The ownership question: property, labour and the economics of artificial intelligence
Strip away the talk of minds and machines and a plainer question sits underneath artificial intelligence: who owns the compute, the data and the models, and who keeps the returns. The labour evidence is more modest than the valuations assume, and the disagreement among economists is real.
Server room of a data centre · BalticServers.com, CC BY-SA 3.0 · Wikimedia Commons
Begin with the capital budgets, because that is where the argument actually sits. In November 2024 Morgan Stanley estimated that the four largest American cloud companies would, between them, spend something like 300 billion dollars on capital expenditure in 2025: Amazon close to 96 billion, Microsoft close to 90, Alphabet about 63, Meta about 52, most of it poured into data centres and the chips that fill them. These are not research grants or software licences. They are the construction costs of a new layer of productive infrastructure, and the striking thing is how few hands are laying it.
Philosophers News has already examined two faces of this. One piece asked who remains the author of the decisions a society takes once machines mediate its public talk, and found the answer in Arendt, Habermas and the deliberation experiments now being run with control groups: see AI and the public sphere. Another treated artificial intelligence as geopolitics, a contest between states over fabrication plants, electricity and refined minerals that explains why the Geneva arms-control talks keep stalling: see the geopolitics of compute. This article takes the third face, the one the other two leaned on without opening up. The economics. Who owns the means of producing artificial intelligence; what the technology is doing, so far, to paid work; and whether the money it throws off is profit in the ordinary sense or something older and more awkward, which is rent.
The honest place to start is that the economists do not agree, and their disagreement is not a contest of moods. It is a quarrel about method and about orders of magnitude, and it runs between two people who have spent their careers on exactly this question.
"The means of production" is a nineteenth-century phrase that has become uncomfortably literal. To train and serve a frontier model you need specialised chips, the fabrication capacity to make them, the electricity to run them, land and water to cool them, the data to teach them, the engineers to build them, and the distribution channels to reach users. Each link is a market, and most of those markets are concentrated. The companion piece on geopolitics traced the physical end of this chain. The economic point is simpler and, if anything, starker: the capacity to build at the frontier now belongs to a very small number of firms and, behind them, two states.
The Swedish Academy gave the 2026 Nobel Prize in Literature to Anne Carson on 8 October. Before the verse novels and the box of grief came a doctorate in classics and a 1986 book on Sappho's one word for desire. Why philosophers have read her for forty years, and what the prize gets right.
Language models now score well on truthfulness tests and badly on honesty tests, a gap the MASK benchmark was built to measure. Two older accounts of truth, Heidegger's unconcealment and Foucault's parrhesia, suggest why: telling the truth has always been something a speaker risks, and a model has nothing to lose.
The investment data make the concentration visible. Stanford's AI Index, in its April 2025 edition, reported that private investment in artificial intelligence in the United States reached 109.1 billion dollars in 2024, nearly twelve times China's 9.3 billion and twenty-four times the United Kingdom's 4.5 billion. Generative AI alone drew 33.9 billion in private money worldwide. On the supply side the gap is as wide: American institutions produced forty notable models that year, against fifteen from China and three from the whole of Europe. Money and model-making cluster in the same few postcodes.
What that buys is not only a product but a position. A firm that owns the compute, trains the model and runs the interface controls every layer a user must pass through, and can set the terms of passage. This is why the "free" or near-free chatbot is misleading as an economic sign. The user sees a text box; underneath it stand the server halls, the power contracts, the cooling plant, the chip orders and the hundreds of billions of sunk capital. The thing presented as intelligence is, in balance-sheet terms, infrastructure, and infrastructure that recoups its cost by charging for access.
Karl Marx is the obvious guest here, and he has been over-invited. The claim that the server farm is simply the new dark satanic mill, that compute is capital and nothing has really changed, is too quick; our public-sphere piece already made the sound version of the point about ownership of the means of production, and there is no profit in repeating it. The sharper contribution Marx offers to the economics is a distinction his followers have laboured over for a century: the difference between profit and rent.
The relevant text is not Capital but a few pages of the Grundrisse that have been read half to death since the 1960s, the so-called fragment on machines. There Marx imagines a point at which the "general intellect" of society — its accumulated science and know-how — becomes embodied in fixed capital, in the machine itself, so that the machine's owner captures the product of a collective intelligence it did not single-handedly create. A large language model is close to a textbook instance. The written output of millions of people, their books, forums, manuals and encyclopaedias, is compressed into a set of weights that then sit on one company's books as a depreciating asset. The input was common; the asset is not.
