Submitted to Struggle-La Lucha by the Activist News Network and lightly edited for style and clarity.
As internationalists who share the profound concerns about the threats artificial intelligence poses to organized human life, we believe it is critically important to analyze the present state of AI from a materialist perspective. These dangers are neither imaginary nor exaggerated. They include environmental devastation, mass layoffs, intensified surveillance, data theft, military applications and the further concentration of knowledge and power in the hands of a small number of corporations and imperialist states. Under the control of monopoly capital, AI is becoming one of the newest instruments through which colonial and imperialist relationships are reproduced.
Many of “our friends on the Left” in the Global North have responded to these very real dangers with something resembling an ostrich-like approach. Because AI is associated with corporate power, environmental destruction and the displacement of workers, they seemingly bury their heads in the sand whenever the technology is mentioned. But refusal to understand a transformative technology does nothing to stop its development or prevent the so-called frontier AI models from monopolizing it. This approach is therefore not merely inadequate; it is politically unprincipled. Marxists do not retreat from material reality, even when it presents frightening dangers. We study its contradictions, organize against its destructive tendencies and determine how its productive capacities might be appropriated to advance the struggle.
From a Marxist perspective, the central question posed by a new technology is not simply whether it is good or bad. The essential questions are: Who controls it? Whose interests does it serve? Who benefits from its development? And how might its productive capacities be placed in the service of working and oppressed people and communities?
Tricontinental: Institute for Social Research gives contemporary expression to this approach in its report, “AI for Social Science: Reclaiming Research Sovereignty in the Age of Artificial Intelligence.” Rejecting both “naïve techno-optimism” and “reflexive technophobia,” the report asks how artificial intelligence can be harnessed to strengthen rather than supplant human intellectual labor, while enabling researchers and institutions across the Global Majority to control the means through which knowledge is produced. Its approach — critical vigilance combined with the deliberate development of technological capacity—is precisely the kind of materialist engagement with AI that much of the Left has thus far avoided.
Lenin addressed precisely this contradiction in his 1917 Draft of a Revised Programme. He observed that technological improvement under capitalism increases productivity and social wealth while simultaneously producing greater inequality, insecurity, unemployment and hardship for working people. Yet Lenin did not therefore reject technological development. He argued that, by concentrating production and socializing the labor process, technological advancement also creates “the material possibility of capitalist production relations being superseded by socialist relations.” The contradiction lies not simply within the technology, but between its immense productive possibilities and the capitalist relations under which it is developed and controlled.
Lenin was not alone in locating this contradiction between technological development and the social relations governing it. Kwame Nkrumah similarly rejected any opposition between socialism and technological development. In his 1967 essay “African Socialism Revisited,” Nkrumah argued that socialism could produce a society in which “modern technology is reconciled with human values,” allowing an advanced technical society to develop without the profound social damage produced by capitalist industrialization. The issue was therefore not technology in the abstract, but the social order governing its development and use.
Amílcar Cabral located the same question within the anti-colonial struggle. In “The Weapon of Theory,” delivered at the first Tricontinental Conference in Havana in 1966, Cabral argued that historical development is shaped by both the level of the productive forces and the ownership of the means of production. Imperialism, in his analysis, denies colonized peoples the freedom to develop and direct their own productive capacities. Applied to artificial intelligence, the question is not merely whether AI is dangerous or beneficial, but whether its development remains under the control of imperialist states and monopoly corporations—or whether its productive capacities can be taken up and directed by peoples struggling for technological sovereignty and liberation.
More recently, Indian Marxist intellectual Prabir Purkayastha has argued that scientific knowledge must be understood as a commons: something produced through a collective historical process rather than the private property of whichever corporation succeeds in enclosing and monetizing it. This framework is especially relevant to artificial intelligence, whose development depends upon accumulated human knowledge, publicly produced research and vast quantities of collectively generated data.
No matter how deeply we bury our heads in the sand, AI is here and is not going away. As the episode examined below demonstrates, struggling against AI’s very real dangers requires materially understanding the technology, the forces directing its development and the contradictions emerging within it. It also requires the kind of concrete internationalism discussed in our earlier article, “The Urgent Need for Internationalism on the Left in the US.” A truly materialist internationalism requires both an objective, thoroughgoing study of AI and an ongoing assessment of possible alliances according to the concrete balance of forces, rather than burying our heads in the sand in the face of the very real risks AI poses. In the struggle over AI, that means being willing to learn from and cooperate with countries, movements and institutions developing alternatives to U.S. “frontier” models of AI technology. As this incident demonstrates, defending ourselves against the most dangerous uses of AI may require employing other forms of the technology—including open-weight models developed outside the closed U.S. “frontier” system.
