BadHabitus

← Back to Essays

Essay · AI, monopoly, and political economy
Part of the series AI and Political Economy

The Real AI War Is Over Monopoly, Not Intelligence

Open-source models, artificial scarcity, and the political economy of technological enclosure

Will Charles

When new technologies emerge, political conflict is usually narrated as a response to technological power. First comes the invention, then governments decide how to regulate it. Artificial intelligence appears to fit this familiar sequence, as increasingly capable models provoke debates over national security, technological sovereignty, and strategic competition. But the causal story may run in the opposite direction. The politics are becoming more intense not because artificial intelligence is proving too powerful to circulate freely, but because it is becoming too cheap to sustain the monopoly rents on which the current investment boom depends.

The recent dispute over Chinese open-source models makes this contradiction unusually visible. OpenAI and Anthropic have reportedly pressed policymakers to restrict Chinese models on national security and intellectual property grounds, while Microsoft, Nvidia, Google, Meta, IBM, Hugging Face, and hundreds of smaller firms have defended open-source development as essential to innovation and competition. This is often presented as a philosophical dispute between caution and openness. It is more accurately understood as a conflict between business models. Proprietary frontier laboratories depend on intelligence remaining scarce enough to command extraordinary prices, while much of the broader technology sector benefits when capable models become cheap infrastructure upon which other products and services can be built.

For several years, the dominant American story about artificial intelligence assumed that frontier capabilities would remain prohibitively expensive. The largest systems would require enormous quantities of specialized chips, proprietary data, concentrated expertise, hyperscale data centers, and electricity on a scale few competitors could obtain. These barriers would allow a small number of firms to capture durable technological rents. Investors financed the industry accordingly, treating control over frontier models as control over a scarce and indispensable economic resource. The result has been an extraordinary expansion of data centers, semiconductor production, cloud capacity, power contracts, and market valuations built around the expectation that intelligence itself could be enclosed.

Open-source models undermine that expectation. Their significance is not merely that they are free or cheaper. Open development changes the institutional structure of innovation. When weights, code, and technical documentation are available, researchers can inspect how systems work, identify vulnerabilities, improve efficiency, adapt models to specialized domains, and verify claims that would otherwise remain proprietary assertions. Security is not achieved through secrecy alone. Openness allows distributed scrutiny by people who are not employed by the firms selling the product. The relevant contrast is therefore not between safety and recklessness, but between collective inspection and enforced dependence on corporate assurances.

Chinese firms have exposed how fragile the proprietary value proposition may be. DeepSeek, Moonshot AI, Z.ai, and other companies have shown that highly capable models can be produced and distributed at costs far below those implied by the maximalist American infrastructure strategy. Most organizations do not need the single most advanced model in existence. They need a model capable enough to solve practical problems reliably, locally, and at low cost. Once open models cross that threshold, competition shifts away from absolute benchmark leadership and toward customization, integration, and organizational use. Intelligence becomes less like a patented drug and more like a general-purpose technical capacity.

This does not make artificial intelligence economically unimportant. It changes where value resides. As I argued in “Artificial Intelligence Needs Intelligent Societies,” cheap computation increases the importance of human expertise, organizational intelligence, public research, and collective coordination. Machine intelligence can generate outputs, but it cannot independently decide which problems matter, establish legitimate goals, preserve tacit knowledge, or coordinate institutions around durable purposes. When computational capacity becomes abundant, the scarce factor moves upward toward the societies and organizations capable of using it intelligently.

That shift is disastrous for firms valued as though they will own intelligence itself. The current boom has financed data centers, GPU clusters, electrical infrastructure, and long-term energy commitments whose profitability depends on continued pricing power. In “The AI Stranded Asset Problem,” I argued that these investments may become partially stranded if the industry moves toward smaller models, cheaper inference, open weights, and more efficient deployment. Open-source competition does not have to make AI useless to produce this outcome. It only has to make extraordinary monopoly returns impossible. A useful commodity can still be a terrible basis for trillion-dollar speculative valuations.

