Redistributed

The dominant model of AI development follows a familiar script. Venture capital funds a startup. Engineers build a foundation model. The company scales. Returns flow to shareholders, executives, and early investors. The workers who label training data, the users who provide feedback, the communities whose culture and knowledge fill the datasets — they are inputs, not beneficiaries.

This is not the only way to build AI. It is just the way that attracts the most capital.

Cooperatives offer a structural alternative: organizations owned and governed by the people who create value within them. The cooperative model is old — older than the corporation, in fact — and it has been applied to agriculture, finance, energy, and housing. Applying it to AI is not a thought experiment. People are already doing it.

The Corporate Model and Its Logic

To understand the alternative, start with what it replaces.

Frontier AI development is extraordinarily capital-intensive. Training GPT-4 reportedly cost over $100 million. The next generation of models may cost billions. This capital comes from venture investors and public markets, and it expects returns — large returns, quickly.

The VC model creates a specific incentive structure. Build fast. Capture market share. Monetize users. Optimize for growth metrics that justify the next funding round. The technology serves the business model, not the other way around.

This structure determines who benefits. OpenAI's valuation exceeds $80 billion. That value was created by researchers, data labelers, users who provided RLHF feedback, and the vast commons of human knowledge scraped for training data. The value flows to investors, executives, and a narrow band of highly compensated engineers.

The question is not whether this model produces powerful AI. It does. The question is whether it is the only model that can — and whether the distribution of returns it produces is inevitable or designed.

Cooperative Principles, Applied to AI

The International Cooperative Alliance defines cooperatives through seven principles. Three are directly relevant to AI.

Democratic member control. In a cooperative, members govern the organization on a one-member, one-vote basis. Applied to AI: users and workers would govern development decisions — what gets built, what data is used, what safeguards are implemented, how the system is deployed. This contrasts sharply with corporate AI, where a board of directors and C-suite make these decisions with no structural accountability to users or affected communities.

Member economic participation. Members contribute to and democratically control the capital of the cooperative. Surpluses are allocated for the cooperative's development, for common benefit, or returned to members in proportion to their engagement. Applied to AI: returns from AI systems would flow to those who build and use them — not to external shareholders seeking maximum extraction.

Concern for community. Cooperatives work for the sustainable development of their communities. Applied to AI: development priorities would be set by community needs, not by what maximizes advertising revenue or enterprise SaaS contracts. An AI cooperative serving healthcare workers would optimize for clinical utility, not engagement metrics.

These are not abstractions. They describe a different ownership structure — and ownership structures determine who gets what.

Existing Models and Precursors

No large-scale AI cooperative exists yet in the formal sense. But several organizations demonstrate cooperative principles in practice, and their experience illuminates what a full cooperative model might look like.

Cooperative-Spirited Research

EleutherAI began in 2020 as a volunteer research collective on Discord. It built GPT-NeoX, the Pile dataset, and a series of open language models — demonstrating that significant AI research could happen outside corporate labs. EleutherAI is not legally a cooperative; it has no formal membership or governance structure. But it operates on cooperative principles: voluntary participation, shared resources, open outputs. Its limitations — difficulty sustaining long-term funding, reliance on donated compute — illustrate the challenges cooperatives face in a capital-intensive field.

BLOOM, trained by the BigScience project, brought together over 1,000 researchers from more than 60 countries to train a 176-billion-parameter multilingual language model. The project was coordinated by Hugging Face and funded largely through a French government supercomputing grant. BLOOM demonstrated that large-model training does not require corporate backing — but it required extraordinary coordination and a one-time public compute allocation that has not been repeated at the same scale.

The Distributed AI Research Institute (DAIR), founded by Timnit Gebru after her departure from Google, pursues community-centered AI research. DAIR's model — independent research directed by community needs rather than corporate priorities — embodies the cooperative principle of concern for community, even if its legal structure is a nonprofit rather than a cooperative.

Data and Platform Cooperatives

Data cooperatives represent a different entry point. Rather than building AI systems, they organize the people whose data feeds those systems.

Driver's Seat Cooperative collects and analyzes data from ride-share and delivery drivers — the same workers whose data Uber and Lyft use to optimize their platforms. The cooperative gives drivers collective ownership of their data and uses it to negotiate better conditions. This is the cooperative principle applied at the data layer: if your labor produces the data that makes AI valuable, you should own that data collectively.

