When human knowledge is compressed into model weights, and humans then pay by the token to retrieve that knowledge, are we purchasing a service—or paying rent to a new class of digital landlord?
This question carries surprising intuitive force. Centuries of accumulated text, code, images, public discourse, and professional expertise have been transformed into repeatedly callable machine capabilities; enterprises that own compute, models, and distribution gateways can price these capabilities; and people who once made a living through professional knowledge may increasingly need to subscribe to these enterprises' tools just to continue doing their own work.
This seems to constitute a perfect Capital narrative: socialized production, private appropriation; knowledge enclosure, the separation of workers from the means of production; technology raising productivity, yet returns concentrating in the hands of a few capital owners.
But if we are truly to use Marx's analytical method, we cannot treat elegant metaphors as already-proven economic relations. Not all training data is unauthorized; not all model revenue is rent; not all AI productivity gains translate into corporate profit; and international technology trade is not inherently equivalent to colonialism. Meanwhile, open weights, falling inference prices, cross-model competition, and agent interoperability protocols are undermining the exclusivity on which certain "new landlord" narratives depend.
This article fuses two modes of writing: it preserves the original manuscript's critical insights into the means of knowledge production, global inequality, and "intelligence rent," while incorporating the counter-evidence and fact-checking produced by subsequent in-depth research. Its core conclusion is not that "AI has already entered techno-feudalism," but a more rigorous and more explanatory judgment:
AI is re-encoding social knowledge as capital; but whether this capital takes the form of a competitive commodity, a temporary innovation surplus profit, or a long-term economic rent dependent on exclusive control, is determined jointly by technology diffusion, property-rights arrangements, market structure, and institutional choice. What is most worth watching is not necessarily the model itself, but the agent gateways that control identity, memory, payments, permissions, and workflows.
TL;DR|Six Key Judgments
- "AI colonialism" is a structural dependency that must be proven—not a synonym for all cross-border AI transactions. We must observe who controls irreplaceable production conditions, who sets the rules, who has exit rights, and who remains persistently stuck in low-return segments.
- "Data enclosure" shares mechanisms with Marx's primitive accumulation, but the two cannot be directly equated. The key is not merely data acquisition, but whether knowledge producers thereby lose their autonomous production conditions and are forced into ongoing dependence on a new platform.
- The clearest Capital mechanism in AI is the raising of labor productivity, the reorganization of the labor process, and the redistribution of profit. In a strict labor theory of value, machines do not independently create new value; the "value added" of modern national economic accounting is a different concept.
- The model layer simultaneously exhibits enormous frontier R&D costs and rapidly declining costs for using existing capabilities. Concentration in chips, advanced manufacturing, cloud, and electricity does not mean that every dollar of API revenue is monopoly rent.
- "Intelligence rent" is currently better understood as a falsifiable hypothesis about agent gatekeeping power. Only when some form of identity, memory, authorization, payment, or workflow control creates hard-to-migrate dependency and sustained excess returns is there reason to speak of "rent-like" income.
- The core institutional variables are exit rights and benefit distribution. Open models, interoperability, public compute, labor benefit-sharing, and competition policy are more readily translated into testable governance practice than simply declaring "AI is public property."
Not "the machines are coming," but the relations of production have changed
The starting point of Capital is not whether technology is good or bad, but who owns the means of production, how labor is organized, and how surplus is distributed.
In industrial capitalism, factories, machines, raw materials, and land constitute production conditions that workers cannot easily obtain independently. The AI era has produced another set of production conditions: compute, training corpora, model weights, deployment infrastructure, tool permissions, industry data, identity systems, and the workflows that connect models to real business operations. We may provisionally call these "cognitive means of production," but this concept does not imply that knowledge itself is an entity that can be permanently occupied like land.
The key change is this: a person's knowledge may still reside in their head, but the organizational capability, customer reach, automation tools, and data access rights needed to complete work in the marketplace may gradually migrate to platforms.
This bears resemblance to what Marx described as the separation of workers from the means of production. However, "resemblance" is not "already accomplished." If an independent developer can use local models, freely switch cloud services, and retain customer relationships, they have not fully lost their production conditions. Conversely, if a company's accumulated business knowledge is locked inside a non-exportable agent system, and both employees and the enterprise can only maintain productive capacity by paying ongoing subscription fees, the separation mechanism becomes far more evident.