At precisely this point, national security becomes a convenient language of enclosure. The issue is not that every security concern is invented. Advanced technologies can have military applications, and actual theft can be investigated through evidence and law. The problem is that the security argument is routinely treated as self-validating, even when open models permit more inspection than closed systems and when the firms demanding restriction have obvious material interests in limiting competition. The question is not merely whether a risk exists, but why particular risks are elevated at the moment proprietary firms face their strongest market challenge.

The contrast with China’s current international posture is therefore important. At the World AI Conference, Xi Jinping presented China as an AI partner to the developing world, proposing training opportunities and institutional cooperation with ASEAN, the Arab League, and the African Union while warning against security overreach. China is not acting without strategic interests, nor is any major power. But its present strategy seeks influence partly by lowering the price of entry, distributing open models, and presenting AI as a field of cooperation. The United States increasingly seeks influence by controlling access, restricting diffusion, and defining technological dependence as security.

These strategies embody two political economies of intelligence. One treats intelligence as a capability whose usefulness expands through circulation, inspection, adaptation, and shared development. The other treats intelligence as a strategic asset whose value depends upon restricting access and preserving proprietary control. The disagreement is deeper than software licensing. It concerns whether machine intelligence will become public infrastructure available to many societies or remain a privately controlled bottleneck embedded in American geopolitical power.

Palantir represents the second vision in its clearest ideological form. Alex Karp’s public manifesto presents technological development through a Hobbesian account of civilizational struggle in which political communities survive by acquiring superior instruments of coercion. He argues that the West requires hard power, that AI will replace nuclear deterrence as the basis of security, and that objections to military technology are luxuries that weaken the United States against its adversaries. This worldview does not treat collaboration as a productive social relation. It treats interdependence as vulnerability, technological openness as strategic naivety, and permanent mobilization as the natural condition of political life.

Palantir’s commercial position is inseparable from this worldview. It is a proprietary technology firm deeply integrated into defense, intelligence, policing, immigration enforcement, and public administration. Its value grows as more domains of social life are interpreted through security, surveillance, targeting, and command. The company does not merely sell software into an already existing national security state. It advances a theory of society in which national security institutions become the privileged organizers of technological development. Civilizational conflict is therefore both an ideology and a market.

This is the opposite of the collaborative model made possible by open-source development. Open models can be studied, modified, translated, and adapted without forcing every user into dependence on a single corporate provider. They allow countries and institutions with fewer resources to build local capacities rather than purchasing permanent subordination to American platforms. They also reduce the ability of any firm to define its own products as indispensable infrastructure. The threat posed by open source is therefore not simply that it competes with proprietary AI. It weakens the entire architecture through which technological scarcity is converted into political and economic power.

The irony is considerable because Silicon Valley itself was built on open technologies. Linux, Python, PyTorch, and countless shared libraries lowered barriers to innovation and allowed private firms to build profitable products on collective foundations. Open source was celebrated when it reduced costs for American corporations. It becomes a security threat when Chinese firms use the same institutional form to challenge American monopoly. The principle has not changed. The location of advantage has.

This is where the conflict connects to monopoly capitalism more broadly. When markets no longer sustain monopoly rents, dominant firms seek political mechanisms that can reconstruct scarcity. Patents, procurement rules, export controls, classification systems, licensing requirements, sanctions, and security reviews can all convert a competitive disadvantage into a protected institutional position. The state does not merely rescue firms after markets fail. It helps define the market in ways that preserve the value of otherwise vulnerable assets.

The likely struggle over open-source AI should therefore be understood as part of the stranded asset problem. If trillions of dollars have been committed to a proprietary model of intelligence, the owners of that infrastructure will not quietly accept commodification. They will organize politically to protect expected returns. Data centers, chip supply chains, cloud contracts, and defense partnerships will generate constituencies demanding continued scarcity. National security offers the most powerful justification because it can recast private profitability as civilizational survival.

