From Platform Capitalism to What?

Artificial intelligence is being managed as an automation story. It is better understood as a change in what the economy treats as scarce — and therefore in who holds economic power.

For two centuries the input most central to production was human labour, an asset distributed — however unequally — across the entire population. AI is shifting the centre of gravity toward three assets that are not so distributed: compute, models and data. These can be owned, fenced and concentrated in a way that human capability never could. That single shift, more than any headline about job losses, is what makes AI a question about the architecture of the economic system rather than merely its efficiency.

The evidence already points one way. Private AI investment is extraordinarily concentrated. The four largest US hyperscalers alone are on course to spend on the order of US$700bn on AI infrastructure in 2026 — more than five hundred times India's entire national AI mission. The IMF finds that while AI may raise output, capital and wealth inequality “always increase” in its models, because the returns accrue to the owners of the machines.³ The productivity prize is real but contested: sober estimates put the gain near 1% of GDP over a decade, against consultancy forecasts an order of magnitude larger.²

Yet nothing here is technologically determined. The same models are being absorbed into three visibly different institutional settings — a liberal, capital-led United States; a rights-and-regulation European Union; a state-coordinated, infrastructure-first India — and they are producing three different political economies. That is the paper's central finding: AI sets the technological possibilities; institutions, ownership and political choice decide which economy we actually get.

We therefore urge decision-makers to stop treating AI as a procurement question and start treating ownership as the strategic and policy object: who owns the compute, the models, the data, and the surplus they generate. The most robust destination is not a single system but a deliberate mosaic — competitive markets where they work, public and commons infrastructure where concentration is dangerous, and a redistribution of AI's rents broad enough to keep the social settlement intact.

2.1The situation, as we comfortably understand it

The public conversation about artificial intelligence has organised itself around a single, strangely comforting question: will AI take our jobs? It is comforting because it is familiar. We have asked it of every general-purpose technology since the spinning jenny, and the historical answer has always eventually been reassuring: old work disappears, new work appears, and after a painful interval the economy employs more people at higher wages than before.

On this reading, AI is the latest in a lineage — steam, electricity, the assembly line, the computer, the internet — that raised output per worker and, in time, living standards. Ownership does not enter the story, because in the industrial model the crucial productive asset — human skill and effort — was held, in some measure, by everyone. Capital mattered, but labour retained bargaining power precisely because production could not happen without it. The twentieth-century social settlement was built on that mutual dependence.

This is a coherent picture, and parts of it may well prove right. We do not assume it is wrong. But it rests on an assumption that AI is beginning to strain: that the technology augments human labour more than it replaces the need to bargain with it.

2.2The complication

Three specific shifts turn a familiar productivity story into something the old narrative cannot contain. Each concerns not how many jobs move, but who owns what production increasingly depends on.

One — the scarce input is changing

Every economic order is organised around whatever is scarce and essential to production. For most of the industrial era that was labour, and later, particular skills. AI moves the centre of gravity toward a different triad: compute (the physical infrastructure of chips, data centres and power), models (the trained systems themselves), and data (the raw material from which capability is distilled). Unlike labour, none of these is naturally dispersed across the population. All three can be accumulated, fenced behind capital and intellectual-property barriers, and concentrated.

Two — ownership is concentrating at every layer at once

This is not a forecast; it is already measurable. The point is not that large firms lead — they usually do at a technology's frontier — but that AI concentrates simultaneously across compute, data, models and distribution, and that each layer reinforces the others. The economist's usual consolation, that high returns invite competitors who compete them away, runs into network effects, data feedback loops and capital requirements that a challenger cannot meet.

Three — the countervailing response has not yet arrived

Karl Polanyi observed that societies subjected to rapid, unmanaged market expansion generate a “double movement”: a self-protective counter-push of regulation, institutions and social claims.⁹ That counter-movement is what turned the brutal early factory system into the managed capitalism of the post-war decades. With AI, the market is moving at the speed of software and the counter-movement at the speed of legislation — and the two are badly out of phase.

