Economic Growth Under Transformative AI: A Computational Analysis of Automation’s Economic Impact

The End of 200 Years of Constant Growth

For more than two centuries, economists have marveled at what they call the “Kaldor Facts”—two remarkably stable patterns that have defined industrial economic growth since the early 1800s: approximately constant long-run growth rates in output per capita (around 2% annually) and approximately constant labor share of output (roughly 67% in most developed countries).

These patterns felt so fundamental that many economists treated them as economic laws. But groundbreaking research from Stanford University and the National Bureau of Economic Research reveals a startling truth: these regularities are historical anomalies, not natural constants. And transformative AI is poised to shatter both patterns in ways that could reshape civilization itself.

As the research demonstrates, “the global economic growth rate unambiguously accelerated with the onset of the Industrial Revolution; we have no reason to insist that it cannot accelerate again.” The mathematical models suggest that once AI enables capital to become a gross substitute for labor in production, we could see growth rates accelerate by more than an order of magnitude—potentially reaching 20% or higher annually while the labor share plummets toward zero.

This isn’t speculative futurism. It’s based on rigorous economic modeling using established frameworks, current AI capability trajectories, and empirical data from the past 200 years. The research identifies three distinct channels through which AI could trigger this transformation, with timeline analysis suggesting critical thresholds could be reached around 2035.

Three Channels of AI Economic Transformation

The research identifies three distinct mechanisms through which AI could fundamentally alter economic dynamics, each with different implications for growth rates, labor share, and societal outcomes.

Channel 1: Automating Production

The Mechanism: AI enables capital to become a gross substitute for labor in production, allowing machines to self-replicate without human intervention. Under this scenario, the elasticity of substitution between capital and labor rises above 1, meaning capital can effectively replace labor rather than merely complement it.

The Result: Growth rates dramatically accelerate while labor share collapses. The research shows that “if AI enables capital to become a gross substitute for labor in production alone, allowing machines to self-replicate, the growth rate would dramatically accelerate and the labor share would plummet toward zero, decisively breaking both Kaldor Facts.”

Current evidence suggests we may be approaching this threshold. Information processing equipment represents less than 2% of US GDP ($0.5T out of $29T in 2024), but AI capabilities are expanding beyond narrow domains into general-purpose automation.

Channel 2: Automating R&D

The Mechanism: AI systems become capable of conducting research and development, effectively substituting for human researchers and engineers. This creates the potential for self-improving AI systems that can accelerate their own development.

The Result: “Automating R&D in isolation has more ambiguous effects: it may accelerate growth if automated research proves highly parallelizable, but it could also leave growth rates bounded if fundamental constraints on the research process persist.”

The key variables are whether research can be parallelized effectively (allowing multiple AI systems to work simultaneously on different aspects of problems) and whether there are diminishing returns to research intensity. Current evidence from semiconductor R&D suggests λ/β ≈ 5, indicating that research acceleration effects could be substantial.

Channel 3: Full Automation

The Mechanism: AI systems automate both production and R&D simultaneously, creating a feedback loop where improved technology increases output, which increases capital accumulation, which enables further technological advancement.

The Result: The most radical transformation possible. Under certain parameter conditions, this produces “hyperbolic growth”—economic output that increases faster than exponentially, essentially approaching infinite growth rates in finite time before hitting natural resource or other physical constraints.

For the Cobb-Douglas case with both production and R&D automated, hyperbolic growth occurs when ξ/(1-α) > β. With natural resource shares under 5% (meaning ᾱ > 0.95), this condition “will not prevent hyperbolic growth in the event of fully automated production given any non-negligible capital share in R&D.”

Mathematical Models Predicting 10x Growth Acceleration

The economic models underlying these predictions rely on well-established frameworks, not speculative assumptions. The core mathematics comes from production functions that have successfully explained economic growth for decades.

The Substitution Parameter Framework

The key insight lies in the elasticity of substitution (ε) between capital and labor, defined by the substitution parameter ρ ≡ (ε-1)/ε. Current empirical estimates from Gechert et al. (2022) show ε ≤ 1, meaning ρ ≤ 0—capital and labor are gross complements, each making the other more productive.

But transformative AI changes this relationship fundamentally. Once ρ > 0 (ε > 1), capital becomes a gross substitute for labor. At ρ = 1 (perfect substitution), machines can completely replace human workers in all tasks.

