Scenarios 1 and 2 document the mechanisms: labor displacement, capital concentration, platform dependency. This scenario asks a different question. At what point do those dynamics become irreversible?
Inequality is a condition. Caste is a structure. The difference is whether the people at the bottom can plausibly reach the top. Nick Bostrom projects that under full automation, the factor share of capital approaches 100% of world product. The remainder of the population becomes structurally superfluous. The enhanced class out-thinks the underclass in every domain. Bostrom’s vision of a post-transition Malthusian state, where impoverished rentiers reduce metabolic costs because even subsistence becomes unaffordable, follows logically from unrestrained automation without redistribution. But the endpoint is less important than the threshold. What matters is the moment when the feedback loops close and the exits seal shut.
The first lock is capital compounding. An IMF working paper published in April 2025 (Rockall, Tavares, and Pizzinelli) found that while AI may reduce wage inequality by displacing high-income workers, it “always” increases wealth inequality, in every scenario modeled, because displaced high-income workers are better positioned to benefit from rising capital returns. They own the assets whose value AI inflates. The workers below them do not. The model predicts a wealth Gini increase of over 7 percentage points under baseline AI adoption. The mechanism is simple: wages can equalize downward while wealth diverges upward. You can flatten everyone’s paycheck and still watch the ownership class pull away, because wealth generates returns and returns generate wealth. A separate study in the Bulletin of the World Health Organization (Occhipinti et al., 2025) proposes the existence of an AI-capital-to-labor ratio threshold beyond which a self-reinforcing recessionary cycle emerges that market forces alone cannot correct. The question is whether we have already crossed it.
Where the capital is landing is visible. The Magnificent Seven now account for 32.6 percent of the S&P 500 as of February 2026, up from 12.5 percent a decade ago. The top ten companies in the index represent approximately 39 percent of total market capitalization, well above the 27 percent peak reached during the dot-com bubble (Columbia Threadneedle, September 2025). The OECD reported in February 2026 that AI firms captured 61 percent of all global venture capital in 2025, with mega-deals above $1 billion representing roughly half of total AI investment value. UNCTAD’s Technology and Innovation Report 2025 found that just 100 firms, mainly in the U.S. and China, account for 40 percent of global corporate R&D spending. The leading tech giants each have a market value rivaling the GDP of the entire African continent. Cresset Capital describes the resulting structure as a reflexive loop: AI investment inflates asset values, asset values underwrite consumer confidence among the wealthy, and their spending props up an economy whose broader foundations are eroding (Cresset Capital, February 2026). The top 20 percent of U.S. households now own nearly 70 percent of financial assets. Their spending contributed approximately 1.1 percentage points to GDP growth in 2025, more than half of total consumption growth. For the bottom 80 percent, spending growth barely kept pace with inflation.
A February 2025 paper by Caleb Maresca at NYU (Strategic Wealth Accumulation Under Transformative AI Expectations) exposes a mechanism that makes this concentration self-accelerating. If wealth at the moment of AI deployment determines each household’s share of future AI-automated labor, then the rational response is to accumulate as much wealth as possible before that moment arrives. Savings become claims on future AI labor. The result is a prisoner’s dilemma: every household tries to save more, driving up interest rates (the model predicts 10-16 percent versus approximately 3 percent without strategic competition), while the collective effort cancels out any individual advantage. Everyone consumes less in the present. Nobody improves their relative position. The wealthy start ahead and stay ahead. The mechanism does not require anyone to act irrationally or maliciously. It follows from the logic of the situation itself.
