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 are not a new type of labor input with infinite supply; they are a capital-to-labor conversion technology that turns compute capital Kc into effective cognitive labor LA. As a result, the elastic margin that anchors equilibrium wages for substitutable cognitive tasks shifts from the household labor-supply margin to the compute-capital market. The paper derives a Compute-Anchored Wage (CAW) bound: on tasks where human and agent cognitive labor are (approximately) substitutes, the competitive human wage WH is bounded above by WH ≤ λ · k · rc, where rc is the rental rate of compute capital, k is the compute intensity (GPU-hours per agent-equivalent hour), and λ is the human-to-agent productivity ratio. With imperfect substitution the result generalizes via a CES aggregator with elasticity σ.

Key Points

• Conceptual recoding: AI agents are a production technology ϕ(Kc) that produces LA (agent-produced cognitive labor), not a labor supply from households.
• CAW bound (perfect substitute case): WH ≤ λkrc. If both human and agent labor coexist on a task, equality WH = λkrc holds in equilibrium.
• Compute market is the price-setter: the wage ceiling for substitutable cognitive tasks is set by compute rental rates and compute intensity, not by the human labor supply curve.
• CES generalization: with Leff = (α LρH + β LρA)1/ρ and σ = 1/(1−ρ), the relative-demand condition yields a continuous mapping from compute-price changes to human wages. Exposure to CAW pressure is governed by σ (elasticity of substitution) for each task.
• Heterogeneity across tasks: only tasks in the substitutable set TS face the CAW ceiling; complementary tasks (TC) remain priced by traditional factor margins. This implies a directional inversion within cognitive labor (some cognitive tasks face downward wage pressure; others do not).
• Factor-share consequences: compute-capital income and firm/IP rents become more central determinants of income distribution within cognitive sectors.
• Sensitivity is linear in λ, k, rc — reductions in k (algorithmic or architectural gains) or reductions in rc (cheaper compute) lower the CAW ceiling proportionally.
• Political-economy note: concentration in compute supply (fabs, cloud providers, geopolitics) affects rc and thus wages on substitutable tasks.

Data & Methods

• Theoretical framework: standard competitive factor-pricing / cost-minimization (Mankiw-style) with augmented production Y = F(Ko, LH, LA = ϕ(Kc)). First-order conditions link WH and rc to marginal products; LA supply is derived from Kc supply.
• Proposition derivation: unit-cost comparison (WH vs. λkrc) under perfect substitution; CES aggregation used to handle imperfect substitution and to introduce σ.
• Calibration (illustrative, not an econometric estimate):
  • rc: on-demand H100 GPU rental ≈ $2–$5/GPU-hour (multi-year contracts nearer $1.50); author uses rc ≈ $2/GPU-hour as midpoint.
  • k: frontier-model inference ≈ 0.5–2 H100-hours per agent-hour; examples use k = 1 (frontier) and k = 0.05 (small distilled models).
  • λ: productivity parity or gains from LLM usage reported in empirical studies; author considers λ ∈ {0.5, 1.0, 2.0}.
  • Resulting illustrative CAW ceilings range roughly from $0.05/hour (distilled small model, agent-favored) up to ~$10/hour (frontier model, human-favored, rc = $5). Table in paper shows full grid.
• Empirical inputs suggested: occupational/task exposure scores (e.g., Eloundou et al., Felten et al.), measured k and rc series (cloud or contract markets), and estimated σ per task/occupation from observed factor demands.
• Limitations acknowledged in methods: calibration is illustrative; IP rents and amortized training costs complicate the decomposition of rc; dynamic supply elasticities and general-equilibrium feedbacks are not fully modeled.

Implications for AI Economics

• Re-frames empirical focus: researchers should track compute-market prices (rc) and task-level compute intensity (k) and estimate the elasticity of substitution σ for tasks, rather than treating agents as an infinitely supplied labor type.
• Testable predictions:
  • Wages on substitutable cognitive tasks should co-move with λkrc (or with rc and k) — falls in compute costs or algorithmic improvements predicting wage declines on high-σ tasks.
  • Occupations/tasks with higher estimated σ will show larger wage/composition responses when compute prices change.
  • The labor share within cognitive work will compress while capital/compute shares and IP rents rise, especially in tasks where substitution is strong.
• Policy-relevant channels:
  • Compute-market policy matters for wage outcomes (antitrust, export controls, fab investment, energy constraints).
  • Interventions that affect rc (taxes/subsidies on compute, public compute provision, R&D to change k) can influence wage ceilings on substitutable tasks.
  • Worker adjustment: retraining and labor reallocation toward complementary tasks (TC) are central; safety nets and redistribution may be needed if workers cannot transition.
  • Regulation of IP rents and pricing (training-weight amortization) can shift how much of compute-related cost appears in rc vs. markup.
• Broader implications:
  • The CAW framework implies concentrated compute supply (e.g., few cloud providers or fabs) can translate into concentrated distributional power over cognitive-labor wages.
  • Macro/GPT models should ensure their short- and medium-run equilibria are consistent with the CAW relation on substitutable margins.
  • Policy debates about “AI and jobs” should be task-specific, focusing on substitutability and compute intensity rather than aggregate statements.
• Caveats for application:
  • CAW applies only to tasks where human and agent outputs are (approximately) substitutable.
  • Assumes competitive factor markets and clear rental markets for Kc; real-world frictions, IP licensing, and dynamic adoption can modify outcomes.
  • Long-run changes in compute supply elasticity, new tasks creation, and endogenous technology adoption are important for dynamics but lie outside the static CAW bound.

If you want, I can: - Extract the paper’s formal propositions and equations into a one-page math-only cheat-sheet. - Propose an empirical strategy and data sources to estimate k, rc, and σ for a selected occupation (e.g., paralegals or customer support).