Growth Rate Mathematics

With a 5% natural resource constraint (ᾱ = 0.95), the steady-state growth rate formula becomes:

g_Y = g_A / (1 – ᾱ)

Currently, with capital share around 33%, growth rates hover around 2% annually. But automation dramatically increases the effective capital share. When ᾱ rises from 0.33 to 0.95, the denominator (1 – ᾱ) shrinks from 0.67 to 0.05—a reduction of more than 90%.

This mathematical relationship means that even modest improvements in labor-augmenting technology (g_A) get amplified by more than an order of magnitude. A 2% technology improvement rate becomes a 40% economic growth rate under full automation scenarios.

Computational Requirements

The research provides concrete estimates of the computational capacity needed to reach these thresholds. The human brain performs approximately 10^15 computations per second (Carlsmith, 2020). Current Nvidia GPU stock could theoretically support around 10 million virtual researchers—roughly equal to the current global researcher population of 10 million.

However, AI systems remain approximately 4 orders of magnitude less sample-efficient than human brains. But with algorithmic efficiency improvements roughly halving computational requirements each year and GPU capacity doubling every 10 months, these gaps could close within 1-2 decades.

The Labor Share Collapse: From 67% to Near Zero

Perhaps the most profound economic shift predicted by this research involves the collapse of labor’s share of economic output. For 200 years, workers have captured roughly two-thirds of economic value created. That could change dramatically under AI automation.

Historical Labor Share Stability

Since the Industrial Revolution, labor share in developed economies has remained remarkably stable around 67%. This wasn’t always true—pre-industrial economies showed different patterns—but industrialization created conditions where human labor consistently captured the majority of economic value.

The research reveals this stability depends on capital and labor being gross complements (ε ≤ 1). When machines enhance human productivity rather than replace it, both factors capture economic gains. But once the substitution threshold is crossed (ε > 1), this relationship breaks down.

The Mathematics of Labor Share Decline

Under a Cobb-Douglas production function, labor share equals 1-α, where α is capital’s share. Currently, α ≈ 0.33, so labor captures about 67% of output. But as AI enables machines to substitute for workers across more tasks, effective α approaches 1.

The task-based model provides even more granular insight. With tasks i ≤ b automated and tasks (b, 1] requiring human labor, the relationship shows that “as b rises from 0 to 1, A falls from ∞ to 1 and B rises from 1 to ∞.” This mathematical structure demonstrates how complete automation (b → 1) drives labor share toward zero.

Transition Dynamics

The transition won’t necessarily be smooth. Rising wages naturally incentivize automation through what economists call endogenous automation effects (dating back to Habakkuk, 1962). There’s a critical threshold where K/L > (1/θ)·(α/(1-α)) triggers marginal perfect substitution.

Once this threshold is crossed, “as long as it is possible at some cost to build machines that can do all work, rising wages will eventually render them profitable, and then ever more so.” This creates a feedback loop where automation success drives further automation investment.

Timeline Analysis: Why 2035 Becomes Critical

While economic models can predict the relationships between AI capabilities and economic outcomes, estimating timelines requires analyzing current AI development trajectories and capability benchmarks.

Current Capability Acceleration

Several key metrics suggest rapid progress toward transformative AI thresholds:

• Task Length Capabilities: METR research shows the maximum task length AI can handle has been doubling approximately every 7 months
• Code Generation: Anthropic CEO Dario Amodei predicted in March 2025: “I think we’ll be there in three to six months, where AI is writing 90% of the code. And then in twelve months, we may be in a world where AI is writing essentially all of the code.”
• Adoption Velocity: Large language models reached over 1 billion users within 2.5 years of introduction
• AGI Forecasts: Metaculus median forecast for AGI is 2033, with some researchers predicting “near-complete R&D automation within several years”

Hardware and Efficiency Trends

The computational substrate supporting AI development continues accelerating:

• GPU computing capacity doubling every ~10 months for the last 6 years
• AI algorithmic efficiency improving with computational requirements roughly halving each year
• Current human-machine sample efficiency gap of ~4 orders of magnitude could close in 1-2 decades at current rates

The 2035 Convergence

Multiple trend lines suggest convergence around 2035 when AI systems could reach the capabilities needed to trigger economic transformation. This includes sufficient computational power, algorithmic efficiency, and task generalization to begin substituting for human labor across broad economic sectors.