The second lock is mobility collapse. Raj Chetty’s research at Harvard, published in Science (Chetty et al., 2017), showed that 92 percent of American children born in 1940 grew up to earn more than their parents. For children born in 1984, the figure was 50 percent. Two-thirds of the decline was attributable to unequal distribution of growth, not slower growth itself (Equitable Growth, 2019). AI accelerates the distributional skew that drove the collapse. The traditional escape route, education, is breaking down. Georgetown’s Center for Security and Emerging Technology found that technical skills now become outdated in less than five years on average, while AI, unlike previous technology waves, disrupts both blue-collar and white-collar workers simultaneously. The retraining system that is supposed to catch displaced workers has a dismal evidence base. Julian Jacobs at Brookings reviewed the record in May 2025 and found that the National JTPA Study, a genuine randomized controlled trial, showed no statistically significant improvement in employment or earnings for participants. A national evaluation of the Workforce Investment Act found that training services had no positive impact on earnings or employment within 30 months. Roughly 40 percent of current WIOA participants are trained into low-wage support roles paying less than $25,000 per year. The U.S. spends approximately 0.1 percent of GDP on active labor market policies, second-to-last among OECD countries, where peers spend up to 0.5 percent. When the ladder is broken and the replacement ladder does not work, the people at the bottom stay at the bottom.
The dynamics described above are U.S.-centric in their evidence base but global in their structure, and in the Global South they operate on populations with far less institutional cushion. UNCTAD’s Technology and Innovation Report 2025 found that the 100 firms dominating global AI R&D are concentrated almost entirely in the United States and China, while the countries most dependent on the service-sector jobs AI threatens, India, the Philippines, Kenya, and others that built economic development strategies around business process outsourcing, call centers, and back-office services, have no seat at the table where AI’s capabilities are defined. India’s IT services industry employs over 5 million people directly and supports roughly 12 million more, and its growth model depends on providing cognitive labor to Western firms at lower cost. AI does not merely threaten individual Indian workers. It threatens the national development strategy itself. The Philippines’ BPO sector, which accounts for roughly 7 percent of GDP and employs 1.3 million people, faces the same structural exposure. Africa, home to 18 percent of the world’s population, accounts for less than 1 percent of global data center capacity and an even smaller fraction of AI research output. The neo-feudal dynamic is not only within nations. It is between them. A world in which AI capability is concentrated in a handful of U.S. and Chinese firms while the Global South provides data, cheap annotation labor, and consumer markets reproduces colonial extraction patterns with a computational substrate. The governed outcome requires not only domestic redistribution within wealthy nations but international frameworks that give developing economies genuine agency in how AI is built, deployed, and governed, a dimension that current governance discussions, centered on the EU-U.S.-China triangle, largely ignore.
The third lock is cognitive asymmetry. The standard counterargument is that AI tools are cheap. A $20 subscription provides analytical capacity previously available only to expensive consultants. Open-source models are free. This has real force on the access side. It has less force on the compounding side. Consumer AI and corporate AI are diverging into separate tiers with meaningfully different capabilities. Average monthly enterprise AI spending reached $85,521 in 2025, a 36 percent increase from the prior year, and the proportion of organizations planning to invest over $100,000 per month more than doubled (CloudZero State of AI Costs 2025). KPMG found that over two-thirds of enterprise teams plan to spend between $50 million and $250 million on generative AI in the next year. Reworked reported in December 2025 that the gap between organizations that can afford frontier AI capabilities and those that cannot is widening, creating market segmentation where cutting-edge AI remains expensive while everyday AI becomes progressively cheaper. The person using free-tier AI to write a resume is competing against a firm using enterprise AI to screen ten thousand resumes per hour. You can have a smartphone and still be structurally poor. You can have ChatGPT and still be structurally outmatched.
There is a more sophisticated version of the access counterargument. Several studies have found that within specific occupations, lower-skilled workers derive greater productivity gains from AI than their higher-skilled counterparts. David Autor at MIT has cited this evidence as support for the hypothesis that AI could boost middle-class wages and reduce inequality. The finding is real. It is also insufficient. A GovAI analysis at Brookings showed that these within-occupation studies are misleading when extrapolated to economy-wide effects, because most of the evidence comes from workers already in high-paying professions. AI may close the gap between the worst and best programmer inside programming, while the programming occupation itself contracts and capital claims the surplus. The task-level equalization and the occupation-level contraction operate on different axes. Anthropic’s own labor market research, published in March 2026, confirms the demographic pattern: workers in the most AI-exposed occupations earn 47 percent more, are nearly four times as likely to hold graduate degrees, and are 16 percentage points more likely to be female than the zero-exposure group. The displacement pressure is concentrated at the top of the income distribution. The underclass does not form through direct displacement of the poor. It forms because AI displaces the educated professional class, the capital gains from that productivity displacement flow to asset owners, the displaced professionals compete downward into service and physical roles, and the workers already at the bottom absorb the compression from above.
