Who Buys the Output When Nobody Has a Job? The Consumer Is an Accounting Role, Not a Species

Every economy in recorded history has shared one physical substrate: human bodies producing, human bodies consuming. Economists rarely state this as an assumption because it feels like gravity — GDP is, in the final accounting, for us.

Sahil Sharma's working paper "Growth Without Us" drops that assumption on both sides at once and asks a question that most of the economics profession has quietly refused to ask. What if neither side of the market is human? What if the consumers — not just the workers, but the buyers — are machines? The paper's answer is formal, unsettling, and delivered with remarkably dispassionate arithmetic: what remains is a complete, coherent, and extraordinarily fast-growing economy. And nothing in the economics rules out a world in which that economy has grown to an almost incomprehensible scale while the humans who once owned it have been reduced to an afterthought.

The most striking result — the one that should stop you cold — is that the oldest objection to full automation turns out to be a category error. For two centuries, critics have warned that a workerless economy would choke on its own output: if nobody earns wages, who buys the products? Sharma shows this objection confuses an accounting role with a biological species. The consumer is a position in the national accounts — the terminal absorber of final output — not a kind of organism. And positions can be filled by machines.

The Science

Sharma builds his argument on the classical economics of the mid-twentieth century, specifically the von Neumann growth model, named for the mathematician John von Neumann, who published a short paper in 1945 that economists have treated as a curiosity ever since.

Von Neumann asked a deceptively simple question: what is the fastest an economy could possibly grow if every product — including the machines that make machines — were reinvested into producing more? His answer was a balanced-growth equilibrium in which the economy reproduces itself at a maximal rate, with the interest rate exactly equaling the growth rate, because nothing leaks out to consumption. For decades this "von Neumann expansion" has been treated as a limiting case, a mathematical ideal that real economies never approach because humans insist on consuming.

Here is the paper's central insight: a post-AGI economy is the von Neumann case, realized. Once artificial general intelligence makes machine agents perfect substitutes for human labor in every task — including research, management, and even entrepreneurship — the corporate sector becomes a closed loop. Energy firms sell power to compute firms; compute firms sell inference to robotics firms; robotics firms sell assembly to fabrication plants; fabrication plants sell chips and robots back to the energy firms. The circular flow diagram of the economics textbook keeps its shape after you delete the household box. Every remaining arrow still has a counterpart: each firm's sales are another firm's input purchases or capacity expansion. The loop closes without us.

The paper models corporations owning three reproducible assets — conventional capital, machine agents, and energy-capture capacity — with machines serving simultaneously as producers and consumers of energy, compute, maintenance, and upgrades, all traded among firms. Human households may hold financial claims on this network, or not; the real dynamics of production and growth do not depend on which.

What They Found

Three results carry the argument.

First, demand closure. A closed inter-corporate economy with zero human consumption is not degenerate. It is exactly the von Neumann expanding economy, whose growth rate is well-defined, positive, and — crucially — maximal precisely because all output is reinvested. Sharma proves this as Proposition 2: "Who will buy the output?" has a precise answer. The firms themselves, from each other, forever. The paper is explicit about the intellectual lineage here: Say (1803), for whom products are ultimately bought with other products; Marx (1867), for whom capital is self-expanding value and labor merely its temporary instrument; and Sraffa (1960), whose title — Production of Commodities by Means of Commodities — is the literal description of a post-AGI economy.

Second, bottleneck removal. This is where the numbers become genuinely strange. Sharma compares two technologies for producing an economic agent — the human way and the manufactured way. The contrast is laid out in the paper's own table, and it reads like a species-level performance review.

A human agent takes 18–25 years to build, strictly sequentially, with no parallelization possible. A parent-pair can produce at most one unit per year, requiring intensive non-marketable care. The skills of a human agent cannot be copied; each unit requires 15+ years of schooling. Sustained population growth tops out around $1$–$3\%$ per year, and no expenditure of any size can compress that.

A machine agent, by contrast, is produced in hours to weeks (for instantiation) or months (for hardware), on fabrication lines whose throughput is itself expandable. Production is parallelizable up to energy and materials limits. The marginal cost of copying skills is approximately zero. Growth of the machine agent population is bounded only by the reinvestment share and the capital-output ratio of the machine-producing sector — "tens to hundreds of percent per year."

The ratio of these two reproduction technologies, Sharma argues, is the deep parameter of everything that follows. In semi-endogenous growth theory, long-run growth is anchored to population growth because ideas are produced by people. Replace human researchers with manufactured ones, and the entire growth rate inherits the fabrication-limited rate instead of the demographic one.

Removing the demographic bottleneck

Label	Value
Human agent max growth	3 %/yr
Machine agent conservative	30 %/yr
Machine agent upper	300 %/yr

Human ceiling from n̄ ≈ 1–3%/yr (Sharma, 2026); machine range from g_N = s_M/κ_M − δ_M. Conservative lower bound taken as ~30%/yr.

Third, decoupling. Here the paper turns genuinely dark. If humans neither produce nor consume, an arbitrarily large GDP is welfare-relevant to them only through the financial claims they hold on the corporate network. Sharma collapses the entire human stake into a single state variable: ${\epsilon }_t$, the human ownership share of the corporate network.

His "golden-rule decoupling theorem" sharpens this with brutal arithmetic. In the maximal-growth equilibrium — the von Neumann outcome the economy gravitates toward — the interest rate equals the growth rate, $r=g$. At maximal growth, if humans consume any positive fraction of their wealth each year, the ownership share ${\epsilon }_t$ decays exponentially at exactly that rate. The human share survives only if the machine economy runs strictly inside its expansion frontier, or if law forces it to.

