AI agents don't make wages vanish — compute prices do. In tasks where human and agent labor substitute, human cognitive wages are capped by the cost of rented compute (scaled by agents' compute intensity and relative productivity), moving the key price-setting margin from labor to compute.

Main Finding

Agents should be modeled not as a new form of labor but as a production technology that converts compute capital K_c into effective cognitive labor L_A. Once that is recognized, the competitive margin that pins cognitive-labor wages moves from the labor market to the compute-capital market. For tasks where human and agent cognitive labor are substitutes, the paper derives a Compute-Anchored Wage (CAW) bound: w_human ≤ λ · k · r_c where r_c is the rental rate of compute capital, k is the compute intensity per effective agent-labor unit, and λ is human productivity relative to an agent. The result is generalized with CES aggregation, separated by substitutable vs complementary tasks, and implies important factor-share and skill-bias reversals.

Key Points

• Conceptual reframing:
  • Agents = production technology (convert compute → cognitive-labor units), not a new labor input with infinite supply.
  • Replicability reduces marginal cost of agents conditional on compute availability; the binding margin is compute supply and its price (r_c).
• CAW bound (w_human ≤ λ·k·r_c):
  • k: compute required to operate one effective agent-labor unit (compute intensity).
  • r_c: rental/price of compute capital (per unit of compute per time).
  • λ: relative productivity of a human compared to one agent (human/agent output per unit cognitive labor).
  • Intuition: the cheapest way to supply an additional unit of effective cognitive labor is to rent compute; competitive human wages cannot exceed the (agent-produced) cost of that unit adjusted for relative productivity.
• Generalizations and structure:
  • CES aggregation allows continuous substitution between humans and agents; elasticity of substitution matters for wage responses.
  • Distinguishes tasks where agents substitute for humans (downward pressure on wages bounded by CAW) vs tasks where agents complement humans (increasing demand or changing skill composition).
  • Yields a potential inversion of classical skill-biased technical change: depending on which tasks agents substitute/complement, wage impacts across skill groups can reverse historical patterns.
• Factor-share consequences:
  • Returns accrue more to compute-capital owners when agents substitute for humans → rising capital share of income.
  • The competitive wage ceiling is set by compute prices rather than pure labor supply elasticity.

Data & Methods

• Methodology: theoretical/modeling paper built on standard factor-pricing framework (Mankiw-style neoclassical factor pricing) and cost-minimization arguments.
• Core derivation:
  • Model treats agents as a technology mapping compute capital into effective agent-labor units; competitive firms choose between hiring humans or producing agent-labor by renting compute.
  • Solve first-order cost conditions to derive the upper bound on human wages in substitutable tasks.
• Extensions:
  • CES production functions used to relax perfect-substitute assumptions and to study varying substitution elasticities.
  • Task-based decomposition to separate substitutable vs complementary tasks and analyze their different equilibrium effects.
• Assumptions and limitations (explicit or implicit):
  • Competitive factor markets (or well-defined rental price for compute).
  • Well-defined compute intensity k per effective agent-labor unit (may vary across tasks in reality).
  • λ and elasticities known or stable enough for comparative statics.
  • Ignores some frictions and dynamics (e.g., transition costs, fixed R&D, input complementarities, market power beyond compute rental, regulation, externalities).

Implications for AI Economics

• Theory and forecasting:
  • Wage predictions for cognitive labor should condition on compute prices and compute intensity per task, not only on replicability of software.
  • Occupations/tasks facing high substitutability to agents will see wage pressure bounded by CAW; tasks complementary to agents may see rising wages/demand.
  • Elasticities of substitution and task-level compute intensities critically shape distributional outcomes (who gains/loses).
• Distribution of returns:
  • Greater returns to compute-capital owners and possibly rising capital share, rather than a pure collapse of labor value.
  • Concentration in compute supply or market power in cloud/accelerator provision can raise r_c and therefore raise the ceiling on wages — but also concentrate rents.
• Policy levers and regulatory implications:
  • Compute supply, price, and ownership become central policy levers for labor-market outcomes (e.g., taxes/subsidies on compute, limits on compute concentration, targeted compute access programs).
  • Traditional labor-market policies (minimum wages, retraining) remain relevant but may be incomplete if compute-market structure is ignored.
  • Antitrust and industrial policy applied to compute markets can alter equilibrium wages indirectly.
• Empirical tests and predictions:
  • Cross-industry/occupation correlations: wages should be negatively associated with task substitutability to agents and positively associated with task compute intensity times local compute rental rates.
  • Variation in cloud/compute prices (geographic or temporal) should predict differential wage pressures in occupations with high agent-substitutability.
  • Factor-share shifts: industries that adopt agent-heavy production should show rising capital income share correlated with compute intensity and compute-price changes.
• Practical consequences for stakeholders:
  • Workers: exposure depends on task substitutability; complementary-skill development may protect wages.
  • Firms: cost-optimization will trade off hiring humans vs renting compute; compute-capital strategy matters.
  • Policymakers: to influence labor outcomes, consider interventions on compute markets (prices, access, competition) in addition to classic labor policies.

Limitations and open questions to address in follow-up work: - Empirical measurement challenges: estimating k and λ at task/occupation level; obtaining fine-grained data on compute rental prices and their pass-through. - Dynamics: investment, endogenous compute supply, learning-by-doing, and market power dynamics could alter the static CAW bound. - Distributional policy design: how best to tax/redistribute rents when returns concentrate in compute capital.