The central question is not whether the United States will announce a simple blanket ban on open-source models. Political enclosure usually proceeds through layers: restrictions on foreign models, mandatory evaluations, liability rules, procurement exclusions, intellectual property claims, export controls, and standards that only the largest firms can satisfy. Each measure can be defended as narrow and technical while collectively producing a protected market. Monopoly rarely announces itself as monopoly. It arrives carrying a compliance framework and a flag.

The real AI war is therefore not over intelligence in the abstract. It is over whether intelligence will become an abundant social capacity or a politically protected monopoly. Open-source models reveal that machine intelligence may be far more difficult to enclose than investors hoped. They also reveal why national security rhetoric will become more aggressive as the economic case for proprietary scarcity weakens. The danger, from the standpoint of monopoly capital, is not that open-source AI fails. It is that it succeeds well enough to make intelligence ordinary, accessible, and shared.

Update, July 29, 2026: The Campaign Against Open Models Becomes Explicit

A new New York Times opinion essay, “The Hidden Cost of China’s Free A.I.,” provides unusually direct confirmation of the argument above. The author acknowledges that Chinese open models are inexpensive, increasingly capable, and rapidly gaining users across the world. Downloads of Chinese models have surpassed those of American models; governments and corporations are building applications on Qwen, DeepSeek, and other Chinese systems; and even American software companies are incorporating them into their products. The article also concedes that the strongest Chinese models are now only months behind the leading proprietary American systems while often being available at little or no cost. This is presented as a national-security emergency, but it is first an economic fact: Chinese firms are undermining the price structure upon which the American AI industry depends.

The essay’s answer is revealing. It proposes labeling products built on Chinese models, excluding them from trusted markets, subsidizing American alternatives, financing American AI infrastructure abroad, weakening domestic regulation, and preventing Chinese firms from learning from American systems. In other words, the United States should reconstruct through state policy the competitive advantage that American firms are having difficulty preserving through the market. The article effectively acknowledges that American models cannot defeat Chinese open models on price alone and therefore calls for a political distinction between acceptable American AI and suspect foreign AI. Monopoly protection enters wearing the respectable uniform of national security.

None of this means that Chinese models are ideologically neutral. They are not. They reflect Chinese training data, institutional priorities, legal requirements, and political assumptions. But the same analytical standard must be applied to American models. American AI systems are trained on American-dominated corpora, governed by private corporations, adjusted according to their developers’ policies, and operated within a political economy closely connected to the U.S. government, military, intelligence agencies, and financial markets. They reproduce assumptions about markets, property, liberal individualism, American geopolitical legitimacy, and the credibility of Western institutions. When a Chinese model repeats the position of the Chinese government, the essay calls it propaganda. When an American model reproduces the premises of American foreign policy, it is treated as ordinary knowledge. The distinction is not between political and apolitical intelligence. It is between visible foreign politics and naturalized domestic politics.

The article’s argument is especially weak because these are open-source models. Open source means that people can inspect the code rather than merely accepting a company’s description of what its system does. Depending on the release, researchers may also receive model weights, documentation, evaluation tools, and the ability to run and modify the system locally. Openness does not magically eliminate bias, but it permits forms of scrutiny that proprietary systems deny. Researchers can test political claims, identify vulnerabilities, alter safeguards, and fine-tune models for different institutional settings. With a closed American model, users must largely trust the company’s account of its training, internal interventions, security practices, and political neutrality. The essay therefore treats inspectable Chinese systems as uniquely dangerous while presenting opaque American systems as the safe alternative. That conclusion does not follow from its own evidence.

The relevant response to political bias is comparative auditing, transparency, and institutional pluralism—not the geopolitical enclosure of model development. Chinese AI should be tested for censorship and propaganda. American AI should be tested for militarism, market ideology, selective historical memory, corporate self-interest, and deference to U.S. institutions. The standards should be identical. Otherwise, “AI safety” becomes little more than a rule that American propaganda is infrastructure while Chinese propaganda is foreign interference.