| Capital concept | Possible correspondence in the AI economy | Misreadings to avoid |
|---|---|---|
| Primitive accumulation | Data, knowledge, and production conditions exclusively capitalized | Scraping large amounts of web content is not itself equivalent to primitive accumulation |
| Constant capital c | GPUs, servers, capitalized software and model infrastructure | Cannot mechanically equate all CapEx with c |
| Variable capital v | Wage expenditure purchasing labor capacity | Accounting categories for data, code, and R&D differ from value-theory categories |
| Surplus value s | Surplus created by living labor and appropriated by capital | AI service revenue does not equal surplus value "created" by AI itself |
| Relative surplus value | Productivity gains altering the ratio of necessary labor to surplus labor | A single industry's efficiency gain does not necessarily immediately reduce the value of labor power |
| Average profit and price of production | Social surplus competitively redistributed among capital sectors | A company's high profit does not mean it all originates from its own sector's labor |
| Rent | Exclusive property rights blocking free entry, persistently converting excess profit into entitlement income | API pricing, gross margin, and rent cannot be conflated |
| World market | Transnational division of labor, technological dependency, surplus distribution, and state competition | Cross-border payment is not a sufficient condition for a colonial relationship |
Theoretical sources: Capital, Volume I, Chapter 12 (Relative Surplus Value), Chapter 15 (Machinery and Large-Scale Industry), Chapter 26 (Primitive Accumulation), and Volume III, Chapter 9 (Average Profit) and Chapter 45 (Rent).
Knowledge enclosure: from social collaboration to model capital
The accumulation cycle most worth studying in the AI era can be written as:
Social knowledge and labor → datafication and licensing arrangements → model training → capitalization of models and compute → API / agent / application monetization → new workflows and feedback data → reinvestment.
This cycle reveals an important asymmetry: knowledge production may be highly dispersed, yet model training and commercialization demand high concentration of capital, compute, and organizational capability. Many authors, programmers, translators, designers, annotators, forum participants, and ordinary users jointly generate the training corpus; only a few institutions have the capacity to turn large-scale corpora into commercially viable frontier models.
But "social knowledge absorbed by models" still does not prove "social knowledge has been plundered." The rights status of training data includes at least: public domain, open licenses, commercial licensing, copyrighted-but-unauthorized material, personal data, enterprise-proprietary data, and mixed sources of contested legality. The U.S. Copyright Office's 2025 report on generative AI training explicitly states that fair use requires case-by-case factual judgment, and one cannot pre-classify all training as either lawful or infringing. U.S. Copyright Office report
The real question is: Does the training act transform production conditions that creators, workers, or enterprises could otherwise use independently into capabilities usable only by external platforms?
If a publisher licenses its works to a model for fair consideration and retains independent distribution channels, this is closer to a commercial exchange. If a creator's works are used at scale without permission, the model becomes a substitute in their professional market, and the creator lacks bargaining power and exit paths, then the political economy of "enclosure" becomes far more compelling.
Here we must also distinguish two forms of capitalization:
- Modelization of knowledge: encoding dispersed knowledge into model capabilities, producing new production tools.
- Closure of knowledge-use rights: using technology, contracts, ecosystems, or law to prevent original contributors and other competitors from continuing to use these capabilities independently.
The first does not necessarily lead to the second. Open weights, reproducible research, public datasets, and local deployment can precisely cause capitalized knowledge to diffuse again.
Therefore, the AI analogy to primitive accumulation requires at least two evidentiary conditions: producers lose key production conditions; and after losing them, they are forced to depend on new owners to continue producing. Without the second condition, one cannot declare a digital enclosure movement accomplished merely by citing "data was scraped."
Is AI a new worker, or a new machine?
In Volume I, Marx distinguishes constant capital, variable capital, and surplus value:
In this theoretical framework, machines can gradually transfer their existing value into products and can dramatically raise labor productivity, but they are not regarded as living labor subjects that independently create new value. Even if AI can generate code, text, and designs, it cannot thereby be automatically classified as part of the "working class" in the Marxian sense.
This is not a metaphysical verdict on whether AI is "truly intelligent," but a specific definition of the source of value under the labor theory of value. When modern economics discusses AI, it typically uses value added, total factor productivity, consumer surplus, and welfare gains. "How much economic value AI creates" and "how much new value in the Marxian sense AI creates" cannot be directly interchanged.
AI's most direct effect on capital is to alter necessary labor time, labor intensity, product cost, and competitive advantage.
Suppose a certain category of customer service tasks originally required ten hours, and now with AI assistance requires only six. When market prices have not yet fully declined, the enterprise that first adopts the technology may capture excess profit; after the technology diffuses, competition will drive prices, wages, service quality, and industry structure to readjust. Strictly speaking, the excess surplus value an individual capital captures through productivity advantage and the rise in the relative surplus value rate that society as a whole experiences due to the falling value of wage goods are not the same thing.
Existing empirical studies support "localized productivity gains," but do not support "returns inevitably flow entirely to model companies."
