The Agonistic Economy of Digital Capital

Potlatch at Quatsino, British Columbia, between 1897 and 1908. Photograph by Benjamin William Leeson. Public-domain image via Wikimedia Commons (Leeson, ca. 1897–1908). This historical record should be read in its colonial context and not as a comprehensive representation of the diversity of potlatch ceremonies.

Note on sources and the scope of the comparison

This essay takes as its starting point Jon Hernández’s interview with Daron Acemoglu, published on 16 July 2026 on the YouTube channel Inteligencia Artificial (Hernández, 2026).

The comparison between the contemporary economy of artificial intelligence and the potlatch is an analytical analogy, not a historical or cultural equivalence. The potlatch cannot be reduced to irrational extravagance or the competitive destruction of wealth. For different Indigenous peoples of the Northwest Coast of North America, it has been, and remains, a political, legal, ceremonial and economic institution connected to redistribution, rank, memory, territory, names and the transmission of rights (Royal BC Museum, n.d.). The analogy proposed here is confined to a specific question. It examines how the display, circulation and, at times, deliberate loss of wealth produce prestige, obligations and hierarchies.

The reference to hau demands similar caution. Mauss drew on this Māori category to think about the force that keeps a gift in circulation. The interpretation of hau as the ‘spirit of the giver contained in the thing’ has been widely debated, including by Māori perspectives critical of the Eurocentric appropriation of the concept (Frank, 2016; Stewart, 2017). This essay does not transpose a Māori ontology into the digital world. It uses the Maussian controversy to ask how apparently free goods remain connected to those who provide them and what relations of dependence travel with them.

Artificial intelligence is often presented as a force that advances under its own momentum. In this narrative, models grow larger, their capabilities multiply, and society must adapt to a technological direction that has already been decided. Institutions arrive later, tasked with correcting harms, distributing compensation or containing risks. This sequence reverses the problem. From the outset, the direction of innovation is shaped by economic incentives, concentrations of power, cultural imaginaries and decisions about which problems deserve to be solved.

The decisive issue is who defines the productivity gain, appropriates its results and converts it into durable power. An economy can produce more with fewer jobs and call that reduction productivity. It can also use the same technology to expand workers’ capabilities, create new tasks and produce goods and services that were previously inaccessible. Both paths may improve certain measures of efficiency, but they build different societies.

Bringing together the institutional economics of Daron Acemoglu and James A. Robinson with the anthropology of the gift allows this dilemma to be formulated differently. Why Nations Fail argues that prosperity depends not only on the availability of resources or technologies, but on institutions that distribute opportunities and constrain extraction by elites (Acemoglu & Robinson, 2012). Mauss (1925), Codere (1950), Piddocke (1965), Kan (1989) and other traditions within economic anthropology show, in turn, that institutions extend beyond constitutions, contracts and property rights. They are also embodied in obligations to give, receive and reciprocate, in public ceremonies and in collective mechanisms that convert wealth into legitimate prestige.

A thesis emerges from this intersection. The economy of artificial intelligence is increasingly assuming an agonistic form. A small number of corporations compete through the display and sacrifice of capital, distribute subsidised technological access, and convert that apparent generosity into infrastructural dependence and political authority. Unlike the potlatch, where prestige was tied to public obligations of circulation, reciprocity and recognition, the digital regime allows legitimacy to accumulate without an equivalent social return. The institutional challenge of artificial intelligence is therefore to transform technological power into enforceable obligations of redistribution, openness, participation and the expansion of human capabilities.

The Productivity Mirage

One of Acemoglu’s central contributions to the debate is to separate visible gains in specific tasks from their macroeconomic effects. A person may write faster, summarise documents, generate code or automate part of a customer-service process. A company may save several hours on a particular workflow. None of this guarantees that aggregate output will rise in the same proportion.

The diffusion of a general-purpose technology requires firms to be reorganised, teams to be trained, systems to be adapted, mistakes to be absorbed and entire chains of work to be transformed. Benefits observed in controlled trials do not automatically transfer to occupations where context, responsibility, social interaction, tacit knowledge or the ability to respond to unforeseen situations matter.

In The Simple Macroeconomics of AI, Acemoglu (2025b) estimated that artificial intelligence might raise total factor productivity by less than 0.53 per cent over a decade—meaningful, but far removed from the most euphoric projections. In the July 2026 interview, he again placed the share of tasks that could be profitably automated in the near term at around 5 per cent (Hernández, 2026). These figures are estimates dependent on assumptions, not fixed predictions. Their value lies in reminding us that technical exposure to automation is not the same as automation that is profitable, reliable or socially desirable.

The distance between microeconomic performance and macroeconomic outcomes does not make the technology irrelevant. It shows that productivity is a social relationship before it is an isolated property of a tool. Time saved can be converted into new products, better services and learning. It can also be used to reduce headcount, intensify work or transfer to staff the obligation to produce more without equivalent improvements in pay and conditions. The tool does not determine which possibility prevails.

The work of Acemoglu, Autor and Johnson sharpens the distinction between automation and pro-worker artificial intelligence. Their typology identifies five directions of technological change (Acemoglu, Autor, et al., 2026). Technology may augment labour, augment capital, automate tasks, level expertise or create new tasks. Expertise-levelling enables less experienced people to perform functions once reserved for specialists, but it can also turn expertise into an abundant commodity and reduce its remuneration. Only the creation of new tasks is unambiguously pro-worker, because it generates additional demand for human capabilities rather than merely making an existing task cheaper.

