-productivity framework for factor pricing, maintaining the core apparatus of competitive equilibrium theory. The critical innovation is a recoding of the agent's economic role: agents operationalize a technology, $\phi$, that converts discrete bundles of compute capital $K_c$ into effective units of task-specific cognitive labor, $L_A$. The marginal cost and supply elasticity of cognitive labor performed by agents are thus governed not by labor-market dynamics, but by the cost structure and supply constraints present in the compute capital market. This reframing effects a pronounced shift: the labor supply curve no longer anchors the wage; instead, compute market parameters do.

The precise wage ceiling is encapsulated by the Compute-Anchored Wage (CAW) bound,

$W_H \leq \lambda\,k\,r_c,$

where $r_c$ is the competitive rental rate for compute capital, $k$ is the compute intensity (i.e., compute required per agent-labor unit), and $\lambda$ captures relative human-to-agent task productivity. In the limiting case of perfect substitution between human and agent labor, human wage equilibrium is exactly this ceiling whenever both $L_H, L_A > 0$.

Substitution Structure and Task Heterogeneity

Recognizing that perfect substitution is a strong assumption, the model generalizes to a CES aggregation, parameterized by the elasticity of substitution $\sigma$ between human ($L_H$) and agent-generated ($K_c$0) labor inputs. This relaxation offers empirical tractability, supporting the estimation of wage compression as a function of both compute cost trajectories and observed substitution elasticities at the task level. The analysis yields that CAW pressure is tightly modulated by the estimated value of $K_c$1 for a given task, predicting varying degrees of human wage compression dependent on empirical substitutability rather than the coarse binary of "automatable or not."

A fundamental implication is a directional inversion of canonical skill-biased technical change (SBTC). Historically, information technology complemented high-skill labor and substituted for routine work. Under the CAW model, the locus of wage pressure within cognitive tasks is defined by the task's exposure to agent substitution—a paralegal conducting rote contract review faces a CAW floor, while a senior associate focused on complementary, non-automatable tasks does not.

Empirical Calibration and Numerical Results

The paper conducts an illustrative calibration using contemporary (2024–2025) compute market data: H100 GPU rental rates, benchmarked inference compute requirements for both frontier and distilled models, and productivity ratios derived from LLM deployment in cognitive tasks. For high-volume, low-judgment tasks addressable by distilled models ($K_c$2 H100-hours/labor-hour), CAW predicts binding wage ceilings as low as \$K_c$3k \sim 1$K_c$41–\$K_c$5r_c$K_c$6\phi$K_c$7k$K_c$8r_c$) propagate directly to downward wage pressure across substitutable cognitive task segments.

Macro-Level Consequences and Policy Instruments

The migration of wage determination from labor to compute capital markets precipitates a nontrivial shift in factor shares, with cognitive labor's share of income compressed and compute capital's share expanded. Importantly, the holders of this capital income may differ from traditional capital owners due to the structural and regulatory features of the semiconductor, energy, and model IP markets.

The identification of CAW reframes the efficacy and distributional impact of several policy levers:

• Compute taxation raises the CAW ceiling proportionally, with incidence initially borne by compute owners (where supply is inelastic) but eventually translating to output prices and wages as supply expands.
• Public compute provision and antitrust on accelerators act oppositely, compressing markups and thus cognitive wage ceilings, but with potentially regressive distributional effects for cognitive labor.
• Energy policy becomes a direct channel for cognitive wage regulation, given energy's substantial role in compute $K_c$9.

Policy interventions targeting labor supply are rendered largely ineffectual for wage compression induced on the substitutable margin; only policies addressing compute market structure and cost can impact the CAW ceiling.

Limitations and Boundary Conditions

The framework explicitly delineates its limits: wage ceilings, not necessarily wage bills; a focus on the marginal-product component, omitting legal, signaling, or relational wage premia; reliance on competitive compute markets (with extensions necessary for monopoly or political rationing); endogeneity in compute efficiency ($L_A$0), and the fluidity of the $L_A$1/$L_A$2 task partition as the technology frontier advances.

The possibility of demand expansion (Jevons effects), Ricardian comparative advantage preserving some human employment, and the role of model-weight IP rents require careful treatment when extending the CAW analysis to aggregate or longer-run outcomes.

Conclusion

The CAW framework mandates a shift in both analytical and policy focus: cognitive wage determination—on substitutable tasks—is no longer anchored by the labor market but by the compute capital market. Quantitative wage ceilings are robustly expressed as functions of compute intensity, productivity ratios, and rental prices, all of which are amenable to empirical calibration. This migration necessitates new priorities in economic analysis and policy formulation, emphasizing the estimation of substitution elasticities at the task level, systematic monitoring of compute cost trajectories, and targeted intervention within the compute supply chain. The model predicts continued and monotonic downward drift of cognitive wage ceilings on the substitutable margin as technological improvements in compute efficiency or cost propagate through the production function, with significant implications for wage structures, income distribution, and the political economy of automation.