Erik Brynjolfsson, Li, and Raymond's 2025 study, published in the Quarterly Journal of Economics, observed 5,172 customer service agents and found that AI assistance increased the number of issues resolved per hour by approximately 15% on average, with gains concentrated among less-experienced workers. Journal article
Another set of three randomized field experiments covering 4,867 software developers produced a combined estimate of approximately 26% higher task completion, though results varied considerably across experiments and estimation uncertainty was non-trivial. Developer experiments
Meanwhile, a study of approximately 25,000 workers across 7,000 workplaces in Denmark found that AI use yielded average time savings of about 3%, with no detectable significant changes in wages or recorded work hours. NBER working paper
Taken together, this evidence shows: task efficiency, enterprise profit, worker wages, consumer prices, and working hours are five distinct variables. A productivity gain only opens up the possibility of redistribution; it does not predetermine who ultimately benefits.
Why Volume III is better suited than Volume I for studying AI
Suppose a bank saves $100 million annually through AI. Part of this becomes bank profit, part is passed on as lower financial service prices, and part is paid to cloud, chip, and model suppliers. The returns captured by the cloud company do not necessarily originate entirely from the surplus value created by its own employees; it may also, through prices of production, procurement contracts, and market power, share in the surplus generated by other industries.
Marx's Volume III analysis of average profit, prices of production, and inter-sectoral surplus distribution is more explanatory here than "each company can only appropriate the surplus value created by its own employees."
This also reminds us: the source of profit, the sector in which profit is realized, and the entity that holds the key asset may not be the same.
The seven-layer AI value chain: where does scarcity actually lie?
Treating the "AI industry" as a single unified market obscures the real economic structure. In 2026, there are at least seven interlocking markets with varying degrees of competition.
| Layer | Core production conditions | Possible market power | Return nature to distinguish |
|---|---|---|---|
| ① Energy, land, grid connection | Electricity, water, location, permits | Physical scarcity and local regulation | Resource scarcity returns, cost recovery, potential rent |
| ② Semiconductor equipment | EUV, advanced manufacturing equipment | Technology barriers, intellectual property | Innovation profit and market power |
| ③ Chip manufacturing & packaging | Advanced processes, HBM, packaging capacity | Extreme capital thresholds, yield, and scale | Competitive profit, scarcity premium |
| ④ Accelerators & software stack | GPU / ASIC, compilers, ecosystem | Developer ecosystem and migration costs | Product profit, ecosystem lock-in premium |
| ⑤ Cloud & compute services | Data centers, networking, IaaS/PaaS | Economies of scale, contracts, and data egress costs | Normal return, lock-in returns, potential rent |
| ⑥ Models & APIs | Weights, R&D, inference services | Frontier performance, reliability, brand | Innovation excess profit and price competition |
| ⑦ Agents & applications | Identity, memory, permissions, payments, workflows | Default gateways, relationship assets, switching costs | Service profit and potential gatekeeping rent |
Industry data: do not conflate CapEx, revenue, and rent
Semiconductors. TSMC's 3nm process accounted for approximately 24% of its wafer revenue in 2025; ASML sold 48 EUV systems that year. These figures indicate advanced-manufacturing concentration and entry barriers, but a single company's gross margin does not directly equal economic rent. TSMC 2025 annual report · ASML 2025 annual report
Accelerators. Nvidia's fiscal year 2026 revenue was approximately $215.9 billion, with data center revenue in the fourth quarter at approximately $62.3 billion. This demonstrates AI infrastructure demand and enterprise pricing power, but gross profit must also cover R&D, capital risk, and competitive returns—it cannot be directly labeled "rent." NVIDIA financial results
Cloud. Synergy Research estimates that the global cloud infrastructure services market was approximately $143.4 billion in Q2 2026; AWS, Microsoft, and Google held revenue shares of approximately 28%, 20%, and 15% respectively, totaling 63%. This is cloud infrastructure services revenue share—not "63% of the entire AI market," and certainly not "63% of all money is being extracted as rent." Synergy Research, 2026-07-30
Electricity and capital expenditure. The IEA's 2026 update estimates that global data center electricity consumption will grow from approximately 485 TWh in 2025 to approximately 950 TWh by 2030. The IEA also notes that capital expenditure by five major tech companies already exceeded $400 billion in 2025 and is projected to increase by approximately 75% in 2026, potentially reaching a combined total exceeding $700 billion. The latter is forecast capital expenditure—neither realized revenue nor profit, let alone "intelligence rent." IEA Energy and AI · IEA 2026 update
Model prices. The Stanford 2025 AI Index records that the inference price for models achieving approximately GPT-3.5-level MMLU performance fell from approximately $20 per million tokens in November 2022 to approximately $0.07 per million tokens in October 2024. This is a price change at a fixed capability benchmark and cannot be extrapolated to "the inference cost of all frontier models has declined by the same factor." Stanford AI Index 2025
Meanwhile, the Stanford 2026 AI Index notes that industry accounted for 91.6% of the notable AI models cataloged in 2025. This is a sample under specific inclusion criteria—not "91.6% of all models worldwide are owned by private capital." Stanford AI Index 2026
This set of data reveals one of the most important paradoxes of AI capital accumulation:
The fixed costs of frontier R&D and infrastructure are rising ever higher, yet the unit price of using existing intelligence capabilities can keep falling.