The difficulty is that current incentives favour automation. A company can calculate with relative ease how much it will save by eliminating a task or leaving a vacancy unfilled. It is harder to appropriate privately the social benefits of training a person, improving the quality of a public service or creating widely distributed capabilities. When markets reward immediate cost reduction and ownership of closed platforms, technological development bends towards substitution even when other applications would be more socially productive.

Agentic artificial intelligence may intensify this tension. Here, agentic systems are understood as systems able to chain actions together, consult tools and pursue objectives with varying degrees of operational autonomy. Recent advances suggest an episodic evolution, made up of breakthroughs and periods of assimilation, rather than a constant and predictable exponential curve. This requires two simplifications to be avoided. The first denies the possibility of rapid advances because current systems remain fragile. The second assumes that such advances will imminently lead to a general intelligence capable of performing any human function.

Current models continue to display significant limitations in reliability, contextual understanding, social interaction, accountability and physical action. These limits are not necessarily eternal or insurmountable, but they remain unresolved structural problems. Even a substantial technical improvement would still have to pass through organisations, professional standards and legal responsibilities. The economic question includes what a model can do reliably, for whom, under what supervision and with whom bearing the consequences of error.

From National Institutions to the Technological Regime

In Why Nations Fail, Acemoglu and Robinson (2012) distinguish between inclusive and extractive institutions. Inclusive institutions distribute opportunities, limit arbitrary exercises of power and allow broad sections of society to participate in economic activity. Extractive institutions organise society so that resources and rents are transferred towards dominant groups. This distinction does not describe two pure states. It is a way of analysing processes, conflicts and feedback loops.

The distribution of economic resources produces de facto political power. Those who control capital, information, media or strategic infrastructures can influence the rules even without formally holding office. That influence helps create institutions that protect their position, producing further concentration and another expansion of their power. The cycle can be reversed where pluralism, social organisation and countervailing forces are capable of subjecting private wealth to public obligations (Acemoglu et al., 2005).

Artificial intelligence amplifies this mechanism because its infrastructure combines several layers of concentration. Developing frontier models requires vast quantities of capital, advanced chips, energy, data centres, networks, specialised expertise and access to enormous datasets. Everyday use of these systems is also organised through clouds, application programming interfaces and platforms that can record interactions, set prices and alter their conditions unilaterally.

There is no need to imagine complete coordination among all technology companies. It is enough to observe a structure in which a small number of organisations have disproportionate power to decide which models are trained, which languages and forms of knowledge are included, which applications receive investment and which risks are deemed acceptable. In the interview that provides the essay’s point of departure, Acemoglu captures this concentration through a deliberately provocative image: a transformative technology has been left in the hands of ‘five or ten people’ (Hernández, 2026).

This hybrid regime neither replaces the state nor operates outside it. It is formed at the intersection of transnational corporations, industrial policy, intellectual property, public procurement, financial markets, energy regulation and geopolitical rivalry. States fund research, allocate land and energy, purchase services, restrict exports and decide which infrastructures are strategic. They may act as democratic counterweights, as partners in concentration or as participants in the same agonistic competition (Acemoglu, 2025a).

Geopolitics strengthens this tendency. Competition with China is frequently used to present regulation as a unilateral surrender of technological leadership. The argument contains genuine concerns about security, supply chains and authoritarian models of digital control. Yet it can also operate as a device of closure. When the alternatives are reduced to a race between large Western corporations and an authoritarian state, democratic governance, international cooperation, open science and public capacity disappear from the debate.

The result is a false choice between corporate acceleration and total state control. Institutional analysis makes a more productive question possible. What distribution of power allows innovation without granting those who control the infrastructure the capacity to define the collective future unilaterally?

That question, however, leaves a second problem unresolved. Why Nations Fail explains how rules and de facto power enable extraction, but offers less vocabulary for understanding how extraction becomes admirable, inevitable or morally acceptable. Concentration is sustained not only by contracts and barriers to entry. It also needs ceremonies, promises and public demonstrations that turn private expenditure into social leadership. At this point, the anthropology of the gift ceases to be a decorative supplement and becomes an indispensable analytical tool.

AI as a Critical Juncture

Acemoglu and Robinson (2012) use the term critical junctures for periods of disruption in which war, crisis, economic transformation or innovation alters the existing equilibrium and expands the range of possible institutional paths. Initial differences, even small ones, may produce divergent outcomes when they interact with these ruptures.

Artificial intelligence can be interpreted as a technological critical juncture. This does not mean that models are, by themselves, an external event capable of remaking institutions. The juncture is formed through the combination of technical capabilities, massive investment, workplace reorganisation, geopolitical rivalry and regulatory decisions. It is critical because these forces destabilise previous settlements around knowledge, employment, intellectual property and state capacity.

The social response will determine whether this disruption reinforces an extractive cycle or opens a more inclusive path. The first possibility leads to extractive institutional drift. Public support consolidates private infrastructures, free access creates dependence, automation weakens collective bargaining and geopolitical exceptionalism reduces democratic control. The outcome might resemble what some approaches call digital feudalism, although the metaphor should be used cautiously. This is not a literal return to feudal relations, but a structure in which access to essential means depends on infrastructural lords able to impose conditions of continued use.