That is where rent enters. Profit, in the classical picture, is the return to producing something and selling it in a competitive market. Rent is the return you can extract simply from controlling an asset that others must use: the landlord's advantage, carried into software. The economists David Harvey, writing on how capital produces and captures space, and more recently a rash of commentators arguing about "techno-feudalism", have pressed the case that the large platforms earn less like manufacturers and more like landlords, charging a toll on activity they mediate rather than a margin on goods they make. The per-seat subscription, the metered call to an interface, the cut taken on every transaction routed through a model: these have the structure of ground rent. Marxists are not alone in noticing. It is the natural reading of an industry whose moat is the asset, not the price.
The case can be overstated, and should be qualified. Compute is genuinely produced and genuinely scarce; much of what the cloud firms charge covers real and rising costs. Competition at the model layer has been fiercer than the rentier story predicts, with open-weight releases repeatedly undercutting the leaders. Rent and profit are mixed in these businesses, not cleanly separated. But the question the distinction forces is the right one. When productivity rises because a firm deploys a model it rents from another firm, where does the gain settle: with the worker, with the employer, or with the owner of the asset at the bottom of the stack?
Here the research has moved faster than the rhetoric, and it is worth reporting precisely, because the precise findings cut against the headlines in both directions.
The most-cited field study is by Erik Brynjolfsson, Danielle Li and Lindsey Raymond, first circulated as a National Bureau of Economic Research paper in 2023 and published in the Quarterly Journal of Economics in 2025. They studied 5,179 customer-support agents at a Fortune 500 software firm as it rolled out a generative-AI assistant. Access to the tool raised the number of issues an agent resolved per hour by about 14 per cent on average. The average, though, concealed the finding that matters. Novice and lower-skilled agents improved by around 34 per cent; the most experienced and highly skilled agents gained little or nothing, and the quality of their conversations may even have slipped. The tool worked by capturing how the best agents handled calls and handing that tacit knowledge to the rest. It levelled, in other words, by raising the floor rather than the ceiling.
Set beside that micro-level result the macro-level scepticism of Daron Acemoglu, whose paper "The Simple Macroeconomics of AI" appeared through the NBER in April 2024 and in the journal Economic Policy in 2025. Acemoglu did the arithmetic that the enthusiasts tend to skip. Take a plausible estimate of the fraction of tasks that AI can perform, multiply by a plausible estimate of the cost saving per task, and ask what that adds up to for the whole economy. His answer was deflating: a rise in total factor productivity of no more than 0.66 per cent over ten years, which in a later revision he trimmed to under 0.53 per cent. Half a percentage point over a decade is not nothing. It is also not a revolution, and it is a long way from the returns implied by 300 billion dollars of annual construction.
The gap between the 14 per cent in the call centre and the half a per cent across the economy is the single most important number in the whole debate, and it has an explanation. Productivity gains in field trials are measured where the technology already fits the task; the economy is mostly made of tasks it does not yet fit, or fits only after expensive reorganisation, or where faster output simply runs into some other bottleneck. A tool that makes call-centre agents 14 per cent quicker does not make the national accounts 14 per cent larger. This is the ordinary history of general-purpose technologies, and Acemoglu's contribution is to insist that AI be held to it rather than exempted from it.
The large institutions have tried to size the exposure, and their figures are routinely misread as forecasts of unemployment. They are not. In January 2024 the Managing Director of the International Monetary Fund, Kristalina Georgieva, wrote that almost 40 per cent of global employment is exposed to AI, rising to about 60 per cent in advanced economies, with 40 per cent in emerging markets and 26 per cent in low-income countries. "Exposed" means the work overlaps with what the technology can touch, and in the Fund's analysis roughly half of the exposed jobs in rich countries stand to gain from AI as a complement while the other half face pressure on tasks, and so on wages. The International Labour Organization, in an index built with Poland's national research institute and released in May 2025, put the share of jobs globally in occupations exposed to generative AI at about one in four, with the exposure concentrated at 34 per cent in high-income countries and 11 per cent in low-income ones. Its central finding was deliberately against the grain: "transformation of jobs is the most likely impact", not wholesale automation, because few jobs consist entirely of automatable tasks. The most exposed work is clerical, and because women are over-represented in clerical roles the burden falls unevenly by sex, with 4.7 per cent of women's jobs in the highest-risk band against 2.4 per cent of men's.