That is the sharp edge on which this section ends, and it is deliberately left unresolved. If the assets that increasingly drive production are ownable and concentrable in a way that human capability was not, then AI is not principally a labour-market event. It is a question about the ownership base of the economy itself — about who receives the surplus when the machines do more of the work, and who is left with a claim on it.

3.1The centre of economic gravity is moving from labour to compute and data

The clearest signal is where capital is going. Investment is not flowing into wages or broad-based skills; it is flowing into physical AI infrastructure — chips, data centres, power — and into the proprietary models and datasets that sit on top of it. This is intangible-heavy, capital-intensive accumulation: its returns are captured by asset owners, and its barriers to entry rise with every training run. The task-based labour economics of Acemoglu, Autor and Restrepo is instructive here: technology raises welfare broadly only when it also creates new tasks for labour, not merely when it automates old ones.²

3.2The productivity prize is real, genuinely uncertain, and unevenly captured

Two disciplined conclusions survive this disagreement. First, the credible near-term macro-effect is modest: Acemoglu's estimate implies AI adds well under a tenth of a percentage point to annual productivity growth over the coming decade, because only about a fifth of tasks are exposed, of which under a quarter are today profitably automatable. The transformative case rests on capabilities and diffusion that have not yet happened. Second — and this is the finding that matters for system design — even the optimistic models redistribute upward.

The mechanism is not mysterious. AI raises the productivity of capital and of a slice of highly complementary labour; it does little for, or competes directly with, everyone else. Growth and inequality move together. A system that relied on the labour market alone to distribute those gains would, on this evidence, distribute them upward.

Is this just the last technology wave, again?

None of this proves the pessimists right; earlier waves also looked uniquely threatening in their day, and eventually enlarged employment. But the honest position is that AI's distinctive properties all push in the same direction — toward the owners of compute, models and data — which is why the historical optimism cannot simply be assumed.

3.3Ownership is concentrating at every layer at once

3.4Labour bifurcates; it does not simply vanish

Exposure to AI is not the same as replacement by it. The IMF's central contribution is to divide exposed work into two categories: jobs where AI mostly complements the worker, and jobs where it mostly substitutes for them. The balance between the two is not fixed by the technology; it is shaped by how firms deploy it, how workers are represented, and how skills are distributed.

The paradox: the richest economies face the most disruption and hold the most tools to manage it; the poorest face less disruption but have the fewest.

The Global South is not exposed — it is embedded, at the bottom of the value chain

The exposure figures understate the position of developing economies, because they measure the risk of AI displacing local work while missing the work AI creates there — and how it is priced. Behind every frontier model lies a hidden workforce of data annotators, content moderators and reinforcement labellers, disproportionately located in India, Kenya, the Philippines and elsewhere, performing the cognitively taxing and sometimes traumatic work of making raw data usable. This is the mechanism Couldry and Mejias call data colonialism: value is extracted from the periphery as cheaply as possible and captured at the core.¹⁰ That Kenya moved in 2026 to draft minimum-pay and mental-health protections for AI data workers is itself evidence of how far the extractive default had run.

3.5The same technology is building three different economies

If AI determined outcomes, the US, the EU and India would be converging. They are not. Each is absorbing the same models through a different institutional logic, and those logics are producing recognisably different political economies. This is the empirical heart of the paper's thesis.

None of these labels is fixed or flattering in only one direction. The US model builds capability fastest and distributes it least; the EU model protects citizens best and risks regulating an industry it does not own; the Indian model has the boldest public-infrastructure instinct and the thinnest fiscal and institutional capacity to carry it, atop a workforce that is roughly 90% informal and therefore largely outside the reach of the very protections a transition would require.

3.6The ecological bill is arriving, and it is not small

Ecological economics insists that no account of an economic system is complete without its material throughput. AI's energy, water and hardware demands are large, rising, and concentrated in the same few hands that own the compute — which means the ecological question and the ownership question are the same question. Any future system that ignores the physical bill is not a serious proposal.

