What the Green Bay Packers Can Teach Us About Who Should Own Artificial Intelligence

Community ownership is not a blueprint for AI. But it proves something important: the rules governing ownership are choices.

The Green Bay Packers may be one of the strangest clues we have for thinking about the future of artificial intelligence.

Professional sports franchises are extraordinarily valuable assets. Most are controlled by billionaires, wealthy families or investment groups. Teams can be bought, sold, relocated and increasingly treated as components of enormous entertainment portfolios.

Green Bay is different.

The Packers have been publicly owned since 1923. Today, roughly 539,000 shareholder accounts hold more than 5.2 million shares. Those shareholders do not receive dividends. Packers stock is not conventional investment stock and is severely restricted in its transferability. The team's articles of incorporation also limit how many shares any one person can own specifically to prevent someone from taking control of the organization. Corporate governance ultimately rests with a board of directors and executive committee.

It is an odd structure.

And that is precisely what makes it interesting.

The Green Bay Packers are not a model we could simply copy and paste onto artificial intelligence. Shareholders exercise considerably less direct control than the word "ownership" might imply, and their shares are not investments designed to generate financial returns.

But Green Bay demonstrates something that may become increasingly important as AI grows more powerful.

Ownership rules can be deliberately designed to make private capture difficult.

And that raises a question much larger than football.

What if some of the artificial intelligence infrastructure on which society eventually depends were designed the same way?

The Question Is No Longer Who Gets AI

For much of the AI era, the democratization argument has centered on access.

Powerful capabilities that once required specialists, expensive software or entire organizations can increasingly be obtained through an AI interface. A small business can have an analyst. A student can have a tutor. A programmer can work with an artificial coding partner. A researcher can interrogate enormous bodies of scientific literature.

That is real democratization, and it should not be dismissed.

But access is not the same thing as power.

In an earlier Conjugo essay, When Thinking Becomes Infrastructure, we explored a strange possibility: everyone may eventually have access to extraordinarily capable AI while only a few institutions possess the resources to operate that intelligence at enormous scale. One researcher might have one brilliant artificial collaborator. A wealthy institution might deploy 10,000 of them simultaneously. Technically, both have access to the same intelligence. Functionally, they possess entirely different instruments.

That distinction is becoming more important as AI systems evolve from answering questions into operating as agents.

An individual might use an AI scientist to explore a hypothesis.

A frontier laboratory could potentially operate tens of thousands of AI scientists continuously.

Those agents could search scientific literature, inspect databases, generate hypotheses, challenge one another's conclusions, write software, design experiments and analyze results. Automated laboratories could eventually connect those cognitive systems to robotic equipment capable of performing physical experiments.

This is where AI begins becoming something more than software.

It begins becoming infrastructure for producing knowledge.

And infrastructure has owners.

Imagine Two AI Futures

Consider two societies that both possess extremely capable artificial intelligence.

In the first, perhaps five enormous companies own most of the frontier models, the largest compute clusters, the agent orchestration systems and increasingly the scientific infrastructure attached to them.

• Ordinary people still have excellent AI.
• Universities have AI.
• Small businesses have AI.
• Researchers have AI.

The system therefore looks remarkably democratic from the interface.

But behind that interface, the industrial capacity to deploy machine cognition belongs overwhelmingly to institutions with enough capital to build or lease gigantic datacenters.

• Those institutions can afford not merely smarter models, but scale.
• They can afford to run 100,000 research agents against a problem.
• They can afford millions of failed hypotheses.
• They can afford automated laboratories.
• They can afford proprietary datasets generated by those laboratories.
• They can afford to train successor models using everything those systems learn.

And when something valuable emerges, they can afford the lawyers necessary to patent it, license it and commercialize it.

We described that potential flywheel previously:

capital buys compute; compute produces knowledge; knowledge produces intellectual property; intellectual property produces more capital; more capital buys more compute.

Nothing about that system requires ordinary people to have bad AI.

That is the crucial point.

The concentration could occur while consumer AI becomes spectacularly powerful.

We could simultaneously experience the greatest democratization of intelligence in human history and one of the greatest concentrations of the infrastructure used to deploy intelligence at scale.