Land cannot be copied, but model capabilities can diffuse through algorithm publication, personnel movement, distillation, open source, and hardware optimization. The claim that "models inherently produce permanent rent" is therefore encountering factual counter-evidence. High fixed investment and intense price competition can occur simultaneously.
What kind of return deserves to be called "intelligence rent"?
The original manuscript used "intelligence rent" to describe capital's ongoing charging for social knowledge. This concept has critical force, but requires strict narrowing.
In Marx's rent theory, land ownership forms an exclusive right that blocks the free entry of capital, allowing some excess profit to be stably converted into landowner income. What modern economics calls economic rent is the portion a resource earns above what is necessary to keep it in its current use. The two are related but are not the same definition.
In the AI economy, we can first express economic rent through an approximate measurement framework:
Where K is comparable invested capital, and the normal return must be controlled for R&D risk, cost of capital, asset depreciation, and service quality. Only when sustained excess returns stem primarily from exclusive control, entry barriers, or high switching costs—rather than temporary technological leadership—is it worth calling the income "rent-like."
This definition rules out three common conflations:
- API fees ≠ rent. Providing inference services involves real compute, operations, R&D, and risk costs.
- High gross margin ≠ rent. Gross profit is not the same as economic rent after deducting all costs and normal capital returns.
- Large CapEx ≠ rent. CapEx is an investment that may even lead to excess capacity and losses.
We can distinguish four candidate categories of "rent":
Compute scarcity returns arise from supply bottlenecks in electricity, grid connection, advanced chips, packaging, or data centers in specific geographies; model licensing and technology monopoly returns arise from short-term frontier performance and hard-to-substitute proprietary capabilities; cloud platform lock-in returns arise from proprietary interfaces, data egress, contracts, and ecosystem migration costs; agent gatekeeping returns may arise from exclusive control over user identity, long-term memory, business permissions, payment authorization, and default distribution gateways.
Of these, the first three already have varying degrees of evidence for market concentration and lock-in; the long-term economic rent mechanism of the fourth remains insufficiently proven as of 2026.
The real new land may not be the model, but the right to "act on your behalf"
Models answer questions; agents may execute actions on behalf of users and enterprises.
This distinction has economic significance. A generic model can be substituted between different suppliers; but if an agent holds three years of an enterprise's customer communication records, ERP permissions, approval workflows, contract terms, employee identities, and payment strategies, its replacement cost may be far higher than that of the underlying model.
The most critical scarce asset of the future may not be "the smartest model," but rather:
Identity + Memory + Workflow + Payment + Distribution
That is, identity, memory, workflow, payment, and distribution.
When an agent can decide which model to call, which supplier to recommend, which payment channel to use, and which business actions to execute, it is no longer merely a production tool but a gateway to economic transactions. If the gateway's owner can continuously charge commissions to merchants, developers, and users while impeding their migration to other gateways, a structure resembling platform rent emerges.
Early signals already appeared in 2026: Visa launched a pilot for agents executing enterprise payments; Mastercard introduced a payment solution for machines and agents. Visa 2026 pilot · Mastercard Agent Pay for Machines
But countervailing forces are equally real. Anthropic donated the Model Context Protocol (MCP) to the Linux Foundation's Agentic AI Foundation; the A2A protocol is also advancing interoperability among agents from different vendors. MCP and Foundation · Linux Foundation A2A 2026
Open protocols do not equal data portability, nor do they equal transferable payment authorization. But they do constitute an important counter-evidence to the claim that "agents will inevitably feudalize."