The cognitive gap compounds across generations. Brookings warned in February 2024 that AI risks creating a “next digital divide” in education, where access to technology becomes universal while access to the human guidance that makes technology effective remains scarce and stratified. Research published by Wilson Wong at the Chinese University of Hong Kong (PA TIMES, September 2025) found that AI lowers the cost of knowledge while raising the premium on the complementary capacities that make knowledge useful: AI literacy, ethical judgment, human-AI collaboration. When those capacities are not developed universally, they are acquired disproportionately by already-advantaged students and institutions. The result is compounding inequity: unequal access to AI, unequal learning outcomes, unequal employment, unequal income. Robin Lake, director of Arizona State University’s Center on Reinventing Public Education, told NPR in August 2025 that the “AI divide is starting to show up in just about every major study” she sees. Chetty’s earlier research established that neighborhood and environment are the strongest determinants of economic mobility. AI adds a new layer of environmental advantage. It follows children through screens.
The corporate sector is not waiting for the debate to resolve. In April 2025, Shopify CEO Tobi Lutke published a company-wide memo declaring that teams must prove AI cannot do a job before requesting a human hire. AI usage was added to performance reviews. Shopify had already cut headcount by roughly 36 percent between 2022 and 2024. Within weeks, Duolingo CEO Luis von Ahn issued a parallel memo declaring the company “AI-first,” stopping the use of contractors for tasks AI can handle and mandating that new headcount would only be granted if teams could demonstrate automation could not achieve the required work. Klarna had already gone further: a hiring freeze, a 40 percent headcount reduction, and public celebrations of AI replacing 700 customer service agents. (By July 2025, Klarna reversed course and announced a hiring spree, admitting that quality had suffered.) The pattern emerging across these companies is consistent: AI adoption as a gate that employees must justify their existence against, with the default assumption running toward replacement. The market rewarded the announcements. Block’s stock rose 25 percent on a single earnings call after announcing AI-driven layoffs. The signal to every other CEO was unambiguous.
The fourth lock is political capture. When the enhanced class accumulates enough wealth and intelligence to shape the rules governing both, redistribution becomes structurally impossible. Scenario 2 documents the lobbying data. What matters here is the implication: the window for policy intervention is not indefinite. Every year that concentration deepens without redistribution makes future redistribution harder, because the concentrated class gains more power to prevent it. This is the feudal parallel. Aristocracies did not merely possess wealth. They wrote the laws that perpetuated their possession. The transition from inequality to caste is the transition from “the rules could be changed” to “the people who benefit from the rules control the rulemaking.”
The fifth lock is geographic, and it operates at every scale simultaneously. Brookings research published in October 2025 found that generative AI’s geographic impact inverts the pattern of previous automation waves. Previous technologies hit lower-paid, less-educated metro areas hardest. Generative AI hits high-skill, high-paying metros like San Jose, San Francisco, and New York, where up to 43 percent of workers could see AI shift half or more of their tasks. The difference is that these metros have the resources, innovation ecosystems, and reskilling infrastructure to adapt. A separate Brookings analysis found that 30 metro areas now account for two-thirds of all AI-related job postings in the United States, with Anthropic reporting per capita Claude usage in Washington, D.C., at 3.82 times the expected rate while Mississippi registered 0.21 times. The concentration locks in through path dependency: early advantages in AI talent, infrastructure, and adoption compound into positions that become increasingly difficult to challenge.
