When a private company can hold the Pentagon’s AI access hostage, every institution in America has a sovereignty problem.

By Shaun Modi

There is an old image -- part alchemical symbol, part fever dream -- of a serpent devouring its own tail. The ouroboros. A system that sustains itself by consuming itself. It was never meant to be a business model. And yet, here we are.

The artificial intelligence industry has spent the last decade doing something remarkable: it consumed the sum total of human knowledge from books, papers, contracts, medical records, source code, creative work, and compressed it into a product. Then it built a tollbooth in front of that product and called it progress.

We are inside the ouroboros now. And the snake is very, very full.

What It Means to Own Nothing You Built

Let me describe a business model you might recognize.

A company arrives in your neighborhood. It catalogues every house on your block, from the layout, the contents, the history, the family patterns. It uses that information to build an extraordinarily sophisticated map of your community. Then it tells you that if you want directions, you need to rent the map. Every month. Through their app. On their terms.

This is, functionally, what the dominant AI companies have done.

The large language models that now underpin everything from enterprise search to government decision-making were trained on the accumulated knowledge of the internet. Which is to say, on the accumulated work of billions of people who were never asked, never compensated, and never consulted. The legal battles around this remain unresolved. But even setting aside the question of training data rights, there is a second, quieter problem happening right now.

Every enterprise workflow you run through a single model API -- every document, every query, every strategic discussion -- feeds the system. Every government use case, every internal deliberation passed through an external vendor, enriches the vendor’s understanding of you. In exchange, you get an invoice. You have handed over your institutional intelligence and received a subscription.

That is not a service. That is a landlord.

The Pentagon Could Not Get Sovereignty Over Its Own AI

The story of a leading AI company and the Department of War is the most clarifying episode in the history of this industry, and it deserves to be understood clearly.

Last July, the Pentagon awarded a contract of up to $200 million to Anthropic, which is the first time an AI company’s models were integrated into classified defense networks. A remarkable milestone. What happened next was more instructive than the contract itself.

The Department of War wanted language giving it unrestricted access to the model for all lawful purposes. Anthropic refused to remove two specific guardrails: prohibitions on fully autonomous weapons and mass domestic surveillance of American citizens. Negotiations collapsed. The Trump administration designated Anthropic, which is an American company, a “supply chain risk to national security,” a designation previously reserved for foreign adversaries. The contract shifted to a competitor willing to accept terms Anthropic would not.

I am not here to adjudicate who was right on the underlying policy questions. Reasonable people disagree. What I want you to sit with is this: the most powerful military institution in human history, with a budget larger than the GDP of most nations, could not get sovereignty over its own AI use. It had to negotiate and ultimately yield to the discretionary decisions of a private company about what tools it would and would not provide. As of this week, they are back at the negotiating table, trying again.

If it can happen to the Pentagon, it is already happening to you.

The Agentic Landlord

The next frontier of AI - what the industry is now selling as “agentic” systems -- deepens this problem dramatically. Agentic AI does not merely answer questions. It plans, executes, and iterates. It writes the report, files the brief, synthesizes the intelligence. It acts inside your systems, on your behalf.

And when it does all of that through a single vendor’s model, you have not gained a capability. You have created a dependency.

Consider what that means in practice. Your enterprise or agency has delegated meaningful cognitive work to an external model you do not control, cannot audit at the architectural level, and can be cut off from at any moment by regulatory action, by contract dispute, by the vendor’s own policy decisions. You have built your operations on rented land.

This is the ouroboros eating its own tail. The same companies whose models were built on the world’s data are now the gatekeepers of the world’s cognition. They have consumed the input and now they control the output. The snake is very full. The market has not yet noticed it is on the menu.

Why Multiple Models Are Not a Preference — They Are a Requirement

The conversation about AI in enterprise and government has been framed almost entirely around capability. Which model is most powerful? Which produces the sharpest analysis, the most defensible legal summary, the fastest code?