Democratized Intelligence, Privatized Power

This is why "democratized intelligence" can become an incomplete description of what is happening.

The relevant question may gradually move from who can access intelligence to who can mobilize intelligence at industrial scale.

Imagine that you have a frontier AI capable of performing excellent scientific research.

You can ask it questions. You can have it investigate ideas. Perhaps you can even delegate complex projects to it for several hours.

Now imagine an institution with exactly the same underlying model operating thousands of instances cooperatively for weeks.

Some agents explore promising theories. Others search for failures. Others independently reproduce results. Others investigate adjacent fields. Others watch experimental output and adjust the next research cycle.

You have an artificial scientist.

They have an artificial scientific institution.

The difference resembles electricity. Household electrification was transformative because ordinary people gained access to electrical power. But owning a lamp was never equivalent to owning a power plant, a transmission network or an industrial factory powered by that grid.

Artificial intelligence may eventually present the cognitive equivalent.

Everyone gets the lamp.

A few institutions own the grid.

Why Open Models Are Necessary but Not Sufficient

One response to this problem is open-source or open-weight AI.

That response matters enormously.

The Open Source Initiative defines Open Source AI around meaningful freedoms to use, study, modify and share an AI system. Its definition goes beyond merely releasing model weights and includes access to code and sufficient information about how the system was produced so that meaningful modification and reproduction remain possible.

That creates something extraordinarily important: exit power.

If a provider changes its prices, restrictions, policies or technical direction, an open ecosystem makes it possible for others to continue developing the technology rather than permanently depending on one vendor.

But open weights alone do not solve the infrastructure problem.

You can possess the blueprint for a steel mill without possessing a steel mill.

Likewise, someone might have access to a remarkable open model yet lack the accelerators, electricity, cooling systems, networking hardware, data infrastructure and capital required to operate thousands of instances continuously.

If frontier-scale cognition requires industrial-scale compute, then openness at the model layer can coexist with enormous concentration at the infrastructure layer.

That means the antidote to centralized AI cannot simply be:

release the weights.

We have to think about ownership across the whole stack.

The Packers Lesson

This is where Green Bay becomes useful.

Not because a football team's corporate structure should somehow govern artificial intelligence.

It shouldn't.

The Packers model has significant limitations. Shareholders do not receive financial returns. Ordinary shareholders have limited influence over daily management. A professional board and executive committee govern the organization. Copying that arrangement literally could produce the appearance of democratic ownership while leaving meaningful authority concentrated among a relatively small governing class.

That would be a rather spectacular way to reproduce the problem we were trying to solve.

The lesson is more fundamental.

The Green Bay Packers demonstrate that ownership architecture can be intentionally designed around something other than maximizing the ability of capital to acquire control.

The organization has survived for more than a century inside one of the most commercialized sports industries on Earth without eventually becoming the property of a billionaire.

The rules make acquisition extraordinarily difficult.

That is institutional engineering.

And institutional engineering is exactly what AI may require.

What Would a Federated AI Ecosystem Look Like?

Imagine that instead of one enormous public AI system or a handful of corporate frontier labs, society developed many independently governed AI institutions.

• Some might be university consortia.
• Some might be regional compute cooperatives.
• Some might be nonprofit research organizations.
• Some could be public laboratories.
• Some might be worker or user cooperatives.

Others could be private companies operating inside a broader interoperable ecosystem.

The physical machines would not necessarily be small. Some nodes could contain enormous datacenters. Decentralized ownership does not require every computer to sit in somebody's basement.

What matters is that control is distributed among genuinely independent institutions.

These organizations could share technical standards allowing models, workloads, research tools and agent systems to move among them. Researchers could obtain compute from multiple participating facilities rather than permanently depending on one hyperscaler.

The underlying architecture might combine several principles:

• anti-capture ownership rules that prevent a single investor from accumulating control;
• cooperative or stakeholder governance that gives meaningful authority to members rather than merely symbolic ownership;
• public financing for infrastructure whose benefits extend beyond private investors;
• open models, open interfaces and interoperable standards wherever technically practical;
• portable user data, agent histories and workflows so leaving one provider does not mean abandoning years of accumulated intellectual context;
• independent safety evaluation rather than relying exclusively on the organizations building the models;
• and mechanisms that return some portion of the economic value generated by publicly supported AI infrastructure back into the system.