At the international level, the pattern is starker. UNDP’s December 2025 report, titled The Next Great Divergence, warned that AI could reverse decades of narrowing development inequality between nations. Fewer than a third of developing countries have national AI strategies. 118 nations, mostly in the Global South, are absent from global AI governance discussions entirely. Internet access stands at 27 percent in low-income countries versus 93 percent in high-income ones. Fixed broadband costs account for 1 percent of monthly income in wealthy nations and 31 percent in the poorest ones. AI-driven automation favors capital over labor, which erodes the competitive advantage that low-cost labor has provided to developing economies for decades. The mechanism by which billions of people in the Global South climbed out of poverty, offering labor at a price that attracted foreign investment, is exactly the mechanism that AI undermines. The Center for Global Development warned in December 2025 that wealthier nations could use AI-driven manufacturing to reshore production, outcompeting developing countries on cost, speed, and product quality. The convergence between nations that marked the period since 2000 could reverse into a new era of divergence.
The sixth lock is the disappearing floor. Anthropic’s March 2026 labor market research classifies 30 percent of U.S. workers as having zero AI task exposure: cooks, dishwashers, bartenders, mechanics, lifeguards. These jobs are invisible to language models because they require physical interaction with the environment. They are not invisible to robots. Boston Dynamics began commercial production of its Atlas humanoid in January 2026, with Hyundai planning 30,000 units per year from a single factory by 2028. Chinese manufacturer Unitree launched the R1 humanoid at $5,900 in July 2025, following its G1 at $16,000 and H1 at $90,000. Goldman Sachs reported manufacturing costs declining 40 percent year-over-year, faster than the projected 15-20 percent annual decline. At scale production costs over a five-year lifespan operating 20 hours per day, a humanoid robot’s effective hourly cost falls to $3-5, below the minimum wage in every U.S. state. Morgan Stanley projects over 1 billion humanoids in service by 2050, with 63 million in the U.S. alone potentially affecting 75 percent of occupations and $3 trillion in payroll. The “safe” jobs are safe only because they are currently cheaper to perform than to automate. That is a cost arbitrage on a declining curve. Each production doubling compresses the gap. The floor is not a floor. It is a price spread with an expiration date.
Daron Acemoglu, the 2024 Nobel laureate in economics, provides the historical frame. In a January 2025 lecture at the University of Zurich, he noted that benefits from the British Industrial Revolution took more than 100 years to diffuse to workers. Three generations lived and died in immiseration before institutional pressure forced redistribution. His research with Pascual Restrepo shows that since 1987, automation has displaced roughly 16 percent of jobs in affected industries while reinstating only 10 percent, a sharp deterioration from the prior period when reinstatement ran at 19 percent (MIT News, 2020). The new jobs that do appear increasingly benefit high-skill workers, leaving displaced low-skill workers with no equivalent replacement. Acemoglu and Johnson’s Power and Progress (2023) documents that diffusion was never automatic. It was fought for, through labor movements, antitrust action, progressive taxation, and public investment. The question is whether those fights are being fought now. The current evidence, documented across the first two scenarios and reinforced by the data above, suggests they are not.
The 6-point governance dividend reflects enormous policy power. Universal AI access, progressive taxation of AI-generated wealth, public ownership stakes in frontier systems, and sustained social investment could prevent lock-in. These are the same mechanisms that diffused the gains of every prior general-purpose technology. The difference is speed. Previous technologies gave societies decades to organize responses. AI capability is compounding on timescales of months. The window between “inequality that policy can reverse” and “caste that policy cannot reach” may be shorter than any previous transition has allowed.
Key tension: Every lock described above is currently tightening. None is yet fully closed. The sixth removes the last structural floor: the category of work where human embodiment provided a final advantage is now on a cost curve converging with automation. The Maresca model reveals something worse: even rational actors trying to prepare for AI are trapped in a competitive dynamic that worsens inequality before AI arrives. The governed outcome requires acting in a window whose duration no one can measure and whose closing no one will announce.