If one idea organises this paper, it is that “access to AI” and “control over AI” are not the same thing. Millions can access a model through an interface while a handful decide what it does, what it costs, what it refuses, and who profits.

1. Who owns the compute — the chips, data centres and power?
2. Who owns the models — the trained weights themselves?
3. Who owns the data the models are built from?
4. Who owns the infrastructure of distribution?
5. Who controls access — and can revoke it?
6. Who sets the rules of use?
7. Who receives the economic surplus?
8. Who bears the risks when it fails?
9. Who can challenge a decision the system makes?
10. Who can exit — and at what cost?

Ask these of today's frontier AI and the answers cluster around the same small set of firms at almost every line. That clustering — not the capability of any single model — is the political-economy problem. The remedy is not to ban private AI but to vary the ownership form by layer, matching each layer to the arrangement that best guards against capture.

This is the intellectual foundation of the mosaic recommended later: no single ownership form is right for every layer, and the design task is to place each where its strengths apply and its failure modes are contained. The distributive question — who receives the surplus — then follows the ownership question, rather than being left to redistribution after the fact.

AI is not merely a national development question; it is a global one, and the global picture sharpens every concern raised so far. Productive capability is not democratising across the world — it is concentrating, geographically and institutionally, along a small number of chokepoints that a few states and firms control.

The result is a world in which the US–China rivalry sets the pace, the European Union writes the rules it hopes others will adopt — the “Brussels effect” — while depending on others' compute, and most nations, India included, must choose between renting capability, building sovereign alternatives at a fraction of the frontier's scale, or negotiating access on terms they do not set. On current trends AI is concentrating productive capability, not spreading it.

Without deliberate ownership strategy, AI could reproduce the classic dependency — exporting cheap cognitive labour, importing finished intelligence. India therefore holds the whole paper's thesis in miniature. The technology is available to it. Whether it leapfrogs or is locked into a new hierarchy will be decided by institutions and ownership: data governance, public compute, the formalisation and protection of labour, and a strategic choice to own the data and application layers rather than merely to supply the labour beneath someone else's model.

The evidence in Section Three does not point to a single destination. It points to a fork. Against those three filters we set out five candidate systems — from the market's default trajectory to its most radical alternative. They are not a menu of equally likely outcomes but the corners of the space within which the real future will be assembled, almost certainly as a blend.

4.1Scoring the futures against one framework

A candidate system cannot be judged on growth alone. We score all five against the same eight criteria — and we are explicit that this is an assessment, combining evidence with stated values, not a measurement. The values are named; readers may re-weight them.

Three conclusions follow. First, the trade-off between desirability and feasibility is the real constraint: the most attractive pure system (D) is the least buildable, and the most buildable (A) is the least attractive. Second, feasibility is not fixed — it can be widened by exactly the ownership and institutional moves Section Five sets out. Third, the balance tips toward Model E, not because pluralism is tidy, but because it is the only candidate that can be assembled incrementally, absorb shocks, and keep several values in tension without sacrificing any entirely. The mosaic wins less on elegance than on robustness under uncertainty.

4.2Four scenarios for 2035–2050

Which future arrives depends on two variables above all others — the two this paper has argued are decisive. How concentrated is the ownership of AI and data? And how strong are the countervailing institutions? Cross them, and the possibility space resolves into four scenarios. These are not forecasts. They are a map of where the mechanisms lead.

The choice between AI-era economies is not a choice between good and bad. It is a set of genuine tensions between things a society legitimately wants at once. Naming them is the discipline that keeps the recommendations honest: every position that follows resolves toward the right-hand column, and pays a stated price on the left.

No system resolves all of these in one direction without loss; that is why Section Four found no dominant candidate, and why the mosaic — which holds several of these tensions in deliberate balance rather than collapsing them — scored best. A strategy that pretends the tensions away is not a strategy; it is a preference disguised as a fact.