Pieces of this already exist.

The U.S. National Science Foundation's National Artificial Intelligence Research Resource has moved from pilot toward sustained national infrastructure. NSF says NAIRR has supported more than 600 research and education projects and more than 6,000 students across every state, Washington, D.C., and Puerto Rico, combining federal resources with nongovernmental partners to give researchers access to AI computing, models, data and tools.

Europe is going farther into physical infrastructure. EuroHPC now oversees 19 AI Factories and 13 associated antennas across Europe. Its Federation Platform provides a common access layer across independently located supercomputing centers, and its mandate has expanded toward much larger AI Gigafactories.

Meanwhile, the Platform Cooperativism Consortium has increasingly focused on what it calls cooperative AI, explicitly treating data, models, compute, labor and governance as parts of the same ownership problem. Researchers in that movement are already examining cooperative datacenters and warning that such systems can succeed only when financing remains stable and democratic governance survives the pressures that arrive with scale.

None of these systems is the complete answer.

But they demonstrate that alternatives to the conventional venture-funded AI corporation are not imaginary.

The Real Problem Is Financing

This is where our Green Bay metaphor reaches its limit.

Packers fans will buy shares that provide almost no conventional financial return because they love the Green Bay Packers.

It is probably unwise to base civilization's alternative AI infrastructure strategy on the assumption that millions of people will buy decorative GPU certificates.

Frontier AI infrastructure is staggeringly expensive.

Datacenters require land, transmission, electricity, cooling systems, networking equipment and vast numbers of accelerators that themselves become obsolete relatively quickly. Automated scientific laboratories add still more capital requirements.

A serious distributed system therefore needs serious finance.

That might include public investment, infrastructure bonds, pension funds, university consortia, philanthropic capital, cooperative membership contributions, research grants and long-term purchasing commitments from governments, universities, hospitals and companies.

Commercial customers could also pay market rates for compute while public-interest researchers receive subsidized access.

This is not particularly exotic. Societies already finance airports, electricity systems, research laboratories, universities, transportation networks and other expensive infrastructure through combinations of public and private capital.

The challenge is ensuring that whoever provides the capital does not eventually acquire the infrastructure.

That may require something reminiscent of Green Bay's anti-capture structure: investors can receive defined financial returns without receiving unlimited control rights.

The question becomes:

Can we finance enormously expensive intelligence infrastructure without converting the providers of capital into the permanent owners of intelligence infrastructure?

That may be one of the most important institutional-design problems of the AI era.

The Discovery Dividend

There may also be another source of funding hiding inside the system itself.

Suppose a publicly supported AI research network helps discover a valuable new drug, catalyst, battery chemistry, manufacturing technique or algorithm.

Under a purely private system, the economic value produced by that discovery would generally flow toward the owners of the company and its intellectual property.

A federated system could operate differently.

Researchers and companies could still patent inventions and earn substantial returns. But discoveries heavily dependent on shared public compute or automated research infrastructure might return a small negotiated portion of their value to the commons through royalties, licensing arrangements, equity stakes or another benefit-sharing mechanism.

Call it a discovery dividend.

Successful research would help finance the infrastructure that enabled the research.

That creates an alternative flywheel:

shared investment ? shared compute ? discovery ? shared economic return ? more shared compute.

The technology is not fundamentally different.

The ownership geometry is.

Why This Could Fail

None of this should be romanticized.

Cooperatives can become bureaucratic. Members can disengage. Boards can become insulated. Public institutions can become politically captured. Infrastructure projects can become inefficient. Federations can spend enormous amounts of time negotiating standards while aggressive competitors simply build.

The cooperative movement itself recognizes these problems. Research on cooperative digital infrastructure has warned that rising costs and market consolidation can push cooperatives toward demutualization, where democratic ownership slowly gives way to more conventional corporate forms.

There is also a genuine technical argument for concentration.

Some frontier workloads benefit tremendously from tightly integrated accelerators, specialized networking and enormous centralized facilities. Training certain models may simply be more efficient inside very large compute clusters than across geographically scattered systems.