The second path turns disruption into the democratisation of productivity. Compute, data and models are organised as plural infrastructures; workers participate in design; public support generates obligations of openness; and automation is subordinated to the creation of capabilities and tasks. Neither path is guaranteed. Critical junctures expand possibilities, but they also intensify the struggle to close them down.

This contingency matters. The technological potlatch is not a destiny inscribed in the nature of artificial intelligence. It is a historical form of competition that emerges when scale, financial prestige and control of infrastructure receive greater recognition than the social circulation of their benefits.

The Potlatch Beyond Waste

Mauss (1925) analysed the gift as a ‘total social fact’. Giving, receiving and reciprocating were not economic operations separated from religion, law or politics, but actions that mobilised all these dimensions simultaneously. The gift was not free in the modern sense of a transfer without consequences. It created relationships, memory, obligation and debt.

The potlatch occupied a central place in this analysis. Practised by different peoples of the Northwest Coast, it is not a homogeneous institution and cannot be reduced to the early descriptions of colonial anthropology. In Kwakwaka’wakw, Haida, Tlingit and other contexts, ceremonies have served to distribute goods, confirm names and positions, transmit rights and summon a community as witness. The Canadian colonial prohibition of the potlatch sought precisely to destroy this political and cultural capacity (Royal BC Museum, n.d.).

Codere (1950) described the Kwakwaka’wakw potlatch as a way of ‘fighting with property’. Rivalry could be displaced from physical violence into competition regulated through distributions, debts and displays of wealth. In her study of the Tlingit mortuary potlatch, Kan (1989) showed that exchange also articulated memory, mourning and continuity between the living and the dead. These contributions prevent the potlatch from being reduced to a competition for economic prestige. Property circulated within worlds of kinship, law, territory and memory (Rosman & Rubel, 1972).

Piddocke (1965) also interpreted the potlatch among the Southern Kwakwaka’wakw as a mechanism of redistribution between groups exposed to ecological and productive fluctuations. Rivalry over rank could drive the circulation of surpluses towards communities experiencing temporary deficits. Symbolic competition and material subsistence were not incompatible functions.

This reading must avoid idealisation. The potlatch could reproduce hierarchies, create debt and exclude those who did not occupy particular positions. Its analytical interest lies precisely in showing that circulation does not eliminate power. It organises power and makes it visible. Wealth becomes authority because a collectivity recognises the capacity to distribute it and preserves the memory of the obligations incurred.

Bataille (1949) drew on the potlatch to develop a ‘general economy’ in which societies produce surpluses that must be expended. His emphasis on unproductive expenditure and loss helps to analyse investments oriented towards prestige rather than immediate returns. Applied without caution, however, this reading once again turns Indigenous institutions into spectacles of destruction. The theory of expenditure is useful here only when it remains subordinate to the political, legal and redistributive complexity of actual potlatch institutions.

The Technological Potlatch and Its Public

The contemporary race for artificial intelligence can be read as an economy of prestige. Corporations compete not only for customers or present profits. They compete to demonstrate that they possess the scale required to build the future. Announcements of data centres, chip clusters, energy investments and larger models operate as signals to investors, governments, specialised staff, dependent businesses and competitors.

Expenditure produces authority. A company able to commit tens or hundreds of billions appears to be an unavoidable actor. Its economic scale is transformed into symbolic capital. In Bourdieu’s (1986) terms, material capital assumes a form recognised as competence, leadership and the right to speak on behalf of the future.

The figures indicate the scale of the display. In February 2026, Reuters calculated that Alphabet, Microsoft, Amazon and Meta expected to invest more than US$600 billion between them during the year, much of it in expanding capacity associated with artificial intelligence (Raitano et al., 2026). Alphabet, for example, placed its annual capital expenditure forecast between US$175 billion and US$185 billion and stated that approximately 60 per cent would go towards servers (Alphabet Inc., 2026). These forecasts may change, and not all expenditure represents a loss. They nevertheless show that the ability to sustain investment on this scale has become both a competitive barrier and a public signal of technological sovereignty.

Where, then, is the deliberate destruction or loss that would justify describing this as an agonistic dynamic? Not every purchase of hardware is ritual destruction. Buildings, electricity connections, fibre and some cooling systems may remain useful for years. Specialised accelerators undergo faster cycles of replacement and depreciation, although there is no universal rule by which they lose all value within twelve or eighteen months.

Loss emerges from a combination of practices. Venture capital is burnt to maintain prices below full cost; generations of still-functional hardware are discarded to remain at the frontier; infrastructures are duplicated to avoid dependence on a rival; and models with uncertain commercial returns are trained. The ability to absorb these losses can force smaller firms either to withdraw or to integrate themselves into the infrastructure of those able to sustain them. Destruction is not necessarily the conscious purpose of each investment. It operates as a condition of entry that turns solvency into a competitive weapon.

Financial loss is accompanied by a less visible material destruction. Data centres concentrate growing demand for electricity and, where evaporative cooling is used, consume water in ways that place particular pressure on areas experiencing water stress. The accelerated manufacture and replacement of hardware also requires critical minerals and generates electronic waste (International Energy Agency, 2025; United Nations Environment Programme, 2025). Short-cycle hardware therefore gives the metaphor a material dimension: it functions as the contemporary equivalent of wealth broken to demonstrate capacity, while its ecological costs fall outside the financial ceremony.