Stocksy, a stock photography cooperative, offers an alternative to platforms like Shutterstock that are increasingly using AI to generate images that compete with their own contributors. Stocksy's cooperative structure means photographers own the platform and share in its revenue. When AI threatens to devalue creative work, cooperative ownership gives workers a governance mechanism that employment does not.

Hugging Face, while structured as a venture-backed company, operates with significant community governance elements. Its open model hosting, community-driven model development, and transparent approach to AI research have created a commons that benefits far beyond its shareholders. Hugging Face illustrates a hybrid: corporate structure with cooperative ethos. Whether the ethos survives the pressures of its corporate structure is an open question.

Historical Parallels

AI cooperatives are not inventing a new form. They are applying an old one to a new domain.

Agricultural cooperatives — Land O'Lakes, Ocean Spray, Sunkist — allowed farmers to collectively own processing, distribution, and marketing infrastructure. Individual farmers could not compete with industrial processors; collectively, they could. The parallel to AI is direct: individual researchers, workers, and communities cannot compete with billion-dollar labs. Collectively, they might.

Credit unions emerged in the 19th century as member-owned alternatives to commercial banks. They now serve over 130 million Americans. Credit unions do not maximize shareholder returns; they minimize costs for members. An AI credit union equivalent would provide AI services at cost, returning any surplus to members rather than extracting it for investors.

Rural electric cooperatives are perhaps the most instructive precedent — and a reminder that cooperative models have succeeded at massive scale before. In the 1930s, private utilities refused to extend power lines to rural America — it was not profitable enough. The Rural Electrification Administration, created in 1935, provided financing for community-owned electric cooperatives. Within two decades, rural electrification went from 10% to over 90%. Today, over 900 electric cooperatives serve 42 million Americans.

The lesson: when private capital will not serve a population because the returns are insufficient, cooperative ownership — often with public support — can fill the gap. AI cooperatives could serve communities, languages, and use cases that venture-backed labs ignore because the market is too small.

Open source software is cooperative in production but not in ownership. Linux, Apache, and Python are produced collaboratively, but they do not have members who share in economic returns. The open source model proves that cooperative production works for complex software. What it does not provide is cooperative distribution of economic value — a gap that AI cooperatives could address.

The Theoretical Foundation

Two bodies of scholarship ground the case for AI cooperatives.

Yochai Benkler's work on commons-based peer production demonstrated that decentralized, voluntary collaboration can produce complex knowledge goods — Wikipedia, Linux, academic research — that compete with or exceed corporate production (Benkler, 2006). Benkler argued that in knowledge-intensive domains, the transaction costs of markets and the bureaucratic costs of firms can exceed the coordination costs of commons-based production. AI, as a knowledge-intensive domain, fits this analysis.

Elinor Ostrom's research on governing the commons showed that communities can manage shared resources sustainably — without privatization and without state control — when they develop appropriate institutional arrangements (Ostrom, 1990). Her eight design principles for long-enduring commons institutions include clearly defined boundaries, collective choice arrangements, monitoring, graduated sanctions, and conflict resolution mechanisms. These principles map directly onto the governance challenges of cooperative AI: who is a member, how decisions are made, how contributions are tracked, and how disputes are resolved.

Together, Benkler and Ostrom provide both the production model (commons-based peer production can work at scale) and the governance model (commons can be governed without privatization). The combination suggests that cooperative AI is not just desirable but structurally viable.

The Redistribution Argument

The case for AI cooperatives is ultimately a redistribution argument.

Economic returns. In the corporate model, AI's economic value flows to shareholders and executives. In a cooperative, it flows to members — the workers who build the systems, the users who depend on them, the communities that provide the context and data. This is not charity. It is a different ownership structure producing a different distribution of the same value.

Governance power. Cooperatives give members democratic control over development priorities. What gets built, for whom, and under what constraints — these decisions are made by the people affected, not by a board optimizing for quarterly earnings. When a cooperative decides to build an AI system for a specific community need rather than a lucrative enterprise contract, that is redistribution of decision-making power.

Development priorities. Corporate AI optimizes for markets that generate revenue. This means English-language models for wealthy users get more investment than tools for minority languages, public health applications, or rural communities. Cooperatives, governed by their members, can direct development toward needs that markets ignore. This is not inefficiency. It is a different definition of what efficiency means.

The challenge to assumption. Most fundamentally, AI cooperatives challenge the assumption that frontier AI development requires venture capital and corporate structure. That assumption is partially true — massive compute requires massive capital. But it is not entirely true. BLOOM was trained with public compute. EleutherAI built meaningful models with donated resources. The assumption that only the VC-to-IPO pipeline can fund AI is a claim made by the people who profit from that pipeline. It deserves scrutiny.