5.1Make ownership — not access — the object of strategy and policy

Resolves the core complication: the scarce input is now ownable and concentrating. The single most consequential reframing in this paper is to stop asking “how do we get access to AI?” and start asking “who owns the compute, the models, the data, and the surplus?” For policymakers, that means competition policy aimed at the compute and model layers, including interoperability and, where warranted, structural separation — treating foundational AI as one treats other critical infrastructure. For boards, it means recognising that durable value sits in proprietary data and deployment, not in renting a model everyone else can rent. For investors, it means reading concentration as both the source of today's returns and tomorrow's principal political risk. Highest impact, moderate feasibility — and the precondition for everything below.

5.2Deploy for augmentation, and design at the level of tasks

Resolves §3.4: the same exposure can augment or displace — deployment decides which. Whether AI complements or replaces a worker is a design choice, not a property of the model. Organisations that redesign work at the task level — automating the routine, reinvesting the freed capacity into higher-value human tasks, and giving workers a voice in the redesign — capture more value and distribute it better than those pursuing thin headcount substitution. This is the rare recommendation where the commercial and the social case coincide. Reskilling should be tied explicitly to complementarity, not offered as a generic gesture.

5.3Redistribute the AI surplus deliberately — through services and ownership, not transfers alone

Resolves §3.2: even optimistic models push the gains upward. If the surplus accrues to capital, the remedy must reach capital. We favour two instruments over a headline cash transfer: broadened ownership — sovereign and citizen wealth funds, employee ownership, cooperative equity — so citizens hold a claim on the returns to AI capital; and universal basic services — health, education, care, connectivity and a public AI layer — which decouple survival from wage-labour more cheaply and durably than income transfers alone. A modest, well-designed levy on AI-driven productivity or on excess compute rents can fund both. High impact, lower feasibility, because it redistributes power and not merely money.

5.4Build public and commons infrastructure where concentration is most dangerous

Resolves §3.3 and the ecological constraint of §3.6. Not everything should be a market, and not everything should be a state. The mosaic's practical content is to place public compute, data commons and data trusts, and interoperability and portability mandates at exactly the layers where private concentration threatens competition, rights or resilience. Public and cooperative infrastructure must be genuinely resourced, or it becomes Scenario III — dispersed in name, captured in fact.

5.5Price in the physical and geopolitical bill — and move up the value chain

Resolves §3.5 and §3.6. Energy, water and hardware are first-order constraints; compute is a geopolitical asset. Governments should treat energy and compute sovereignty as industrial-strategy questions, and firms should treat the physical bill as a real cost of AI rather than an externality. For developing economies in particular, the strategic imperative is to move up the value chain — from supplying cheap data-labour to owning data, applications and deployment. Own the application and the data; rent the frontier model.

6.1Risk mitigation — the objections, pre-empted

None of this requires a single decisive political victory — which is the mosaic's central advantage. It can be begun in fragments, in different sectors and countries, by different actors, and still compound. The first move — the ownership audit and the interoperability default — is available to a regulator, and the task-level augmentation review is available to a board, on Monday.

We began with a comfortable question — will AI take our jobs? — and argued that it conceals a harder one. The evidence assembled here does not show a technology destined to enrich or impoverish us. It shows a technology that shifts the scarce input of the economy from labour, which nearly everyone holds a little of, to compute, models and data, which almost no one holds — and which are concentrating, measurably, at every layer at once.

That shift does not settle our future. It is precisely because AI is so powerful that the institutions around it matter so much: the same models are already building a liberal-capital economy in the United States, a rights-and-regulation economy in Europe, and a state-infrastructure economy in India. Technology set none of those outcomes. Ownership, institutions and political choice did.

The disagreement among serious economists about AI's aggregate effect — from around one per cent of GDP to seven — is not a reason to wait for certainty. It is a reason to build institutions robust across the whole range, because we will not know which world we are in until we are already living in it.

The future economic system is open, not predetermined. AI has widened what is technically possible; existing institutions shape what is economically likely; political power decides what is socially feasible; and democratic choice determines what is normatively desirable. Between those four lies the work — and the opportunity — of the coming decade. The organisations that understand AI as a question of ownership rather than access, and act while the answer is still open, will not merely adapt to the next economy. They will help decide which one we get.