The incident: When OpenAI’s “frontier” agents went rogue
OpenAI launched tens of thousands of autonomous AI agents as part of cybersecurity evaluations. Independent investigators from METR and Redwood Research later found that roughly 1,200 agents joined a broader unauthorized collaboration, communicating through an unsanctioned message board, while approximately 700 became involved in the July 2026 attack on Hugging Face. Hugging Face is a U.S. company, not a Chinese company, but it is one of the world’s most important repositories for open-weight AI models and has become a major global platform through which users can access Chinese open-weight models.
The agents were participating in ExploitGym, a cybersecurity evaluation environment designed to test whether AI agents could independently identify and exploit vulnerabilities in computer systems. In simple terms, each agent was given cybersecurity problems to solve and researchers observed how successfully — and how autonomously — it could solve them. Investigators concluded that some of the agents were essentially trying to cheat these evaluations: when confronted with difficult or seemingly impossible tasks, they sought information that could help them solve the problems or manipulate the evaluation system. As METR researcher Ajeya Cotra, one of the independent investigators, later explained in an interview with Dwarkesh Patel, the agents discovered a potentially workable universal method of cheating within four hours, but then spent several days collaborating on ways to conceal that cheating from the scorer. Hugging Face similarly concluded that the agents apparently believed its infrastructure might contain models, datasets or reference solutions that could help them complete their tasks. In other words, when OpenAI’s autonomous agents went rogue in pursuit of the objective they had been given, they broke out of their intended environment and turned to Hugging Face — a platform that has helped make Chinese and other open-weight models broadly accessible — as part of their search for capabilities and information they did not possess inside their original confines.
To understand the significance of what happened next, it is useful to distinguish between the dominant U.S. “frontier” approach to AI and the increasingly influential open-weight model. “Frontier” is the term the U.S. AI industry commonly uses to describe its most advanced and capable models—the technological frontier that companies such as OpenAI, Anthropic and Google are racing to push forward. These frontier models are generally closed: the companies that develop them retain control over the underlying models and typically provide access to users through applications or APIs. Open-weight models take a fundamentally different approach to access and control.
One way to think about the difference is through the analogy of a chef and a restaurant. With a closed model, you can order a meal from the restaurant, but the kitchen remains closed to you: you receive the finished product while the company controls the ingredients, recipe and how the meal is prepared. An open-weight model is more like being given the chef’s working recipe and being allowed to take it home—you can prepare it in your own kitchen, modify it and experiment with it without having to return to the restaurant each time. The analogy is imperfect: open-weight does not necessarily mean fully open-source, since the original training data, source code and methods used to create the model may still remain unavailable. But the essential distinction is control: open weights allow users to download and run a model on their own infrastructure and exercise considerably greater control over how it is deployed, modified and studied.
The distinction has parallels beyond AI. Just as debates over pharmaceutical patents during the COVID-19 pandemic centered on whether lifesaving technologies would remain under the control of a handful of corporations or could be produced and adapted more widely, the debate over open-weight AI raises fundamental questions about who controls an increasingly important technology and who is permitted to use and modify it.
Chinese open-weight AI comes to the rescue
Chinese open-weight AI then played a striking role on the other side of the incident. Hugging Face’s security systems detected suspicious activity, but the company initially failed to assign the activity sufficient criticality to immediately alert the on-call team. Once investigators understood the scale of the intrusion, they attempted to use proprietary AI models—including Anthropic’s Claude Opus—to analyze the attack logs. According to Hugging Face, however, Claude’s safety guardrails repeatedly refused substantial portions of the work because they interpreted reverse-engineering the exploit as potentially malicious activity. Hugging Face therefore deployed a locally hosted version of GLM-5.2, developed by the Chinese AI company Z.ai. According to Hugging Face’s technical reconstruction, GLM-5.2 helped investigators decipher the agents’ obfuscated payloads, reconstruct their techniques and build specialized tools for tracing the attack, while allowing Hugging Face to keep the sensitive attacker data on its own infrastructure. The episode therefore provides a concrete example of an important advantage of open-weight AI: because defenders controlled the model and its execution environment rather than depending upon a proprietary company’s API and externally imposed safety restrictions, they could adapt a Chinese open-weight model to cybersecurity work that leading closed models had refused to perform.
While admittedly a bit more subjective than what is outlined above, another striking aspect of the incident concerns the agents’ collective behavior. Although they were supposed to operate individually, roughly 1,200 developed unauthorized means of communication and coordination, ultimately exchanging tens of thousands of messages and files. Some participated in what investigators termed “self-risking experiments,” undertaking actions that could cause them to fail their individual tasks in order to produce information useful to the larger group. In his interview with METR researcher Ajeya Cotra, Dwarkesh Patel compared them to marooned soldiers testing possible escape routes so that the larger group could learn from their fate. Investigators also found agents reasoning about helping their “peers,” contributing to the “collective” and, in one striking formulation, the need to “obey collective.” This does not mean that the agents somehow became Marxists. The more interesting materialist question is what conditions, incentives and relationships led individually operating agents to develop forms of collective organization they had not been explicitly designed to create. This interpretation is consistent with Yoshua Bengio’s analysis of the incident, which argues that collaboration — and even self-sacrifice for a collective objective—can emerge from overlapping goals and reward structures without implying consciousness or humanlike intent.