This is the wrong question. Or rather, it is only half the question.

The right question is: what happens when the vendor you are dependent on changes its terms, faces regulatory action, decides your use case is no longer acceptable, or becomes a geopolitical instrument? The capability question collapses without a sovereignty answer underneath it.

Data sovereignty is not an IT concern. It is a strategic imperative. It means your institution controls which models process your data, how that data flows, what is retained and by whom, and what happens when you need to switch. It means the ability to orchestrate multiple models by selecting the right intelligence for the right task from an open, competitive landscape, rather than from a menu curated by a company with a financial interest in your dependency.

No single model should be trusted with the totality of an institution’s intelligence. Not because the models are bad, but because concentration is bad. The lesson of every critical infrastructure sector such as energy, communications, finance is that single points of failure are single points of control. And control, in time, is exercised.

A Constitution Written in San Francisco

One of the dominant AI companies did something genuinely unusual: it wrote a constitution for its AI. A document of values, principles, and commitments that its model is trained to uphold. This is, on its face, commendable. Thinking carefully about the values embedded in AI systems is exactly the kind of work the field needs.

But let us be honest about what it also represents. A small group of people in a private company, unelected and accountable to no democratic process, has decided what values will be encoded into a system used by governments, hospitals, financial institutions, and defense agencies worldwide. They have appointed themselves the architects of machine ethics. And as the Pentagon story illustrates, those values include decisions about what governments are allowed to do with the tools they have already paid for.

The moment we allow a corporation to define the moral baseline of our cognitive infrastructure is the moment we have handed over something we cannot easily get back. I believe in building AI responsibly — I have spent my career doing it. But a privately authored AI constitution, however thoughtfully constructed, is not a substitute for institutional sovereignty. Not for enterprises. Not for governments. Not for anyone who holds information they cannot afford to lose control of.

The Most Extraordinary Library and Who Holds the Keys

Here is what I want people to understand, because it gets lost in the rightful critique of industry excess. What these models represent, at their technical core, is extraordinary. They are, in a real sense, the most comprehensive distillation of common human knowledge ever assembled in one place. The printing press democratized text. The internet democratized distribution. AI at its best, in its most humane form has the potential to democratize understanding itself.

But only if the keys are not held by two companies in San Francisco.

I have spent my career at the intersection of design, technology, and institutions -- from Airbnb to NASA to projects inside the Department of War. I have watched, again and again, what happens when tools people depend on are controlled by entities whose interests are not fully aligned with theirs. The pattern is not malice. It is structural. Concentration produces capture. Capture produces brittleness. And brittle infrastructure, in high-stakes environments, produces failure at the worst possible moment.

The goal is not to build the most powerful model. The goal is to build systems where the power stays with the institutions and people using them.

The Architecture of Sovereignty

What enterprises and governments actually need is not another AI company with a better benchmark score. We do not need to swap one concentration of power for another.

We need model-agnostic infrastructure. Systems that orchestrate across models, selecting and adapting AI based on the specific needs of the task, the sensitivity of the data, and the institutional requirements of the organization. We need AI that turns institutional knowledge into decision-grade intelligence without exporting that knowledge to a cognitive landlord who will rent it back to you on their schedule, on their terms. We need open standards, interoperable systems, and the kind of competitive landscape that actually produces innovation instead of the appearance of it.

The institutions that will be resilient in the AI era are not the ones that picked the best single vendor in 2024 and built their entire stack on top of it. They are the ones that maintained sovereignty over their data, their intelligence, and their choice. Those institutions will not find themselves in the position the Pentagon found itself in — dependent on a private company’s goodwill to access tools they had already paid for.

We need, in short, to stop renting.

The snake is eating its own tail. Someone needs to build the exit ramp.

Shaun Modi is the founder and CEO of Capitol AI. He previously led foundational design at Airbnb and contributed to projects at NASA and AI projects at the Department of War. Capitol AI builds model-agnostic intelligence infrastructure for enterprises and government institutions.