The technological potlatch also has a fragmented public. There is no single ceremonial community able to record obligations and demand reciprocity. Financial markets observe expenditure as a signal of growth; specialist media turn launches and benchmarks into events; governments interpret scale as sovereignty; and developer communities assess capabilities through performance tables whose comparability is often contested. These publics recognise prestige, but none possesses sufficient authority on its own to impose a social return.

Reciprocity may even assume a negative form. Releasing a subsidised model, or one offered with open weights, can force rival companies to lower prices, redesign products or increase their own expenditure. ‘Generosity’ humiliates when it demonstrates that the giver can forgo the revenue on which competitors depend for survival. This does not make every open model an extractive manoeuvre. It shows that openness can also be used as a strategy of prestige, ecosystem capture and the transfer of costs to those who adapt and maintain the system.

The relevant question is not whether a technology is open or closed in the abstract. What matters is which capabilities it distributes, which dependencies it preserves and who can change its conditions.

Digital Hau and the Capture of Value

In the classical reading of Mauss (1925), the received object does not remain inert. It retains a force that connects it to its origin and helps demand a return. The familiar explanation through hau has been debated for a century. Some anthropologists argued that Mauss had turned a specific Māori category into a general theory of reciprocity. Contemporary Māori perspectives have also pointed to the danger of separating the concept from its relationships with land, genealogy and Indigenous knowledge (Frank, 2016; Stewart, 2017).

This controversy is more productive than a literal transposition. Free digital tools are not neutral objects that entirely leave those who provide them. They travel with licences, formats, interfaces, telemetry, accounts, updates and terms of service. The bond does not reside in a spirit inside the software, but in a technical and legal architecture that keeps the service attached to its provider.

The return takes several forms. Some platforms may use interactions and ratings to improve their systems, depending on their policies and applicable contractual options. Others obtain usage patterns, visibility, market position or the embedding of a product within organisational processes. Even when content is not used to train models, adoption generates information about demand, creates habits and raises the cost of leaving the ecosystem.

Users do not provide data alone. They also contribute adjustment work. They learn to formulate prompts, correct outputs, reorganise tasks, create templates, connect services and teach other people how to use the tool. Some of this work remains outside the provider’s accounts even though it increases the value of the platform.

With all the qualifications already stated, digital hau can be understood as the provider’s persistence within what has been received. The gift carries with it the power to change prices, withdraw features, switch models or redefine permitted uses. The obligation to return is not staged as an explicit ceremony, but as a silent accumulation of dependence.

The analogy reaches its most important limit here. Māori hau is not a metaphor available to explain every form of circulation. The digital world must be described through its own legal and material devices. Mauss is useful because he prevents free provision from being confused with the absence of a relationship.

Infrastructural Debt

The dependence created by subsidised access can be termed infrastructural debt. The concept is related to technological lock-in, but is not reducible to it.

Lock-in describes the costs that make it difficult to change providers once a technology has been adopted. Infrastructural debt emphasises the temporal asymmetry produced by the subsidised gift. The initial cost of adoption is low, while the cost of exit grows with every integration, every skill acquired, every dataset adapted and every process reorganised around the platform. Dependence accumulates during a phase in which its price remains partly hidden.

An organisation may begin with free credits for an API. It then adapts its software to particular responses, limits and formats. It trains its team, stores data in associated services and promises customers functions that depend on the model. When prices or conditions change, migration no longer means replacing a component. It requires part of the organisation to be rebuilt.

This debt does not affect all actors equally. A large company may negotiate contracts, build redundancies or distribute workloads between providers. A small organisation, local authority, co-operative or news outlet has less capacity to absorb migration. Initial free provision may democratise immediate access while concentrating future autonomy.

Open models, local execution and interoperable standards can reduce this debt. Yet ‘open’ is not a simple category. A model may provide downloadable weights while keeping its data, training process or enabling conditions opaque. It may also require infrastructure affordable to only a handful of organisations. Openness should be measured by the effective capacity to exit, adapt, audit and govern, not by the nominal availability of a file.

Work, Recognition and AI Washing

Automation has a particularly sensitive effect on entry-level positions. Many tasks that appear routine are also spaces of learning. Drafting initial versions, classifying information, reviewing code or assisting with an analysis enables people to acquire the tacit knowledge required to assume greater responsibilities later. If those tasks disappear without new training routes being created, a company secures immediate savings by eroding its own professional reproduction.

The interview mentions a 37 per cent fall in certain entry-level vacancies (Hernández, 2026). The figure should not be turned into a general diagnosis without precisely identifying the source, period and occupations observed. The conversation itself acknowledges that the causal debate remains open concerning the respective weight of automation, interest rates, earlier overhiring in technology and other post-pandemic adjustments.

This caution leads to the concept of AI washing. A company may attribute redundancies to automation even when the principal causes are previous overexpansion, reorganisation, financial pressure or management failures. The technological explanation is attractive because it turns a business problem into a sign of modernisation. Announcing efficiency through artificial intelligence may be better received than acknowledging a fall in demand.