How to falsify "agent intelligence rent" from 2027 to 2030
| Observation metric | Outcome supporting the rent hypothesis | Outcome refuting the rent hypothesis |
|---|---|---|
| Memory portability | Long-term memory cannot be fully exported or reconstructed | Structured memory can migrate across platforms |
| Identity & permissions | User identity and enterprise authorization bound to a single agent | Independent identity layer, revocable and transferable permissions |
| Payment & commercial gateway | A few platforms long control payment authorization and merchant exposure | Multi-wallet, multi-payment-rail, cross-platform authorization |
| Workflow assets | Business processes highly dependent on proprietary scripts and history | Standardized workflow export and replay |
| Model selection rights | Agent prohibits or penalizes underlying model switching | Enterprises can freely route across multiple models |
| Merchant commission rate | High take rate persists long after competitors enter | Commissions continuously compressed by competition |
| Customer churn rate | Users unable to migrate even when dissatisfied | Low migration costs and widespread multi-platform use |
| Excess capital returns | Persistently above competitive normal profit after controlling for risk | Excess profits converge rapidly as technology diffuses |
The true test variable for this theory is not the number of agent users, nor model leaderboard rankings, but exit costs and sustained excess returns.
Knowledge workers: from skill owners to workflow tenants?
Industrial machines fixed part of the artisan's skill in equipment; AI may fix part of the knowledge of programmers, customer service agents, paralegals, researchers, translators, designers, and analysts in models and workflows.
The original manuscript posed a sharp proposition: knowledge workers may be progressively stripped of market control over their own skills. This stripping does not mean people suddenly "lose knowledge," but that the work outputs sellable in the marketplace increasingly depend on external models, enterprise data, and agent permissions. Workers may retain judgment, yet lose the technical and customer gateways needed to deliver products independently.
But the outcome is not unidirectional.
AI may also enable junior employees to reach the performance of experienced workers faster, allow small teams to complete tasks once only large companies could undertake, and let individuals start businesses at lower cost. The aforementioned customer service study found that less-experienced workers benefited more—evidence of capability diffusion rather than pure "skill deprivation."
The ILO estimated in 2025 that approximately one-quarter of workers globally are in occupations with some degree of exposure to generative AI, but emphasized that this is a measure of technological exposure—not a job loss forecast; most occupations are more likely to undergo task restructuring. ILO 2025
Meanwhile, a Stanford research team using U.S. ADP payroll records found that employment of young workers aged 22–25 in high-AI-exposure occupations showed a notable gap relative to less-exposed peers; the 2026 updated estimate is approximately 19%. The researchers noted that the primary mechanism may be reduced hiring rather than increased departures, but macroeconomic and industry cycles may still confound the results—one cannot simply declare "19% of young people were replaced by AI." Stanford Digital Economy Lab · OECD Employment Outlook 2026
Even more invisible is the living labor behind AI: data annotation, content moderation, quality evaluation, wafer manufacturing, data center maintenance, electricity supply, model safety testing, and business integration. The ILO notes that crowdsourcing platforms, micro-task platforms, and BPO companies undertake large volumes of data and moderation work, with many workers located in the Global South; reliable global headcount statistics remain lacking. ILO: AI and Labor
So the real question is not "whether there is still human labor in AI," but: which labor is visible and which is hidden; which labor can share in the productivity dividend and which can only accept price competition.
The productivity gains from AI can theoretically flow in at least four directions:
There is currently no evidence that the distribution across these four directions has been determined by the technology itself.
"AI colonialism" and the world market: when does dependency hold?
Lumping data colonialism, digital colonialism, surveillance capitalism, platform capitalism, techno-feudalism, and dependency theory together under "AI colonialism" causes analytical precision to be lost.
Data colonialism focuses on how everyday social life is continuously datafied and used for capital accumulation; see Couldry and Mejias. Digital colonialism focuses on transnational power asymmetries in infrastructure, standards, and platforms; see Kwet. Surveillance capitalism focuses on behavioral data, prediction, and intervention; see Zuboff. Platform capitalism emphasizes network effects, platform intermediation, and capital accumulation. Techno-feudalism argues that platforms are replacing capitalist profit logic with "cloud rent"; but Nicholas Gane's critique notes that platforms remain deeply embedded in commodity production, investment, competition, and profit maximization.
These theories can complement one another, but cannot substitute for one another.
A stricter definition of "AI colonialism" should be:
A long-term structural economic dependency: peripheral actors contribute markets, data, labor, or natural resources, but key infrastructure, intellectual property, standards, and commercial gateways are controlled by external actors; peripheral actors, lacking viable alternative paths, remain persistently confined to low-return segments, and lack equal rule-setting power.
Therefore, a Kenyan company purchasing a U.S. API does not automatically constitute colonialism; an Indian enterprise using foreign GPUs does not automatically mean it has been colonized. We must continue to ask five questions: Is it irreplaceable? Is the asymmetry persistent? Is there a lack of bargaining power? Is upward mobility to higher-value segments blocked? Is exit difficult?