Importantly, the research notes that “a notable characteristic of the R&D-initiated ‘singularity’ hypothesis is its prediction of rapid, nearly discontinuous transformation in economic conditions. Consequently, the absence of observable effects in current economic data cannot definitively refute the possibility of imminent dramatic change.”

Computational Capacity vs. Human Brain Power

Understanding when AI might trigger economic transformation requires comparing machine and human computational capabilities across relevant dimensions.

Raw Processing Power

The human brain performs approximately 10^15 computations per second, according to estimates by Carlsmith (2020). Current Nvidia GPU stock in early 2026 could theoretically support around 10 million virtual researchers if AI software achieved brain-level efficiency.

This matches the current global researcher population of roughly 10 million people (UNESCO, 2024), suggesting we’re approaching hardware parity for research-focused AI systems. However, raw computational capacity alone doesn’t determine AI capabilities.

Sample Efficiency Gaps

Perhaps more important than raw compute is sample efficiency—how much data AI systems need to learn tasks compared to humans. Current AI systems require approximately 4 orders of magnitude more data than human brains to achieve comparable performance on learning tasks.

But this gap is closing rapidly. Deep learning shows algorithmic efficiency improvements where computational requirements roughly halve each year (Eth and Davidson, 2025). At current improvement rates, AI could reach human-level learning efficiency within 1-2 decades.

Research Parallelizability

One area where AI could potentially exceed human capabilities involves research parallelization. The research identifies three reasons automated R&D might deliver λ > β (explosive growth):

1. Larger vs. more numerous systems: AI can form larger integrated systems rather than just more discrete brains, potentially achieving higher effective λ
2. Copies of Einstein: Multiple copies of the best AI models avoid the selection effects that limit expanding human researcher pools
3. Slow diminishing returns: Semiconductor R&D exhibits λ/β ≈ 5, and software efficiency improvements also show λ/β > 1

Constraints on AI Research Acceleration

Not all research can be parallelized effectively. Some constraints could limit AI’s ability to accelerate R&D:

• Serial Dependencies: Some experiments must be conducted sequentially, bounding growth rates regardless of researcher count
• Creative Destruction: Accelerating innovation could shrink expected monopoly durations toward zero, limiting investment incentives
• Learning by Doing: Real-world economic data may be gross complements to R&D, creating bottlenecks even for large numbers of virtual researchers

The Automation Threshold: Capital as Labor Substitute

The critical economic threshold occurs when capital becomes a gross substitute for labor—when machines can replace rather than merely augment human capabilities. Understanding this threshold helps predict when AI might trigger economic transformation.

Current Complementarity Relationships

Empirical research from Gechert et al. (2022) shows that capital and labor currently maintain elasticity of substitution ≤ 1, meaning they are gross complements. Each factor makes the other more productive: better tools enhance worker output, while skilled workers can utilize more sophisticated capital equipment.

This complementarity explains the stability of labor share over 200 years. When both factors benefit from technological progress, both capture economic gains proportionally.

The Task-Based Transformation

The research employs a task-based framework building on work by Zeira (1998) and Acemoglu & Restrepo (2018). Production gets decomposed into a continuum of tasks, where tasks i ≤ b are automatable and tasks (b, 1] require labor only.

As automation expands (b increases toward 1), a fascinating mathematical relationship emerges: automation proves equivalent to labor-augmenting rather than capital-augmenting technology. “Each time 1-b halves, capital can be spread more thinly across the widened range of automatable tasks, but labor is concentrated twice as heavily in each non-automatable task.”

The Inevitability Argument

The research presents what it calls an “inevitability argument”: rising wages will naturally incentivize automation until the substitution threshold is crossed. The critical point occurs when K/L > (1/θ)·(α/(1-α)), triggering marginal perfect substitution.

Once this happens, “as long as it is possible at some cost to build machines that can do all work, rising wages will eventually render them profitable, and then ever more so.” This suggests that transformation becomes economically inevitable once the technological capability exists.

Distributional Consequences and Wage Inequality

Even if AI automation dramatically increases total economic output, the distributional consequences could be severe. The research reveals why aggregate prosperity doesn’t necessarily translate to widespread benefit.

The Aggregate vs. Individual Paradox

The mathematics show that even rising aggregate wage bills can mask extreme inequality. As the research notes, “even if the aggregate wage bill stagnates or rises, wage inequality may grow such that many people cannot earn enough labor income to live on.”