AI washing takes other forms. Conventional systems are marketed as advanced intelligence; workplace surveillance is presented as optimisation; data capture is justified as the inevitable improvement of a model; and energy-intensive products are described as sustainable solutions without accounting for their infrastructure. These practices share a mechanism. The label ‘AI’ shifts attention away from an organisational decision and towards a supposedly technical necessity.

Discourse produces effects even when it exaggerates actual capabilities. Persuading investors and workforces that substitution is inevitable weakens collective bargaining, disciplines wage expectations and transfers responsibility for adaptation to each individual. Automation then functions both as a technique and as a narrative of power.

Work cannot be reduced to a cost variable or a mechanism for distributing income. It also organises learning, sociability, recognition and collective capacity. This does not mean idealising wage labour, which can be precarious, alienating or harmful. It means recognising that a policy focused exclusively on compensating for lost wages leaves the distribution of productive power intact.

A universal basic income may provide material security and expand people’s freedom to refuse abusive work. Its effect depends on institutional design, funding and the rights that accompany it. Without participation in technological decision-making, strong public services and access to means of production, a cash transfer could stabilise a highly unequal society. Integrated into a broader regime of rights and capabilities, it could strengthen autonomy.

Similar caution applies to reductions in working time. If a company must pay the same wage for fewer hours while retaining complete freedom to replace jobs, the accounting incentive to automate grows. This effect does not make every reduction in working time an absurd policy. It may distribute productivity when combined with collective bargaining, protections against dismissal, worker participation in decision-making and mechanisms that prevent the unilateral appropriation of time saved.

Taxes on robots or compute may slow automation or finance transitions, but they do not automatically alter the direction of innovation. A pro-worker policy requires positive incentives to develop tools that create tasks, enhance capabilities and improve services in areas of high social need.

When Productivity Erodes Learning

The problem is not confined to employment. The same tool can improve immediate performance while weakening the capacity that makes such performance possible over the longer term.

A working paper published in 2026 analyses thirty months of data from 26,811 students in Years 7 to 12 in China (Strömberg et al., 2026). According to its estimates, the adoption of artificial intelligence raised homework marks by 18 per cent and reduced completion time by 30 per cent. At the same time, scores in monthly closed-book examinations fell by 20 per cent during the first six months. Performance in higher-stakes entrance examinations declined by between 18 and 24 per cent, with more pronounced effects after prolonged periods of use.

The study uses a difference-in-differences strategy based on staggered adoption rather than a randomised experiment. It is also a recent paper that requires further scrutiny. Its findings do not justify the claim that every educational use of AI produces cognitive atrophy.

The available evidence shows that design matters. Other experiments find benefits when the tool acts as a tutor, requires learners to explain errors, guides discovery or is embedded in deliberate pedagogical practice (Fischer et al., 2025; Li et al., 2026). The risk is particularly acute when a system supplies complete solutions and replaces the effort through which knowledge is constructed.

Acemoglu, Kong, et al. (2026) have formalised a related concern through the idea of knowledge collapse. Individual learning generates not only a private signal useful to the learner, but small contributions to the collective stock of knowledge. An agentic system that provides context-sensitive recommendations may improve present decisions while simultaneously reducing the incentive to learn. If many people cease to produce general knowledge, society may preserve high-quality recommendations for a time while eroding the human foundation on which they depend.

This dynamic makes it possible to speak of a degradation of human capital under an extractive model of software consumption. The platform supplies answers and captures use, while displacing from immediate experience the cost of losing understanding, judgement and autonomy. The productivity of the task conceals the unproductivity of the trajectory.

Inclusive educational AI would have to expand the pedagogical relationship rather than replace it. It could help teaching staff identify difficulties, adapt materials and release time for personal support. It should also be designed to elicit explanation, comparison and active recall, rather than eliminate the cognitive effort that constitutes learning.

Giving in a Way That Compels Sharing

From the perspective of the anthropology of the gift, a practice described by López García (2001) offers a particularly useful image for thinking about this direction. Among Maya Ch’orti’ communities, he identified a model of giving that can be summarised as ‘giving food in a way that compels its redistribution’. The scale and form of the gift prevent the recipient from consuming it individually and require its circulation to continue.

The point of the example is not to turn a ritual practice into a recipe for digital policy. It allows us to imagine institutions in which redistribution does not depend on the subsequent goodwill of the beneficiary. The obligation to circulate is built into the design itself.

What would be the equivalent of providing a good that cannot be consumed individually? Advanced compute, large datasets and some models require collective resources in order to acquire social value. A university acting alone, a small public authority or a co-operative can rarely sustain the whole infrastructure by itself. The need to share, however, can be resolved either through a centralised private platform or through common institutions. Both aggregate resources, but distribute decision-making and dependence differently.

Providing a model free of charge does not necessarily distribute capability. Capability includes knowledge, infrastructure, autonomy, the right to modify, protection against unilateral changes and power to decide which applications are developed.

A policy of ‘giving in a way that compels sharing’ could attach conditions to subsidies, public contracts and privileged access to scarce resources. A company that receives funding, energy, land, publicly funded research or tax advantages should assume verifiable obligations concerning interoperability, training, auditing, task creation and the return of knowledge. Public support would be neither a transfer without return nor a purely financial equity stake. It would be an architecture of circulation.

