Most discourse on AI economics these days looks at “will AI take my job?”, i.e. where will “capital substitute for labour”, and assumes automation is an incremental diffusion. Philip Trammell and Dwarkesh Patel’s “Capital in the 22nd Century”, one of the most thought provoking pieces published over the holiday break, goes further and farther and suggests that the AI capital winners could take all. Thomas Piketty’s Capital in the Twenty-First Century argued that without strong redistribution, inequality compounds indefinitely… the rich save more, earn higher returns, and pull ever further ahead. Critics challenge this as historically unsubstantiated and argue that capital and labour have been complementary. Trammell and Patel’s “Capital in the 22nd Century” (the title is no accident) is subtitled: “Piketty was wrong about the past, but AI will make him right about the future.”
Historically they argue, capital accumulation has been self-correcting as “more hammers” mean more demand for hands to use them, pushing wages up. But once AI can do everything humans can, this correction breaks. Returns accrue entirely to wealthy capital owners, and inequality compounds without limit. In our year-end analysis, we also identified “an imperative for capital capture rather than worker distribution” arising from the intensive up front AI investment model. We warned about “efficiency gains stranded in an economy with insufficient purchasing power” if automation outpaces adaptation. We noted the hidden displacement already underway: job losses, frozen hiring, and severed entry-level career paths creating a “silent iceberg” of workforce disruption.
We framed the challenge as adoption: 2026 is a tightrope between two existential risks, a disorderly unwinding of the productivity overhang triggering a job displacement crisis, or a persistent adoption lag exploding the debt-fuelled off-balance sheet bubble. We argued for “genuine human-machine teaming” so capital captures its necessary return without suffocating the economy. But Trammell and Patel describe a bigger risk: even after safely navigating this narrow path, the platform may be unstable. Our framing assumed complementarity, but the authors believe this will only be temporary. Full labour substitutability will lead to “Jevons world,” where capital accumulation no longer faces diminishing returns that historically raised wages; human-AI teaming will be transitional, not equilibrium.
The essay has drawn substantial pushback. Ben Thompson’s Stratechery rebuttal argues that human preferences for artisanal services, authentic experiences, and human connection could sustain labour’s share even in full automation. But will preferences alone preserve labour’s slice of the pie? In a world of genuine abundance, or one where AI-accelerated scientific progress reshapes what humans can do and want, the zero-sum distributional logic may not apply. Their model assumes we’re fighting over shares of a fixed or steadily growing output. What if the nature of value generation changes so fundamentally that “labour share” becomes the wrong question?
Zvi Mowshowitz, AI blogger, struggles with the fundamental assumptions of the analysis, that human institutions persist: property rights enforced, legal systems functioning, AIs serving human capital owners rather than accumulating for themselves. Why would any of that hold? He quotes Eliezer Yudkowsky: “What is with this huge, bizarre, and unflagged presumption that property rights, as assigned by human legal systems, are inviolable laws of physics? That ASIs remotely care?” In Mowshowitz’s and Yudkowsky’s world of artificial superintelligences, the question isn’t which humans own the capital, it’s whether human ownership means anything at all.
But the variable that perhaps matters most may be what we collectively decide to value in the future. The scarcity frontier will move to what AI cannot provide, which may be something we have yet to understand. Whether the correction of inequality can persist may depend more on our consumption choices than on the automation of production. The question isn’t just what AI will replace. It’s what we decide to want.
Takeaways: Trammell and Patel’s essay is worth reading because it asks the bigger questions and explores the more radical models most AI commentary avoids. But its title is misleading. It’s not about capital in the 22nd century, capital concentration dynamics are operating now. The practical question for 2026 is how we build for human-machine complementarity, not substitution, and create new scarcities where human contribution remains vital.