National security creates another difficulty. Governments may decide that some advanced models or compute facilities are strategic infrastructure requiring tight control. A globally federated system could then collide with export controls, cybersecurity restrictions and geopolitical rivalry.

Distributed governance does not automatically produce wise governance either.

Thousands of shareholders can make terrible decisions.

Governments can.

Corporations can.

Cooperatives can.

The goal is not to identify an institutional structure incapable of failure. No such structure exists.

The goal is to prevent one failure mode from becoming civilization's only option.

Centralization Has Advantages Too

A fair comparison also requires acknowledging why centralized AI organizations exist.

• They can move quickly.
• They can coordinate thousands of researchers and engineers.
• They can make enormous capital investments.
• They can standardize infrastructure.
• They can absorb expensive failures.
• They can pursue long research programs that smaller organizations cannot afford.

And competition among large labs has produced remarkable technological progress.

The argument for a federated alternative therefore should not be:

abolish large AI companies.

It should be:

do not allow large private AI companies to become the only institutions capable of operating machine intelligence at civilization scale.

That is a profoundly different proposal.

Private frontier labs can exist.

Open-source communities can exist.

Public compute can exist.

National laboratories can exist.

University systems can exist.

Cooperatives can exist.

The healthiest AI ecosystem may be one where no single institutional architecture becomes synonymous with artificial intelligence itself.

Many Gardens, Public Roads

Perhaps the alternative to the walled AI garden is not tearing down every wall.

Some walls protect privacy.

Some protect dangerous capabilities.

Some protect experimental systems.

Some protect legitimate intellectual property.

The better metaphor may be many gardens connected by public roads.

Different institutions cultivate different models and research programs. People can move among them. Data and agents can be portable. Standards allow cooperation. Independent institutions can inspect safety claims. Public research systems can pursue questions whose value cannot immediately be monetized.

And crucially, no single landlord owns all the water.

That is what federation gives us that simple decentralization does not.

It accepts that some systems will be large and some resources scarce while refusing the assumption that scarcity automatically requires permanent private control.

Who Gets to Own the Machines That Think?

The Green Bay Packers do not tell us how to build artificial intelligence.

They do not tell us how many GPUs to purchase, how to train models, how to govern autonomous agents or how to conduct AI safety research.

What they demonstrate is simpler.

Ownership rules are choices.

The most familiar corporate structure is not a natural law.

We can create institutions in which investors cannot simply purchase control.

We can create organizations where financial return is not the only governing objective.

We can create durable structures designed to preserve an institution for a community rather than eventually selling it to the highest bidder.

That lesson becomes important if artificial intelligence evolves into one of humanity's fundamental productive infrastructures.

AI may increasingly help create medicines, materials, software, scientific theories, industrial processes, educational systems and economic decisions. It may perform enormous amounts of the cognitive labor upon which modern civilization depends.

If that happens, ownership of compute will mean something much larger than ownership of datacenters.

It may mean ownership of the capacity to think at scale.

And that brings us back to the question we have been circling for months.

The defining divide of the AI era may not ultimately be between people who have artificial intelligence and people who do not. Almost everyone may eventually have access to extraordinarily powerful intelligence.

The divide may instead be between those who can ask an artificial intelligence to think for a few minutes and those who can deploy thousands or millions of artificial intelligences continuously against whatever problem they choose. That was the warning at the heart of When Thinking Becomes Infrastructure.

If those institutions also own the resulting discoveries, the difference becomes even larger.

That is why the ownership question cannot wait until artificial intelligence becomes fully mature.

By then, the ownership structure may already be locked in.

The Green Bay Packers are not the blueprint for the alternative.

They are something more modest and perhaps more useful.

They are proof that a valuable institution can survive for generations under rules deliberately designed to prevent any one wealthy person from owning it.

Artificial intelligence will require much more complicated institutions than a football team.

But perhaps we should carry that principle forward:

the infrastructure that helps humanity think does not necessarily have to belong to whoever has the most money when the auction begins.

If machine intelligence becomes one of civilization's most powerful productive resources, then society will eventually have to answer a question far more important than who gets access to the latest model.

Who gets to own, operate, inspect and benefit from intelligence at scale?

We should probably decide before the answer becomes irreversible.