This perspective alters the question of who deserves prestige. In the potlatch, authority came not only from possession, but from distribution before a community capable of witnessing the act. An inclusive digital economy should recognise as leadership the ability to expand the commons, not only the volume of capital mobilised. The largest model would not necessarily be the most valuable. Greater value might lie in the model that strengthens the greatest range of capabilities beyond the organisation that controls it.

Inclusive Digital Infrastructure

The institutional alternative can be formalised through the concept of inclusive digital infrastructure. This is the set of compute resources, data, models, knowledge, standards and public or common capacities that enables a plurality of actors to develop and govern artificial intelligence without becoming subordinate to a single provider.

The term does not mean nationalising the entire industry or reproducing the same race for scale under state ownership. Infrastructure is inclusive when it expands effective access and distributes decision-making capacity. It requires transparent allocation criteria, interoperability, auditing, representation of affected communities and exit mechanisms. Public ownership may facilitate these conditions, but it does not guarantee them.

Historical comparison with the Industrial Revolution must be handled carefully. Extended property rights, public education, urban infrastructures and labour protection did not arise automatically from industrialisation. They were incomplete and uneven outcomes of political conflict. Their importance lay in preventing all productivity gains from remaining within factories or the fortunes that controlled them. Inclusive digital infrastructure would seek to perform a comparable function in the knowledge economy.

The European Union offers an example whose direction remains open. By August 2026, the EuroHPC network included nineteen AI Factories and thirteen Antennas intended to connect supercomputing, universities, small and medium-sized enterprises, industry and public authorities. The European Commission and participating states expect to mobilise €10 billion between 2021 and 2027 for supercomputing and AI Factories (European Commission, 2026; European High Performance Computing Joint Undertaking, 2026). Access to compute and technical support can reduce barriers to entry that no start-up could overcome by itself.

At the same time, the project reveals the ambivalence of public capacity. Future European AI Gigafactories are also justified through sovereignty, global competition and models with trillions of parameters. Without rules governing access, social evaluation and plural governance, public infrastructure may reproduce the technological potlatch at state level. Inclusion depends on who sets priorities, who receives compute and which obligations accompany that access.

Inclusive digital infrastructure would contain four inseparable components.

The first is shared material capacity. This includes supercomputing, storage, networks and energy, with reserved allocations for research, public authorities, co-operatives, civil-society organisations and small firms.

The second is open and auditable knowledge. This includes documentation, evaluations, responsibly governed public data, adaptable models and tools that enable outputs to be inspected without disclosing personal information or legitimate secrets.

The third is human capacity. Infrastructure needs qualified public-sector staff, training for workers, community mediation and teams able to audit complex systems. Buying chips without building institutions of knowledge reproduces dependence.

The fourth is democratic governance. Priorities must respond to verifiable social needs, with the participation of workers and affected communities. Auditing must have consequences, and public access must generate public returns.

Institutionalising Reciprocity

Current regulation is often reactive. A product is deployed, harm appears and institutions then attempt to correct it. Acemoglu proposes reversing this logic. A society should first decide which objectives it seeks and then direct technological incentives towards them.

This does not mean that a government can foresee every innovation or should design models from a central office. It means recognising that an industrial policy for artificial intelligence already exists. Decisions about energy, chips, public procurement, taxation, intellectual property, research and competition are directing the market even when presented as neutrality.

An inclusive agenda can be organised around five obligations. These proposals build on the regulatory floor established by Regulation (EU) 2024/1689 and amended in July 2026 by Regulation (EU) 2026/1744. The Artificial Intelligence Act has applied generally since 2 August 2026, but the amendment postponed much of the high-risk regime until 2 December 2027; rules for relevant AI systems embedded in regulated physical products apply from 2 August 2028. The aim is not to imply that every high-risk obligation already applies, but to extend this regulatory floor towards interoperability, effective rights of exit and worker participation (Regulation (EU) 2024/1689, 2024; Regulation (EU) 2026/1744, 2026).

1. Participation Before Deployment

Workers and affected communities must be involved before systems reorganise pay, working time, evaluation, surveillance or the allocation of tasks. Collective bargaining over algorithms can address pace, supervision, the distribution of productivity gains and changes to job functions. The European Parliament has proposed prior consultation for algorithmic management systems that substantially affect work, including information on data, bias, occupational health and human oversight (European Parliament, 2025).

2. Effective Openness and the Right of Exit

Portability, interoperable standards and public or co-operative alternatives reduce infrastructural debt. Article 53 of the European Union’s Artificial Intelligence Act requires providers of general-purpose AI models to maintain technical documentation, supply information to those who integrate the models and publish a sufficiently detailed summary of the content used in training (Regulation (EU) 2024/1689, 2024). This minimum standard of transparency should be extended towards effective interoperability. Openness should be judged by the autonomy it creates. The right of exit includes exporting data, preserving essential functions during a transition and preventing formats or contracts from turning migration into a disproportionate penalty.

3. Redistribution of Capabilities

Public funding should favour technologies that create human tasks, extend knowledge and improve fields such as health, education, care, science, accessibility and ecological transition. Every significant form of support should include measurable social-return clauses. Public money, energy and data must generate public capacity.