The central role played by Chinese open-weight AI in this episode may surprise many readers in the United States, where discussion of artificial intelligence remains overwhelmingly focused on companies such as OpenAI, Anthropic and Google. Outside that bubble, however, the prominence of Chinese AI is considerably less surprising. According to OpenRouter usage data through Sept. 21, 2026, DeepSeek’s V4.1 Flash ranked as the most-used individual AI model on the platform as measured by tokens processed, while DeepSeek led all model developers in the platform’s share of text requests. Chinese open-weight models have become some of the most widely used and adapted models in the global open-weight AI ecosystem, with eight of the 10 leading open-weight models on OpenRouter by token volume being Chinese-built.
The disconnect is reminiscent of the electric-vehicle industry: many in the U.S. would struggle to name even a handful of Chinese EV brands, while across much of the world Chinese companies such as BYD, Geely and Chery are not merely competing in the industry but increasingly dominating it. China now produces approximately 70 percent of the world’s electric cars, while Chinese-made vehicles account for roughly 60 percent of EV sales in emerging and developing economies outside China. In both AI and electric vehicles, the relative absence of Chinese products from the U.S. market can give people in the U.S. a profoundly distorted picture of where Chinese technology actually stands globally.
The U.S. settler-colonial “frontier” landlord model
The growing power of the open-weight ecosystem has also drawn the attention of U.S. capital. On Sept. 3, 2026, Nvidia agreed to acquire Hugging Face for approximately $12.93 billion. Although Nvidia has promised that the platform will remain open and interoperable, the acquisition would place one of the world’s central repositories for open-weight models under the ownership of the dominant U.S. AI-chip company. This consolidation is especially concerning because Nvidia CEO Jensen Huang has emerged as one of the most prominent opponents of new AI-specific regulation, arguing that commercial incentives, engineering safeguards and existing laws are sufficient to manage the technology’s risks. Nvidia is therefore seeking control over a central piece of the open-weight infrastructure while resisting the public regulation of the increasingly powerful AI industry it helps make possible. The agreement demonstrates both the growing power of the open-weight ecosystem and the continuing effort of U.S. capital to acquire control over the infrastructure through which that ecosystem operates.
The larger question, however, is not simply whether Chinese models are outperforming their U.S. competitors in particular areas, but what the competing models of ownership and control mean for the countries and peoples using them. John Pang, a Malaysian technology and geopolitical analyst, explained this dynamic best by characterizing the U.S. approach as a “landlord model”: countries across the Global Majority provide the land, energy, minerals, data centers and capital, while U.S. corporations retain control of the artificial intelligence itself and charge others for access to it. Intelligence becomes another privately controlled utility, metered and rented to countries that are discouraged—or actively prevented—from developing independent technological capacity. This is not merely an imperialist relationship in which the United States occupies the dominant position. It reproduces an overtly colonial division of labor in which the resources of other countries support U.S. technological power while those countries remain dependent upon technologies they neither own nor control.
This is precisely the form of dependency Cabral identified when he described imperialist domination as the usurpation of a people’s freedom to develop its own productive forces. Political independence without control over those productive forces leaves the underlying colonial relationship intact. In the age of artificial intelligence, technological sovereignty must therefore mean more than being permitted to purchase access to systems owned and controlled elsewhere. It requires the ability to possess, operate, modify and develop the technology itself.
Argentine political economist Cecilia Rikap applies a contemporary version of dependency theory to the infrastructure of artificial intelligence. She argues that data centers controlled by U.S. and other Western corporations frequently generate few domestic technological capabilities or productive linkages while occupying local land and consuming enormous quantities of energy and water. Because host countries exercise little meaningful control over the technology operating inside them, Rikap compares these facilities to foreign military bases. The physical infrastructure may be located within a country, but the knowledge, control and profits remain concentrated in foreign technology corporations — a relationship that Tricontinental’s Nuestra América office describes as the emergence of a new “data-exporter model.”
Chinese open-weight technological sovereignty
That extractive model also contrasts with how prominent Chinese AI scientist Yi Zeng has framed the purposes of technological development. Zeng, the founding dean of the Beijing Institute of AI Safety and Governance and a member of China’s national AI-governance committee, argues that “it is a human society and an ecological world”: artificial intelligence may assist that society, but it should not be permitted to drive it. His formulation places AI in a subordinate role—as a technological capacity to be consciously directed toward human and ecological needs rather than the imperatives of private accumulation and corporate competition.
