Anthropic's 2030 Economic Model: Growth, Displacement, and Capital Share

Source material: Scenarios for our Economic Future

The Anthropic Institute · 2026-09-09

Research and editorial team: Anton Korinek, Charles I. Jones, Szymon Sacher, Tess Cotter, Peter McCrory, and Santi Ruiz

Technical report and supporting materials: Anthropic: Economic Scenarios for Transformative AI Technical Report · Anthropic: Labor Interventions Policy Research · Anthropic: Economic Index

Public debates around artificial intelligence and the future of work often collapse into rigid binaries: unprecedented abundance or catastrophic unemployment.

The Anthropic Institute Economics team, alongside Stanford macroeconomist Chad Jones and collaborators, built a quantitative macroeconomic model evaluating the economic effects of AI from 2026 through 2030. The model moves past speculative headlines to trace the transmission channels through which AI alters growth, wages, occupational churn, and factor shares.

Its core finding is stark: even when advanced AI sparks historical records in output growth, aggregate prosperity does not translate into widespread wage gains for human workers. Because automation targets cognitive labor, the expansion in economic wealth flows overwhelmingly to capital, leaving knowledge workers with stagnant wages and prolonged displacement.

The framework operates not as an infallible forecast, but as an analytical lens. By decomposing jobs into specific tasks and tracking factor accumulation, the authors quantify the structural tension between accelerating technological capabilities and labor market outcomes.

Jobs As Bundles of Tasks

To evaluate how algorithmic advances diffuse through production, the model adopts the occupational task-based taxonomy established by the US Department of Labor O*NET database.

Every occupation across the economy represents a dynamic bundle of dozens of distinct tasks. In a hospital setting, a registered nurse completes daily patient rounds, draws blood, triages incoming cases, logs vital signs, and manages supply orders.

Technology reshapes this bundle along different dimensions. Machine intelligence cannot physically bathe a sick patient, leaving that task entirely to humans. AI can assist nurses in drafting discharge summaries, monitoring telemetry data, and organizing care shifts, thereby augmenting task productivity. Other tasks, such as transcribing charts or tracking inventory, face full end-to-end automation.

Technological progress also generates brand-new tasks. Nurses must supervise algorithmic recommendations, verify clinical suggestions, and oversee automated care workflows. The center of gravity of the profession shifts toward patient-facing empathy and complex clinical judgment.

Aggregated across every worker nationwide, these recurring activities comprise the more than $30 trillion annual US economy. How AI shapes future prosperity depends directly on whether it augments human labor, substitutes for existing tasks, or generates sufficient new cognitive demands.

Three Scenarios: From Modest Diffusion to Supercharged Growth

Evaluating variations in capability growth, organizational adoption, and task reallocation, the researchers project three distinct paths through 2030.

The benchmark case assumes an economy without AI. In this baseline, annual US GDP expands at 2.0%, labor captures 60.0% of national income, capital receives 40.0%, aggregate unemployment holds at 3.8%, cognitive unemployment sits at 2.9%, and the net return to capital remains 6.5%.

Under the modest scenario, AI generates productivity gains comparable to the diffusion of the commercial internet during the 1990s. By 2030, real GDP rises 1.6% above baseline to $34.1 trillion in 2025 dollars, with annual growth accelerating slightly to 2.4%. Average wages climb 0.7%, the labor share edges down to 59.4%, the capital share reaches 40.6%, and the net return to capital rises to 6.6%. Cognitive employment dips by 0.5%, and total unemployment remains steady at 3.9%.

Under the substantial scenario, AI becomes a pivotal general-purpose technology surpassing the historical impact of the railroad and the internet. Real GDP reaches $36.3 trillion in 2030, standing 8.3% above baseline, with annual growth hitting 5.4%. This pace exceeds the 4.7% peak seen in 1999 during the dot-com expansion. The labor share falls from 60.0% to 56.1%, the capital share rises to 43.9%, cognitive employment drops 3.9%, non-cognitive employment rises 4.6%, total unemployment reaches 4.6%, and the net return to capital climbs to 7.0%.