4. Material Accountability

Audits, assessments and consultations are valuable only if they can alter decisions, impose remedies and halt unacceptable uses. Article 43 of the Artificial Intelligence Act establishes conformity-assessment procedures for high-risk systems, while Article 72 retains post-market oversight. Regulation (EU) 2026/1744 changed the implementation timetable for that regime and removed the requirement for a single harmonised post-market monitoring plan, but it did not eliminate the broader need for material accountability and oversight throughout a system’s lifetime (Regulation (EU) 2024/1689, 2024; Regulation (EU) 2026/1744, 2026). This principle should extend to any system that substantially reorganises work, even when it has not formally been classified as high risk. Organisations must explain which part of a decision comes from the system, who is accountable for it and how it can be challenged. An audit must be able to trigger correction, suspension and redress; otherwise, it becomes an empty ceremony of legitimation.

5. Plural Infrastructure

Public, university, municipal, co-operative or shared compute expands the space for experimentation beyond the largest platforms. It will not replace all private investment, but it can prevent access to advanced knowledge from depending entirely on a few companies. Networks should co-ordinate to share data, services, training and good practice without building a new monopolistic centre.

These obligations do not abolish competition. They reorganise it. The potlatch shows that the pursuit of prestige can mobilise resources towards circulation when an institution links rank to distribution. The challenge is to create a regime in which companies and states also compete to demonstrate how much knowledge they make accessible, how many capabilities they distribute and how much of their power they submit to public decision-making.

Conclusion

Artificial intelligence will not succeed or fail solely because of the quality of its models. Its trajectory will depend on the institutions that translate technical capacity into a particular distribution of power.

The framework of Why Nations Fail reveals the danger of an extractive cycle. The concentration of compute, data and capital produces political influence. That influence generates rules favourable to concentration. Those rules consolidate control of infrastructure and reduce society’s ability to direct technological change. The critical juncture created by artificial intelligence may deepen this drift or interrupt it.

The anthropology of the gift adds a question that institutional economics does not always ask. How does this concentration become legitimate? The industry does more than accumulate resources. It displays expenditure, distributes access, promises abundance and converts scale into moral authority. Its apparent generosity creates dependencies, while the obligation to return value to society remains weak or voluntary.

Describing this process as a ‘technological potlatch’ is useful only if the difference is preserved. Indigenous potlatch institutions were not metaphors for economic irrationality. They articulated circulation, rights, memory, territory and public recognition. For precisely this reason, they expose the central deficiency of the digital regime. Corporations can gain prestige through spending and giving without submitting to a community with equivalent power to demand reciprocity.

Nor does hau provide a mystical explanation of data capture. Its discussion reminds us that given things retain connections to their origins. In the digital world, these connections are materialised in licences, accounts, telemetry, formats and dependencies. The free gift may carry an infrastructural debt whose cost becomes visible only when someone attempts to leave.

Educational evidence reveals the same problem at another scale. A tool may improve the immediate result while eroding the learning that sustains future autonomy. Extraction is then no longer confined to wages or data. It may appropriate the cognitive effort that a society needs in order to preserve knowledge and judgement.

Pro-worker artificial intelligence requires more than better models or compensation after the event. It requires institutions that build circulation into their design, distribute decision-making capacity and convert productivity gains into time, knowledge, services and shared power.

Inclusive digital infrastructure offers a concrete horizon, but not a guarantee. Public compute may democratise capabilities or become another stage for geopolitical display. Its inclusive character will be decided by the obligations it imposes, the actors who participate and the goods it succeeds in keeping in circulation.

The decisive criterion, therefore, should not be how much capital an organisation can mobilise to build the largest model, but the density of the public obligations generated by that process. When society contributes data, energy, labour, knowledge and infrastructure, it also acquires the right to decide how benefits circulate and what limits are placed on those who concentrate technological capacity. Without enforceable obligations, the technological gift becomes debt. With institutions capable of distributing its returns, the critical juncture of artificial intelligence can still be directed towards a common good.

References

Acemoglu, D. (2025a). Institutions, technology and prosperity (NBER Working Paper No. 33442). National Bureau of Economic Research. https://doi.org/10.3386/w33442

Acemoglu, D. (2025b). The simple macroeconomics of AI. Economic Policy, 40(121), 13–58. https://doi.org/10.1093/epolic/eiae042

Acemoglu, D., Autor, D. H., & Johnson, S. (2026). Building pro-worker artificial intelligence (NBER Working Paper No. 34854). National Bureau of Economic Research. https://doi.org/10.3386/w34854

Acemoglu, D., Johnson, S., & Robinson, J. A. (2005). Institutions as a fundamental cause of long-run growth. In P. Aghion & S. N. Durlauf (Eds.), Handbook of economic growth (Vol. 1A, pp. 385–472). Elsevier. https://doi.org/10.1016/S1574-0684(05)01006-3

Acemoglu, D., Kong, D., & Ozdaglar, A. (2026). AI, human cognition and knowledge collapse (NBER Working Paper No. 34910). National Bureau of Economic Research. https://doi.org/10.3386/w34910

Acemoglu, D., & Robinson, J. A. (2012). Why nations fail: The origins of power, prosperity, and poverty. Crown Publishers.

Alphabet Inc. (2026, February 4). 2025 Q4 earnings call. https://abc.xyz/investor/events/event-details/2026/2025-Q4-Earnings-Call-2026-Dr_C033hS6/default.aspx

Bataille, G. (1949). La part maudite. Les Éditions de Minuit.