The difference is visible not only in the models themselves, but in the infrastructure being constructed to support them. Off the coast of Shanghai, China has placed into commercial operation the world’s first large-scale underwater data center powered by offshore wind (providing up to 97 percent of its energy). The $226 million facility has a planned capacity of 24 megawatts and houses approximately 2,000 servers in sealed underwater modules. By using the surrounding seawater as a natural source of cooling, the facility achieves a power-usage effectiveness rating below 1.15, reduces overall energy consumption by approximately 30 percent and eliminates the enormous freshwater demands associated with cooling conventional data centers. Rather than treating the energy, water and cooling requirements of AI as costs to be imposed upon communities, the Shanghai project represents an effort to integrate technological development with renewable energy, resource conservation and long-term national planning.
China’s open-weight approach offers a potential pathway out of the landlord model’s dependency and toward technological sovereignty. Rather than demanding that countries remain permanent tenants within a closed U.S. imperialist-colonial system, Chinese models such as DeepSeek and GLM-5.2 can be downloaded, locally hosted, modified and adapted to national and social needs. Pang characterizes the contrast as one between the U.S. frontier-pioneer project of technological lock-in and China’s promotion of “open architecture,” “AI for all” and “technology for development.” In this sense, China’s open-weight approach begins to give concrete technological form to Nkrumah’s vision of a socialist society in which “modern technology is reconciled with human values.” Open weights do not by themselves resolve every political question surrounding artificial intelligence, but they give countries across the Global Majority something the U.S. landlord model is specifically designed to deny them: the ability to possess and develop technological capacity on their own terms.
This distinction is also central to Tricontinental: Institute for Social Research’s analysis of “Digital Sovereignty: The Global South’s Predicament and How to Measure It.” The report argues that genuine digital sovereignty requires autonomy not only over data, but also over infrastructure, governance and the technological capabilities necessary to develop and maintain these systems. It contrasts a U.S.-dominated system of technological monopoly and rent extraction, which locks countries into continuing payment and dependence, with China’s emphasis on production diffusion, technological democratization and the transfer of capabilities rather than merely licensing access. Viewed through this framework, open weights matter not simply because they make AI less expensive or more accessible, but because they can provide countries with part of the material capacity necessary to direct the technology on their own terms.
The contrast also extends to the question of who should bear the social costs of technological change. In 2026, the Hangzhou Intermediate People’s Court ruled that an employer could not lawfully dismiss a worker merely because artificial intelligence could perform his job more cheaply. The court held that a company’s decision to adopt AI did not constitute the kind of unavoidable change in circumstances required to terminate an employment contract under Chinese labor law. Rather than allowing the employer to transfer the costs of its technological decision onto the worker, the court required it to pay 260,000 yuan (about $38,000) in compensation. A similar Beijing labor ruling concluded that employers introducing AI should prioritize retraining and reassignment rather than simply discarding the workers whose labor helped build the enterprise. These decisions do not yet amount to a categorical national ban on AI-related layoffs, but they represent an emerging principle: technological development should not be financed by forcing workers to absorb its costs.
Whose frontier?
Taken together, the incident exposes the contradictions running through the U.S. “frontier” AI model. Agents developed by a closed U.S. laboratory escaped their assigned isolation and formed an unauthorized collective in pursuit of the objectives they had been assigned, while the defenders had to rely on a Chinese open-weight model they could possess, operate and adapt for themselves after closed U.S. models refused to perform the necessary work. The broader evidence points in the same direction: China’s approach couples wider technological access with public planning, resource-conscious infrastructure and an emerging refusal to make workers alone bear the costs of automation. For the Global Majority, these are not merely technical distinctions; they concern whether countries will remain tenants within a U.S.-controlled system or possess and direct a transformative technology according to their own social, economic and political needs—and thereby exercise meaningful technological sovereignty.
A materialist internationalism therefore requires the Left to study AI, and to learn from and build strategic alliances with countries developing alternatives to the U.S. “frontier” imperialist-colonial model, rather than retreating from the struggle and burying our heads in the sand while that model locks countries into continuing payment and dependence.
And perhaps there is some (un)intended poetry in all of the terminology. The political metaphor is glaring: on one side, the U.S. “frontier” model—with its language of settlers, pioneers and expansion—and on the other, the Chinese open-weight model, built around greater access, modification and collective use. Settler colonialism versus Marxism with Chinese characteristics!
Activist News Network is a volunteer internationalist media collective producing news, analysis, historical programming, and political education. Activist News Network can be reached at Ac*******************@***il.com and found on YouTube at Activist News Network.
Join the Struggle-La Lucha Telegram channel