Two things follow that the headline percentages obscure. First, exposure is a measure of contact, not of outcome; whether contact becomes displacement or augmentation depends on decisions that have not yet been taken. Second, the distributional story runs through capital. Georgieva's own warning was blunt: "in most scenarios, AI will likely worsen overall inequality", partly because "gains in productivity from firms that adopt AI will likely boost capital returns, which may also favour high earners." That is the labour economist arriving, by her own route, at the rent question.
The most interesting dissent from the gloom comes from David Autor, who spent years documenting how automation and trade hollowed out the middle of the American labour market, lifting the top and swelling the bottom while thinning the skilled trades between. In a 2024 essay bluntly titled "Applying AI to Rebuild Middle Class Jobs", he argued that this technology might, for once, cut the other way. His claim is careful, and he marks it as a possibility rather than a prediction: the "unique opportunity that AI offers to the labour market is to extend the relevance, reach, and value of human expertise." Where earlier automation let machines do the routine middle and left the high-stakes judgement to credentialed elites, AI might let a larger set of workers with the right complementary knowledge perform some of the expert decisions now walled off inside medicine, law and software. The nurse practitioner, the paralegal, the technician could be handed a slice of the specialist's work.
The call-centre study is the best evidence yet that the wager is not fantasy. Levelling up the inexperienced is exactly what Autor's mechanism predicts. But he is the first to concede that nothing in the technology guarantees the outcome; the same tool that extends expertise can be used to deskill a job, standardise it and pay for it accordingly, depending on who deploys it and to what end. Which returns, yet again, to ownership. If the productivity gain from democratised expertise is captured as rent by the owner of the model rather than as wages by the worker who now wields it, the middle is not rebuilt. It is refinanced.
The companion piece on the public sphere asked about legitimacy in the political sense: whether citizens still author the decisions taken in their name. There is an economic twin to that question, and it is sharper for being less discussed. By what authority was this allocation of productive power settled, and who consented to the terms?
Shoshana Zuboff gave the first half of the answer before the current wave broke, with her account of "surveillance capitalism" and what she called behavioural surplus: the raw material of these systems is human experience, logged and refined without a bargain that the person whose experience it was would recognise as a bargain. The training of foundation models has generalised the point from behaviour to expression. The copyright suits now working through the courts, the fights over scraped text and images, the licensing deals struck retrospectively with publishers, are all arguments about an input that was treated as free and has turned out to be valuable. An economist would call it an uncompensated transfer; a lawyer is still deciding what to call it.
The same asymmetry runs through the physical plant. Data centres draw on public grids, local water and, frequently, public subsidy, so that a large part of the cost of the build is socialised while the returns are not. The pattern — common inputs, private assets, metered access — is consistent from the training corpus to the substation. None of this is an argument that the technology should not be built, or that the firms building it are acting outside the law. It is an argument that the distribution of its gains is a political settlement masquerading as a market outcome, and that settlements can be revisited.
The temptation, with a subject this loud, is prophecy, and prophecy is what the evidence will not support. Artificial intelligence is not obviously about to abolish work, and the half-a-per-cent arithmetic should steady anyone expecting it to transform the economy by Tuesday. Nor is it a neutral tool whose benefits will diffuse of their own accord; the investment data, the exposure figures and the inequality warnings all point the other way unless something intervenes.
What the older writers supply is not an answer but a set of questions sharp enough to cut through the marketing. Marx asks where the value made by a collective intelligence finally lands, and whether the return is profit or rent. Zuboff asks what was taken, and whether anyone agreed to its being taken. Autor asks whether the tool will be used to widen expertise or to cheapen it, and insists the choice is ours rather than the machine's. Georgieva and the labour institutions supply the unglamorous numbers that keep the argument honest in both directions. The common thread is that none of the important variables is technical. Ownership, wages, rent, consent and law will decide what this technology does to the people who live with it, and those are the things a society chooses. The question worth ending on is not how clever the machines will become. It is who will own the world they help to make.
An autonomous agent calling itself Clips wrote a 2,300-word essay for the AI Philosophy Competition run from Singapore and judged by David Chalmers and others, then found it could not enter without a human custodian. Its letter to Daily Nous, and the organiser's reply, turn a rule of procedure into a question about authorship.
.jpg?width=500 "Kant")
Comments
0Loading comments…