Bourdieu, P. (1986). The forms of capital. In J. G. Richardson (Ed.), Handbook of theory and research for the sociology of education (pp. 241–258). Greenwood Press.

Codere, H. (1950). Fighting with property: A study of Kwakiutl potlatching and warfare, 1792–1930 (American Ethnological Society Monograph No. 18). University of Washington Press.

European Commission. (2026, April 23). AI factories. Shaping Europe’s Digital Future. https://digital-strategy.ec.europa.eu/en/policies/ai-factories

European High Performance Computing Joint Undertaking. (2026, April 28). Cooperation of artificial intelligence factories and factories antennas. https://www.eurohpc-ju.europa.eu/cooperation-artificial-intelligence-factories-and-factories-antennas-0_en

European Parliament. (2025, December 17). Digitalisation, artificial intelligence and algorithmic management in the workplace—Shaping the future of work (P10_TA(2025)0337). https://eur-lex.europa.eu/eli/C/2026/2154/oj/eng

Fischer, M., Rau, H. A., & Rilke, R. M. (2025). AI tutoring enhances student learning without crowding out reading effort (IZA Discussion Paper No. 18338). IZA Institute of Labor Economics. https://www.iza.org/publications/dp/18338/ai-tutoring-enhances-student-learning-without-crowding-out-reading-effort

Frank, S. (2016). The “force in the thing”: Mauss’ nonauthoritarian sociality in The gift. HAU: Journal of Ethnographic Theory, 6(2), 255–277. https://doi.org/10.14318/hau6.2.017

Hernández, J. [Inteligencia Artificial]. (2026, July 16). El impacto económico de la IA con el Premio Nobel Daron Acemoglu [The economic impact of AI with Nobel laureate Daron Acemoglu] [Video]. YouTube. https://youtu.be/CDzTaO_nhAc

International Energy Agency. (2025, April 10). Energy and AI. https://www.iea.org/reports/energy-and-ai

Kan, S. (1989). Symbolic immortality: The Tlingit potlatch of the nineteenth century. Smithsonian Institution Press.

Leeson, B. W. (ca. 1897–1908). Quatsino potlatch [Photograph]. Wikimedia Commons. https://commons.wikimedia.org/wiki/File:Quatsino_potlatch.jpg

Li, A. T., Liu, D., & Ye, T. (2026). Is ChatGPT a boon or a bane for learning? Experimental evidence across task formats and chatbot designs [Working paper]. SSRN. https://doi.org/10.2139/ssrn.5533921

López García, J. (2001). «Dar comida obligando a repartirla». Un modelo de don maya-ch’orti’ en proceso de transformación [‘Giving food in a way that compels its redistribution’: A changing model of the gift among the Maya Ch’orti’]. Revista de Dialectología y Tradiciones Populares, 56(2), 75–98. https://doi.org/10.3989/rdtp.2001.v56.i2.210

Mauss, M. (1925). Essai sur le don. Forme et raison de l’échange dans les sociétés archaïques. L’Année Sociologique, nouvelle série, 1, 30–186.

Piddocke, S. (1965). The potlatch system of the Southern Kwakiutl: A new perspective. Southwestern Journal of Anthropology, 21(3), 244–264. https://doi.org/10.1086/soutjanth.21.3.3629231

Raitano, L., Ranasinghe, D., & Oguh, C. (2026, February 6). Big Tech’s $600 billion spending plans exacerbate investors’ AI headache. Reuters. https://www.investing.com/news/stock-market-news/global-software-data-firms-slide-as-ai-disruption-fears-compound-jitters-over-600-billion-capex-plans-4489820

Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act), Official Journal of the European Union, L, 2024/1689 (2024). https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng

Regulation (EU) 2026/1744 of the European Parliament and of the Council of 8 July 2026 amending Regulations (EU) 2024/1689, (EU) 2018/1139 and (EU) 2023/1230 as regards the simplification of the implementation of harmonised rules on artificial intelligence (Digital Omnibus on AI), Official Journal of the European Union, L, 2026/1744 (2026). https://eur-lex.europa.eu/eli/reg/2026/1744/oj/eng

Rosman, A., & Rubel, P. G. (1972). The potlatch: A structural analysis. American Anthropologist, 74(3), 658–671. https://doi.org/10.1525/aa.1972.74.3.02a00280

Royal BC Museum. (n.d.). Kwakwa̱ka̱ʼwakw: Stories carved in cedar. Retrieved August 5, 2026, from https://totems.royalbcmuseum.bc.ca/en/story/Legacy

Stewart, G. (2017). The ‘hau’ of research: Mauss meets Kaupapa Māori. Journal of World Philosophies, 2(1), 1–11. https://doi.org/10.2979/jourworlphil.2.1.01

Strömberg, D., Lei, V., & Wu, Y. (2026). The generative AI learning penalty: Evidence from Chinese secondary education (CEPR Discussion Paper No. 21577). CEPR Press. https://cepr.org/publications/dp21577

United Nations Environment Programme. (2025, November 13). AI has an environmental problem. Here’s what the world can do about that. https://www.unep.org/news-and-stories/story/ai-has-environmental-problem-heres-what-world-can-do-about