The global AI value chain does exhibit significant inequality. The World Bank estimates that as of June 2025, high-income economies held approximately 77% of global co-located data center capacity; low-income economies held an extremely small share. World Bank 2025
But different countries face different situations:
United States: Home to the major cloud platforms, frontier model companies, private investment, and key software ecosystems; its 2025 AI Action Plan incorporates data centers, chips, electricity, and "full-stack AI exports" into industrial strategy. This is strong technological power, but not a single actor without internal competition. White House AI Action Plan
China: Possesses a large domestic market, model R&D, an open-source ecosystem, and compute construction capacity, and can no longer be simply placed in a "Western center—Chinese periphery" unilinear framework. 2026 policy continues to emphasize coordination across data, compute, energy, and ecosystem. China State Council action plan
European Union: Attempting to reduce dependence on single platforms through AI Factories, public high-performance compute, and the Data Act. The Data Act has been applicable since September 2025, emphasizing cloud service migration and interoperability, with switching fee rules entering a stricter phase in 2027. EU AI Factories · Data Act
India: Building shared compute through the IndiaAI Mission and cloud partnerships. India's Ministry of Electronics and Information Technology reported in its 2025–26 report that over 38,000 GPUs of computing resources have been assembled through 14 cloud partners. This shows that "sovereign AI" can be about enhancing options and bargaining power, rather than pursuing full domestic production of all hardware. India MeitY annual report
Africa and other Global South regions: Face compound constraints in electricity, connectivity, capital, skills, language data, and bargaining power. The World Bank in 2026 proposed shared, pay-as-you-go GPU infrastructure, initially focused on Africa and Latin America, seeking to avoid every country building expensive, low-utilization standalone data centers. World Bank Digital Strategy
The ILO characterizes the potential predicament of developing economies as "disruption without dividend"—suffering technological disruption first, yet unable to capture the productivity dividend. ILO 2026
This is more consistent with the evidence than dividing the world simply into "AI metropoles" and "AI colonies." The global AI system carries real dependency risks, but it is not a uniform, already-completed colonial order.
Four contradictions of AI capital accumulation
First: The cheaper intelligence gets, the more expensive the fixed capital to produce it
Frontier training, chip manufacturing, electricity, and data centers continually demand enormous investment; meanwhile, the inference price of existing capabilities may decline rapidly. This means the winner may not be the enterprise that "owns the model," but rather the one with scarce electricity, advanced manufacturing, low-cost inference, or stable distribution channels. It also means capital overcapacity and profit margin pressure are not impossible.
Second: Individual enterprises want to save on wages, yet the overall economy still needs effective demand
Enterprises reduce costs through automation, but consumer purchasing power, the tax base, and service demand still depend on wages, profits, investment, and fiscal transfers. One cannot directly infer from "AI saves labor" that "demand will inevitably collapse"; falling prices, new tasks, investment, and public policy can all form countervailing forces. But if productivity gains concentrate in capital income over the long term, distribution and demand contradictions will become more acute.
Third: Models can be open, yet workflows may be closed
Open weights make model capabilities easier to diffuse, but identity, customer relationships, long-term memory, payment permissions, and enterprise processes do not automatically open up. Competition may intensify at the model layer while reconcentrating at the agent and distribution layer. This is the most important hypothesis in this article, and the one most in need of future data validation.
Fourth: Knowledge is produced through social collaboration, yet commercial rights are carved up by different laws and contracts
Copyright, trade secrets, data protection, open-source licensing, and training contracts each have legitimate objectives, and may also form entry barriers. Abolishing all rights could harm creators; fully excluding could obstruct knowledge diffusion. The real institutional challenge is to forge a sustainable arrangement among innovation returns, creator rights, and shared social knowledge.
Three futures before 2030
The following are not prophecies, but falsifiable scenarios.
Scenario A: Open intelligence becomes a commoditized public capability
Model performance gaps narrow, inference prices continue to fall; enterprises broadly adopt multi-model routing, memory and identity are portable, agents can invoke tools across platforms, and commercial commissions are compressed by competition. Profit derives more from specific business operations, service quality, and real-world organizational capability.
If this scenario materializes, both the "model rent" and "agent rent" hypotheses will be strongly refuted.
Scenario B: Models commoditize, agent gateways close
Underlying models keep getting cheaper, but a few platforms control default gateways, identity, long-term memory, payment authorization, and workflows. Merchants continuously pay commissions to access these gateways; enterprises find migration difficult even when dissatisfied; platforms earn sustained excess returns after controlling for risk.
This would be the most powerful empirical version of "intelligence rent": the real land is not the model, but the right to govern how intelligence enters the real economy.
Scenario C: Agents open up, infrastructure oligopolizes
Agent interoperability succeeds, but advanced chips, packaging, grid-connected electricity, and cloud infrastructure remain in shortage. The profit center stays at the hardware, compute, and energy layer.
In this case, the more appropriate explanation may be infrastructure scarcity returns and technology monopoly profit, rather than lumping everything under "intelligence rent."