Under the extreme scenario, systems capable of recursive self-improvement diffuse rapidly through corporate workflows. Real GDP expands 32.4% above baseline by 2030, reaching $44.4 trillion. Annual GDP growth accelerates to 15.4%, a velocity that doubles national output every 4.5 years.

Survey data collected in late summer 2026 by Anthropic and Morning Consult across more than 10,000 American adults reveals that public expectations mirror these projections. The median respondent anticipated outcomes aligned with the substantial scenario, while approximately 10% expected extreme transformation.

The Divergence Between Cognitive and Physical Wages

Macroeconomic expansion does not distribute gains symmetrically across the workforce. The model identifies a pronounced wage divergence separating cognitive occupations from non-cognitive physical trades.

In the substantial scenario, overall average wages increase 2.1%. However, while non-cognitive manual wages rise 5.9%, wages for cognitive knowledge workers experience a slight contraction of 0.3%.

In the extreme scenario, this structural bifurcation becomes dramatic. While economy-wide average wages rise 9.7%, wages for physical roles including electricians, construction trades, and bedside clinicians jump 33.6%. In contrast, cognitive wages plunge 11.5%.

This diverging trajectory reflects shifting factor scarcity. Automated code generation, legal drafting, data analysis, and administrative processing commoditize human knowledge work. As the supply of synthetic cognition multiplies, the relative economic scarcity of human brainpower diminishes.

Physical execution becomes the binding bottleneck of the production chain. Even if software can generate architectural blueprints and process permitting documents in seconds, building physical facilities requires electricians and specialized laborers. Surging cognitive efficiency heightens demand for physical execution, bidding up non-cognitive wages.

A Growing Pie Captured Primarily by Capital

In historical US national accounting, for every $1 of output produced, about 60 cents goes to labor and 40 cents goes to capital.

The model demonstrates that pervasive automation raises the marginal productivity and pricing power of capital assets. In the substantial scenario, the capital share climbs to 43.9%. In the extreme scenario, the capital share surges to 54.8%, while the labor share contracts to 45.2%.

Under the extreme scenario, despite the total economy growing by nearly a third, aggregate earnings for all workers increase by merely 0.5% relative to the baseline path.

Because GDP expands by 32.4% (an index multiplier of roughly 1.32) while the labor share falls to 45.2%, their product remains approximately 0.60. The entire incremental gain generated by transformative growth accrues to owners of compute clusters, power facilities, and proprietary algorithms. The net return to capital jumps from 6.5% to 8.3%, an increase of 28%, driving total capital income 81.4% above baseline.

An expanding economic pie does not automatically guarantee broad prosperity. Without deliberate public policy interventions and redistributive mechanisms, rapid growth can coexist with economic distress among knowledge professionals.

Occupational Friction and Model Boundaries

In addition to shifting factor shares, real-world labor markets confront friction that slows occupational adjustment and generates persistent unemployment.

Displaced software engineers and corporate analysts cannot seamlessly transition into electrical contracting or surgical nursing. Re-skilling, professional credentialing, geographic relocation, and occupational preference require significant time.

In the substantial scenario, cognitive employment shrinks by 3.9%, pushing cognitive unemployment from 2.9% to 4.5%, a rise of more than 50%, while overall unemployment reaches 4.6%. In the extreme scenario, cognitive headcount drops 21.5%, driving cognitive unemployment to 17.9% and total unemployment to 11.9%. This surpasses the near 10% peak observed after the 2008 financial crisis and the 8% peak during the Covid-19 recession.

The authors acknowledge explicit structural boundaries in the simulation. The framework excludes autonomous physical robotics, fiscal stabilization measures, tax adjustments, business cycles, and the Keynesian aggregate demand stimulus created by data center construction.

External economists who reviewed the technical paper, including Daron Acemoglu, David Autor, Ben Moll, and Emi Nakamura, noted that the model abstracts from micro-level worker tenure and firm heterogeneity. Even within these boundaries, the analysis underscores a central insight: technological progress does not resolve inequality on its own, and managing the AI transition requires policy frameworks capable of matching the velocity of productive transformation.