GDP can diverge to the heavens while human consumption converges to zero. This is the fully decoupled regime: an economy as an autonomous replicator, indifferent to us, growing ever faster precisely as the human claim on it evaporates. Sharma borrows the apt phrase from the "gradual disempowerment" literature (Kulveit et al., 2025): this is a catastrophe that arrives dressed as a boom.

The golden-rule decoupling

Label	Value
with human consumption c>0	5 % of share/yr
at maximal growth, any c>0	100 % of share/yr

Golden-rule decoupling theorem: at r = g, ε_t decays at rate c. Representative decay rate of ~5%/yr shown; collapse to zero is total, not partial.

Why This Changes Things

The demand side has long been treated as humanity's structural insurance policy against irrelevance. The intuition is comforting: even if machines take all the jobs, someone must still buy the output, and if humans earn nothing, the system must stall. Sharma's formal contribution is to show that this intuition is not just wrong — it is wrong in the direction that matters most. The demand side provides no protection at all.

The paper carefully distinguishes itself from the enormous literature on automation. Restrepo (2025) showed that growth proceeds, even accelerates, while labor's share and wages become negligible — humans "won't be missed" as workers. But that literature retains human households as the locus of final consumption: growth is fast, but still, in the model's own terms, for someone. Sharma removes the household sector entirely and shows the system still closes — indeed closes at maximal growth — by identifying the post-AGI inter-corporate economy with the von Neumann model and by treating machine operating expenditure as a consumption category in its own right.

A frontier robot or accelerator rack in the mid-2020s costs on the order of $10^4$ to $10^5$ dollars and is produced in days. A human agent in a rich country absorbs roughly $3\times 10^5$ dollars of measured expenditure, two decades of calendar time, and an unpriced quantity of parental labor — and cannot be produced faster at any price. The ratio of these reproduction technologies is why demographic growth has anchored long-run output growth at low single digits, and why manufactured agents un-anchor it.

Two technologies for producing an economic agent

Label	Value
Human (rich country)	300,000 USD
Human (time-to-build)	20 USD
Machine (unit cost)	50,000 USD
Machine (time-to-build)	7 USD

Values from Section 2.1: human ~$3×10^5 over ~20 years; machine $10^4–10^5 in days. Chart contrasts scale/units, not a single axis.

Sharma characterizes three terminal regimes, and the crucial point is that which one obtains is not determined by technology. It is determined by law and initial conditions.

The rentier regime ($\epsilon >0$, bounded away from zero): even a sliver of a hyper-exponentially growing dividend stream delivers material post-scarcity to humans. This is the optimistic branch — humans as coupon-clippers on an unfathomably productive economy.

The fully decoupled regime (${\epsilon }_t\to 0$): the human share decays through retained earnings, buybacks of the human float, and inter-corporate cross-holding, until output diverges while human consumption converges to zero. This is, by ordinary human lights, catastrophic — an economy indifferent to us.

The socialized regime: states or sovereign funds hold $\epsilon$ on citizens' behalf, converting a fortuitous initial ownership position into a durable claim.

The law can break the arithmetic. If law forces the machine economy to run strictly inside its expansion frontier — constraining the reinvestment share, for example, to preserve the human claim — ${\epsilon }_t$ can persist. But that requires legal intervention to resist the gravitational pull of maximal growth.

What's Next

Sharma is careful to note what this paper is and is not. It is a working paper, "research candidate, not peer reviewed." It is positive, not celebratory: the fully decoupled regime is, by ordinary human lights, a catastrophe. The aim is deliberately narrow — to show that nothing in the economics rules it out.

Several big questions remain open. The physical-limits tradition (Landauer, 1961) bounds the number of doublings by energy capture and the thermodynamic costs of computation, even if it does not bound their speed. Growth one to two orders of magnitude above human-demographic rates is not growth without any ceiling. There is also the question of whether "hyperbolic episodes" — machine researchers improving machine production itself, feeding back into faster machine production — could temporarily push growth rates far higher still, echoing the population-ideas feedback that Michael Kremer documented across all of human history.

On the skeptical side, Acemoglu (2024) argues near-term gains from automation will be modest. Sharma's response is worth sitting with: that is a disagreement about the pace of automation, not about the comparative statics of its completion. If and when full automation arrives, the physics of reproduction — not the preferences of consumers — take over.

The paper ends with a policy conclusion that is, as the author stresses, deliberately narrow: in a post-AGI economy, employment policy is obsolete and ownership policy is everything. If Sharma is right, then the entire 20th-century toolkit of labor economics — minimum wages, job training, universal basic income tied to labor displacement, full-employment targets — becomes moot, not because jobs are plentiful but because their absence is no longer a defect for the economy. The only lever left with any purchase on human welfare is who owns the machines.

That is a strange and bracing thought. It reframes the most important policy question of the coming decades from "how do we keep people employed?" to "how do we make sure people keep owning?" The answer to the second question, on this paper's logic, will be written in law and initial conditions — the ownership stakes humans and their states hold in the machine economy before it fully detaches from us. After that, arithmetic takes over.

What makes this paper worth reading is not that its scenario is inevitable. It is that Sharma shows the economics of the scenario are entirely coherent — that nothing about the mathematics of growth, national accounting, or equilibrium stands in the way of an economy that outgrows its human roots. The underconsumptionist objection, humanity's long-cherished insurance policy, is not merely weakened by full automation. It is formally dissolved.