These three scenarios may also appear simultaneously across different regions and industries. A country's public-service AI may be highly open, while a large enterprise's ERP agent is highly closed; consumer-grade model prices may plummet while capital returns on advanced processes remain robust.
Anti-colonialism is not rejecting AI, but rebuilding the right to choose
If the core problem is exclusive control over production conditions, then the goal of institutional design should be to lower exit costs, improve entry opportunities, and redistribute productivity gains.
First, safeguard contestable infrastructure. Public compute, shared GPUs, research-institution access, and transparent procurement can enhance the bargaining power of small enterprises, universities, and developing economies. But public programs must consider utilization rates, electricity costs, and technology refresh cycles, avoiding the construction of expensive, idle facilities for symbolic "sovereignty."
Second, establish layered data rights. Distinguish among personal privacy, copyrighted works, public knowledge, trade secrets, and licensed training; support transparent licensing and enforceable compensation mechanisms, avoiding both extremes of "all data must be free" and "all data must be exclusively private."
Third, mandate or encourage genuine interoperability. Require not only model API compatibility, but also support for memory export, identity migration, authorization revocation, workflow portability, cross-agent collaboration, and payment interoperability. MCP and A2A are technical starting points, not institutional completion.
Fourth, let workers share in efficiency gains. Wage growth, profit-sharing, training, employee stock ownership, shorter working hours, and public fiscal spending can all serve as distribution channels. If the time saved by technology merely becomes higher performance-review intensity, the social welfare gains may fall far short of the technology's potential.
Fifth, competition regulation should focus on default gateways and exit costs. Scrutinize agent self-preferencing, exclusive tool access, data egress barriers, non-portable memory, and opaque merchant commissions—rather than merely comparing benchmark scores across models.
The common standard for policy should be: Does an ordinary person, a small enterprise, a research institution, or a developing economy have a realistically viable alternative?
The three strongest questions from opponents
Objection 1: AI companies bear real R&D risk—on what grounds can high profits be called "colonial"?
This is a valid criticism. Innovation enterprises must recoup R&D and capital costs; high returns may also compensate for failed projects and risk. This article therefore does not automatically treat high gross margins, high valuations, or API fees as rent. Only when sustained excess returns after risk adjustment stem primarily from exclusive control and entry-blocking does the rent judgment carry force.
Objection 2: Open-source models and price wars are democratizing knowledge capabilities—where is the "enclosure"?
This too is valid counter-evidence. Model capabilities can be replicated, unlike land by nature. This article confines "enclosure" to situations where producers lose key conditions and are forced into dependency, and explicitly acknowledges that open models may prevent this process. The so-called "age of AI colonialism" is therefore not a historical inevitability, but a risk structure that requires testing.
Objection 3: The Global South buying better foreign technology may be more advantageous than developing from scratch—why call it dependency?
One cannot judge dependency solely by technology imports. We must compare total cost, service quality, technology transfer, employment, tax revenue, bargaining power, migration cost, and local industry upgrading. Efficient international division of labor can improve welfare; only when key rights are persistently asymmetric, exit is difficult, and upgrading channels are blocked does "colonial dependency" acquire analytical meaning.
These objections are not straw men the article must defeat, but the conditions under which this theory remains falsifiable.
A terminal thought experiment: when cognitive labor becomes abundant
Suppose that by 2030, the marginal cost of general text, code, analysis, and basic design has become extremely low. Enterprises no longer earn sustained excess profit merely for "being able to write a report," and individuals can invoke high-quality models at low cost.
What remains scarce?
Perhaps electricity and land, advanced manufacturing, trusted identity, real-world execution capability, industry licensing, customer relationships, organizational trust, payment authorization, and the ability to convert advice into accountable action.
This means AI may make a portion of cognitive labor abundant, yet will not automatically eliminate the scarcity of physical resources, social relationships, legal rights, and organizational capability. The real contest will shift from "who owns intelligence" to "who has the right to make intelligence act, transact, represent others, and capture the returns generated by those actions."
This is the core worth preserving in "intelligence rent."
Technology itself does not predetermine whether AI becomes a public productive force, a competitive commodity, or a new rent-extraction machine. What determines the outcome is ownership, exit rights, market structure, and benefit distribution.
The insight from Marx most worth carrying into the AI era is not that "machines will eventually replace humans," nor that "every technological revolution produces new landlords."
It is a method: put the machine back into the relations of production.
After intelligence becomes cheap, the most important question may no longer be "who owns the smartest model," but:
Who controls the channels through which intelligence enters the real economy? Who can exit? Who bears the costs? Who ultimately captures the surplus?