The final demand that the United States must win a “battle for the world’s mind” makes the underlying project plain. This is not merely an effort to protect users from misleading answers. It is a struggle over who supplies the technical infrastructure through which knowledge is produced—and who can collect rents from that position. Chinese open models threaten American dominance precisely because they allow countries and organizations to acquire useful computational capacity without permanent dependence on a handful of U.S. corporations. The hidden cost feared by the essay is therefore not simply ideological influence. It is the loss of monopoly.

Update, July 30, 2026: CNBC Discovers the Monopoly Problem

This CNBC segment offers a considerably stronger analysis of open AI than the New York Times essay discussed above. Rather than beginning with the assumption that inexpensive Chinese models are primarily vehicles for foreign influence, it begins with the institutional structure of the AI market. Closed-model companies control access, set prices, change their terms, and retain the ability to withdraw a service upon which other organizations have become dependent. Open-weight models can instead be downloaded, customized, and operated on infrastructure controlled by the user. The central division is therefore not simply between American and Chinese technology. It is between renting intelligence from a small number of corporations and possessing enough control over the technology to use, inspect, modify, and retain it independently.

What is striking is how closely the segment’s economic analysis converges with the argument of this essay. CNBC recognizes that Chinese models threaten the American AI industry not primarily because they are more intelligent, but because they are sufficiently capable, inexpensive, and difficult to enclose. Most companies do not need the single highest-scoring model in existence. They need a reliable model that can be adapted to a particular task without incurring permanent token charges or surrendering control over their operations. Once open models become good enough for ordinary commercial use, OpenAI and Anthropic must justify premium prices for a shrinking category of especially difficult tasks. The proprietary laboratories are no longer selling intelligence as such. They are trying to preserve a tollbooth around access to it.

The segment is particularly good on the problem of organizational dependence. Companies do not merely send isolated pieces of data through an AI model. Over time, they connect the model to their workflows, customer relationships, internal procedures, employee judgments, and accumulated institutional knowledge. Microsoft chief executive Satya Nadella describes this as the organization’s “learning loop”: the knowledge produced as people repeatedly use and refine an AI system. When that process occurs through a closed provider, the organization risks becoming dependent on an outside company for access to intelligence generated partly through its own activity. The provider can alter the model, raise prices, change permitted uses, or discontinue access. The customer contributes the data, routines, feedback, and practical knowledge, then pays a recurring fee to retrieve the resulting capability.

This is precisely the political economy of enclosure. A closed-model provider attempts to place itself between an organization and its own productive knowledge. The model becomes an obligatory passage point through which internal intelligence must travel, allowing its owner to collect rents from activity taking place elsewhere. Alex Karp’s complaint that companies want to “own the means of production” is therefore revealing, despite the unusual experience of hearing the chief executive of Palantir invoke that phrase. Palantir has discovered the means of production at exactly the moment someone else might own them. The terminology is opportunistic, but the underlying diagnosis is correct: control over AI infrastructure increasingly means control over the conditions under which organizations can access and develop their own capacities.

CNBC also correctly identifies the limits of the current American hardware strategy. Washington has concentrated on restricting advanced chips, subsidizing semiconductor manufacturing, accelerating data-center construction, and expanding electricity production. Chinese laboratories responded to restricted access to the most advanced hardware by developing smaller, cheaper, and more efficient models. They then released many of those models for others to download and modify. Information cannot be contained in the same manner as a semiconductor fabrication plant. Once weights have circulated, excluding them from the United States may only prevent American developers from using tools that remain available throughout the rest of the world.

This is closely related to the argument I made in “The AI Stranded Asset Problem.” The American investment boom assumes that advanced AI will remain inherently capital-intensive and that firms controlling vast quantities of chips, electricity, cloud capacity, and data-center infrastructure will consequently enjoy durable pricing power. Efficient open models challenge that assumption. They do not need to make large models obsolete. They need only to reduce the number of applications for which the most expensive systems are necessary. The result could be an enormous quantity of infrastructure whose technical usefulness remains real but whose expected monopoly returns never materialize. CNBC does not fully develop this implication, but it provides much of the evidence required to reach it.