Real Obstacles

Advocacy for cooperatives that ignores their challenges is not advocacy. It is wishful thinking.

Capital access. Frontier model training costs hundreds of millions of dollars. Cooperatives cannot sell equity to venture capitalists — their ownership structure precludes it. This is a genuine constraint, not a solvable inconvenience. Without alternative capital sources, AI cooperatives will operate at smaller scales than corporate labs.

Talent recruitment. Top AI researchers receive compensation packages worth millions in equity. Cooperatives cannot offer equity-based wealth creation. They can offer meaningful work, community governance, and the absence of corporate extraction — but they cannot match the financial incentives of a pre-IPO AI startup. Some researchers will choose mission over money. Most will not, until cooperatives can offer competitive compensation.

Scale economics. AI systems benefit from scale in data, compute, and user feedback. Cooperatives, by nature, may be smaller than corporations. Federated approaches — multiple cooperatives sharing resources — could address this, but federation creates coordination costs.

Coordination costs. Democratic governance is slower than executive fiat. When a cooperative must decide whether to train a model on a controversial dataset, the deliberation process takes longer than a CEO deciding unilaterally. This is a cost — and it is also a benefit, since the resulting decisions incorporate more perspectives. But in a competitive market, speed matters.

Possible Paths Forward

Several models could address these obstacles without abandoning cooperative principles.

Hybrid cooperative-corporate structures. A cooperative could own a controlling stake in a company that operates commercially, ensuring democratic governance while accessing conventional capital markets for non-controlling investment. This model exists in other sectors — the Mondragon Corporation in Spain operates a network of cooperatives with combined revenues exceeding $12 billion.

Public funding for AI cooperatives. Just as the Rural Electrification Administration funded electric cooperatives, governments could fund AI cooperatives — providing compute access, research grants, and infrastructure support. The NAIRR (National AI Research Resource) in the United States is a step in this direction, though it supports academic researchers rather than cooperatives specifically.

Cooperative consortiums for shared compute. Multiple cooperatives could pool resources for compute infrastructure, achieving scale economies that individual cooperatives cannot. This is how agricultural cooperatives work: individual farms are small, but their collective processing facilities are large.

Revenue from services. Cooperatives need not depend on venture capital if they generate revenue from services. An AI cooperative could sell API access, consulting, or customization services — returning surplus to members rather than shareholders. The constraint is building sufficient capability to compete, which brings the discussion back to capital and talent.

What You Can Do

AI cooperatives will not emerge from theoretical arguments. They will emerge from people building them.

Support existing cooperative-spirited initiatives. Contribute to EleutherAI, DAIR Institute, or open-source AI projects. Use Hugging Face's open models rather than proprietary alternatives when feasible. Support data cooperatives in your industry.

Advocate for public funding alternatives. The biggest structural barrier to AI cooperatives is capital. Public funding — compute grants, research support, infrastructure investment — can fill this gap. Advocate for expanded NAIRR funding, public compute infrastructure, and government support for cooperative AI development.

Participate in open-source AI development. Open-source AI is not identical to cooperative AI, but it builds the commons from which cooperatives can draw. Contributing to open models, open datasets, and open tools expands the foundation for cooperative development.

Organize in your sector. If AI is transforming your industry, consider how a cooperative might serve your community's needs better than a corporate vendor. Data cooperatives, platform cooperatives, and worker cooperatives all offer models that can be adapted to specific contexts.

Challenge the assumption. The most powerful thing cooperatives do is demonstrate that alternatives exist. Every successful cooperative challenges the claim that the corporate model is the only model. Building alternatives is itself a political act.

Sources

Benkler, Yochai. The Wealth of Networks: How Social Production Transforms Markets and Freedom. Yale University Press, 2006. Publisher
Ostrom, Elinor. Governing the Commons: The Evolution of Institutions for Collective Action. Cambridge University Press, 1990. Publisher
Schneider, Nathan. Everything for Everyone: The Radical Tradition That Is Shaping the Next Economy. Nation Books, 2018. Publisher
Scholz, Trebor. Own This! How Platform Cooperatives Help Workers Build a Democratic Internet. Verso Books, 2023. Publisher
International Cooperative Alliance. "Cooperative Identity, Values & Principles." ICA
BigScience Workshop. "BLOOM: A 176B-Parameter Open-Access Multilingual Language Model." 2023. Hugging Face

General

AI Cooperatives: Alternatives to Corporate Models