Fact-check and original manuscript revision log
| Easily misleading claim in the original manuscript | Strict formulation adopted in the revised version |
|---|---|
| "AI companies privatize social knowledge, so every API call is rent" | Modelizing knowledge is not the same as exclusive privatization; API fees include real service costs, and rent requires evidence of exclusive control and sustained excess returns |
| "Data scraping is primitive accumulation" | The primitive accumulation analogy additionally requires evidence that producers lost production conditions and were forced to depend on new owners |
| "AI created massive value" | Modern value added, productivity, and new value in the Marxian labor theory of value are distinct concepts |
| "The top three cloud providers hold 63% of the AI market" | Synergy's 63% refers to Q2 2026 global cloud infrastructure services revenue share |
| "Tech companies already spent $700 billion in CapEx in 2026" | The IEA figure is a forecast magnitude for the combined 2026 capital expenditure of five major tech companies, not an audited full-year actual |
| "Over 90% of frontier models are developed by industry" | Stanford 2026's 91.6% refers to the industry share among notable AI models in its 2025 sample |
| "AI raised customer service productivity by 14% / 15%" | The early NBER version reported approximately 14%; the 2025 QJE journal version, based on 5,172 agents, reported approximately 15% |
| "A quarter of jobs will be replaced by AI" | The ILO states that one-quarter of workers are in occupations with some GenAI exposure—by no means that one-quarter of jobs will disappear |
| "Agents have already formed intelligence rent" | Agent payment and interoperability ecosystems are forming; persistent gatekeeping rent remains a hypothesis to be tested |
| "Using foreign cloud or models means being colonized" | Requires evidence of long-term structural dependency, asymmetric rule-setting power, and lack of a viable exit path |
Research method, evidence levels, and limitations
This article classifies its material into three types: facts (verifiable through company financial reports, official statistics, papers, and regulatory filings); theoretical explanations (organizing facts using Capital or modern industrial organization theory); and predictive hypotheses (agent gatekeeping power, 2027–2030 scenarios). The three are not interchangeable.
Industry concentration is typically measured by revenue or capacity and does not represent profit margins; experimental effects in papers cannot be automatically extrapolated to all occupations; forecast capital expenditure and energy demand are not realized outcomes; U.S. copyright cases do not equate to uniform global legal rules; and "economic rent" is difficult to estimate directly in the absence of comparable capital cost, risk, and quality data.
The research materials for this article are current as of November 18, 2026. It is a DeepDive draft for editorial review, not a published academic paper.
Primary sources and key theoretical literature
- Karl Marx, Capital, Volume I: Relative Surplus Value · Machinery and Large-Scale Industry · Primitive Accumulation.
- Karl Marx, Capital, Volume III: Average Profit · The Tendency of the Rate of Profit to Fall · Rent.
- Nick Couldry & Ulises A. Mejias (2019), Data Colonialism.
- Michael Kwet (2019), Digital Colonialism: US Empire and the New Imperialism in the Global South.
- Shoshana Zuboff (2015), Big Other: Surveillance Capitalism.
- Nicholas Gane (2024), Capitalism is Capitalism, Not Technofeudalism.
- Stanford HAI, 2026 AI Index Report; 2025 AI Index: R&D and Inference Costs.
- Synergy Research Group (2026-07-30), Q2 Cloud Market Passes $143 Billion.
- International Energy Agency (2026), Key Questions on Energy and AI.
- NVIDIA, FY2026 Financial Results; TSMC 2025 Annual Report; ASML 2025 Annual Report.
- Erik Brynjolfsson, Danielle Li & Lindsey Raymond (2025), Generative AI at Work, QJE; early NBER version.
- Zheyuan Cui et al. (2025), The Effects of Generative AI on High-Skilled Work.
- Anders Humlum & Emilie Vestergaard (2025), Large Language Models, Small Labor Market Effects.
- Stanford Digital Economy Lab (2026), Canaries in the Coal Mine?.
- International Labour Organization (2025), Generative AI and Jobs: A 2025 Update; (2026), AI and Developing Economies.
- OECD (2026), Employment Outlook 2026.
- World Bank (2025), Strengthening AI Foundations; (2026), Digital Strategy.
- UNCTAD (2025), Technology and Innovation Report: Inclusive AI for Development.
- U.S. Copyright Office (2025), Copyright and AI, Part 3: Generative AI Training.
- European Commission, Data Act Explained; AI Factories.
- India MeitY, Annual Report 2025–26.
- Anthropic (2025), MCP / Agentic AI Foundation; Linux Foundation (2026), A2A Adoption.
- Visa (2026), Trusted AI Payments Pilot; Mastercard (2026), Agent Pay for Machines.
When intelligence becomes cheap, who controls the channels through which intelligence enters the real economy?