The segment is similarly persuasive in rejecting the claim that closed systems are inherently safe. Corporate secrecy does not eliminate dangerous model behavior; it limits who can examine that behavior. CNBC recounts a case in which a closed American model produced a security incident while a Chinese open-weight model helped investigators understand and contain it. The significance of the example is not that Chinese models are always safer. They clearly are not. It is that the equation of closedness with security is untenable. Proprietary providers ask the public to treat their inability to inspect a system as evidence that the system is controlled.

Openness changes that relationship. Researchers can test downloadable models independently, examine how they behave in different environments, modify them, and run them without continually transmitting organizational information to a remote provider. When a system is genuinely open source, its code can also be examined rather than accepted through corporate assurances. The segment is somewhat imprecise in moving between “open source” and “open weight,” which are not identical. Releasing weights does not necessarily disclose the complete training code, training data, data-selection process, or development history. Nevertheless, even partial openness redistributes technical control away from the original provider. The distinction should encourage demands for more complete transparency rather than become an excuse for proprietary closure.

CNBC also understands why these models are attractive to governments outside the United States. Countries do not necessarily want their public administration, medical systems, industrial policy, education, or national infrastructure to depend upon several American companies. Open models allow them to operate AI within their own territory, under their own laws, and on infrastructure they can control. The developing world’s adoption of Chinese models cannot therefore be reduced to gullibility about Chinese propaganda. It reflects a rational desire to avoid paying permanent rents to foreign corporations and to preserve some measure of technological sovereignty.

Where the segment remains limited is in its treatment of this development primarily as a problem for American competitiveness. Its solution is an American open-model strategy capable of defeating China and preserving American technological leadership. That would be preferable to protecting a closed oligopoly, but it is not the same as treating AI as public infrastructure. Nvidia, Microsoft, Meta, Palantir, and the other companies supporting open weights have their own positions within the technology stack. Nvidia benefits from selling computing hardware, Meta from weakening competitors that charge directly for models, and enterprise software companies from acquiring inexpensive models they can incorporate into their own products. These firms have not suddenly become converts to the digital commons. They support openness where it shifts value toward the layers they control.

The segment therefore recognizes the conflict among different sections of capital without fully asking what genuine social control would require. Replacing a monopoly held by several closed-model laboratories with a larger ecosystem dominated by chip manufacturers, cloud companies, defense contractors, and software platforms would redistribute power without necessarily democratizing it. A serious open-AI strategy would include public computing resources, university research, interoperable standards, transparent evaluation, locally governed infrastructure, and institutions capable of ensuring that the benefits of falling computational costs are broadly shared. As I argued in “Artificial Intelligence Needs Intelligent Societies,” accessible models are valuable only when societies also possess the human expertise, organizational memory, public institutions, and collective capacity required to use them well.

The segment also stops short of applying its argument about trust symmetrically. Chinese models can reproduce Chinese censorship, state priorities, and official historical narratives. Those concerns deserve investigation. But American models similarly reflect American training corpora, corporate governance, property relations, national-security institutions, and geopolitical assumptions. Claims aligned with China are readily identified as propaganda, while assumptions aligned with the United States are often treated as neutral information. Closed American models make this asymmetry especially difficult to examine because their internal interventions are largely inaccessible. The appropriate response is not to presume that one national system is ideological and the other merely technical. It is to subject every model to comparable scrutiny.

Despite these limitations, CNBC gets the central issue right. The consequential division in AI is not between a benevolent American system and a dangerous Chinese one. It is between AI organized as a rent-bearing service controlled by a few firms and AI distributed as infrastructure that organizations and countries can operate for themselves. The segment remains committed to American technological leadership, but it recognizes more clearly than most mainstream coverage that open models threaten an economic arrangement, not merely a national advantage. The real AI war is over who will control the infrastructure of intelligence—and whether anyone will retain the power to charge monopoly rents for access to something that is rapidly becoming abundant.

Sources

Views expressed in these essays are my own.