This occurs because demand might remain strong for certain types of human services—”the world’s best athletes, for example, could keep the aggregate wage bill positive or rising while the majority, competing with robots… find their wages driven to zero.”

Where Human Value May Persist

Research by Korinek (2026) suggests demand may continue for services requiring:

• Authentic human connection: Childcare, therapy, and personal services where human identity matters
• Competitive/performative activities: Sports, arts, and entertainment where human participation is intrinsic to value
• Religious services: Spiritual guidance and religious leadership
• Human oversight: Serving as “the ultimate arbiter of values for AI alignment”

However, these sectors may employ only a small fraction of the current workforce, leaving most workers without viable income sources.

Capital Concentration Concerns

The research suggests AI automation might vindicate “Piketty (2014)’s concern about twenty-first century capital concentration.” If capital captures an increasing share of economic output while most workers find their skills obsoleted, wealth inequality could reach unprecedented levels.

Unlike historical technological transitions where new jobs emerged to replace automated ones, AI’s general-purpose nature means “humans possess some finite set of cognitive and physical capabilities, and once machines can replicate these underlying capabilities, they will necessarily be able to perform all feasible human tasks, including any newly created ones.”

Policy Responses to Economic Disruption

Recognizing the potential for dramatic economic transformation raises critical questions about policy preparation and response strategies.

The Challenge of Steering Technology

The research reveals a sobering reality about attempts to guide technological development: “It will be very inefficient (and so very difficult) to steer technology in a labor-augmenting direction permanently”—meaning keeping both growth rates and labor share bounded above zero.

This challenges optimistic hopes expressed by researchers like Korinek and Stiglitz (2020), Acemoglu and Restrepo (2020), and Brynjolfsson (2022) that technology could be directed to complement rather than substitute for human labor.

Mitigating Distributional Consequences

The research acknowledges the need for policies to “mitigate extreme distributional consequences” (Korinek and Lockwood, 2025) but doesn’t prescribe specific solutions. Potential approaches might include:

• Progressive taxation of capital returns to fund universal basic income
• Public ownership stakes in AI and automation technologies
• Retraining programs for workers displaced by automation
• Shortened work weeks to distribute remaining human-necessary labor

AI Safety and Governance

The rapid pace of transformation creates governance challenges. The research emphasizes advancing “technical AI safety research” for system alignment and robustness, along with developing “governance frameworks” for managing rapid capability transitions.

However, it warns against excessive caution that might “paradoxically increase long-term risks through delayed beneficial development” (Trammell and Aschenbrenner, 2025).

Building Economic Systems for the Post-Labor World

The research implications extend beyond predicting change to considering how economic systems might be restructured for a post-labor world.

Historical Precedent and Limitation

The Industrial Revolution provides both inspiration and caution. It “unambiguously accelerated the global economic growth rate” and created unprecedented prosperity, but also involved decades of social disruption and adjustment.

The key difference is scope. Historical automation was partial—humans remained essential in most economic roles. AI automation could be comprehensive, crossing “a fundamental threshold” where machines replicate all human capabilities.

Alternative Economic Models

A world where labor share approaches zero requires fundamentally different economic structures. Traditional models assuming labor-capital complementarity break down when ε > 1. New frameworks might need to account for:

• Capital as the primary production factor
• Human services in specialized, non-automatable roles
• Resource constraints becoming the primary growth limiter
• Distribution mechanisms independent of labor market participation

The Transformation Scale

The research suggests this transformation “may represent humanity’s most consequential transition since the Industrial Revolution, or perhaps since the Neolithic.” Such language might seem hyperbolic, but the mathematical models support dramatic claims.

When economic growth rates could increase by more than an order of magnitude while the foundation of income distribution (labor share) approaches zero, the scale of change could indeed rival agriculture’s invention or industrialization’s impact.

The question isn’t whether this transformation will occur—the economic incentives make it virtually inevitable once the technological capability exists. The question is how quickly it unfolds and whether societies can adapt institutions and policies fast enough to manage the transition constructively.

As the research concludes, understanding these dynamics now allows for better preparation. The models suggest we may have until approximately 2035 to develop appropriate governance frameworks, safety measures, and distribution mechanisms for an economy where traditional labor becomes economically irrelevant.