Table of Contents
The Concentration Problem Hiding Inside the AI Boom
In 1748, a French aristocrat published a book that quietly rearranged the political architecture of the modern world. Montesquieu argued that whenever executive, legislative, and judicial powers collapse into a single set of hands, liberty dies. Not because the person holding those hands is evil, but because the structure itself makes tyranny convenient.
Now look at your browser tabs. A handful of labs train the frontier models. The same labs write the safety guidelines, evaluate their own outputs, lobby the regulators, and sell the enterprise licenses. If Montesquieu were alive, he would not need a briefing document. He would recognize the shape of the problem within an afternoon.
This essay applies the separation of powers argument to the modern ai regulation monopoly question. The claim is simple. No single company, and no single alliance of a company and a state, should hold the combined roles of building, judging, and governing artificial intelligence. The reason is not ideological. The reason is structural, and Montesquieu spelled it out more than 250 years ago.
Montesquieu in One Page: Why Structure Beats Virtue
Most people remember Montesquieu as the man who inspired the American constitution. Fewer remember what he was actually reacting against. He had watched the French monarchy, the Ottoman court, and the Venetian republic, and he had reached a conclusion that horrified his contemporaries. Good rulers were not enough. Good laws were not enough. Only competing centers of power could keep any of them honest.
His idea in The Spirit of the Laws runs like this. Human beings holding power will extend that power until something stops them. This is not a moral failing. It is closer to a physical law, like water finding the lowest point in a garden. If you want liberty, you must build the garden so the water pools in different places at once.
Constant experience shows us that every man invested with power is apt to abuse it, and to carry his authority as far as it will go. To prevent this abuse, it is necessary from the very nature of things that power should be a check to power.
Notice what he is not saying. He is not saying rulers are wicked. He is not saying we should replace bad kings with good ones. He is saying the location of power matters more than the character of the person holding it. This is the frame we need for AI, because the current debate is stuck on the wrong axis. We keep asking whether the founders of frontier labs are trustworthy human beings. That question is almost irrelevant. The question Montesquieu would ask is whether the architecture allows any single actor to combine functions that ought to remain separate.
Three Powers, One Company
Translate the classical trio into the AI economy.
- Executive power becomes the ability to train, deploy, and update frontier models at scale.
- Legislative power becomes the ability to write the terms of service, the usage policies, the safety specs, and the internal guidelines that decide what the model will and will not do.
- Judicial power becomes the ability to evaluate the model, adjudicate misuse, ban users, and interpret whether a given output crossed a line.
In the leading labs today, all three powers sit inside one building. The company builds the model, writes the rules, and judges the outcomes. Montesquieu would recognize this configuration instantly. It is the eighteenth century French monarchy, dressed in Patagonia vests and speaking through Slack.
Why Consolidation in AI Is Not a Normal Monopoly
People sometimes wave off concentration in AI with a familiar argument. Big Tech has always been concentrated. Google won search. Meta won social. Amazon won retail logistics. What is different this time?
The difference is that previous tech monopolies controlled a distribution layer. They decided which shops you saw, which friends you heard from, which videos autoplayed next. That was already alarming enough to fill a decade of antitrust literature. But an AI monopoly does something one level deeper. It shapes the cognitive substrate that individuals and institutions use to think, decide, write, and evaluate.
When one company controls the model that half the professional world uses to draft memos, analyze contracts, screen candidates, and summarize research, you have not just created a market monopoly. You have created a monopoly on inference itself. And unlike Google search, which mostly points you to other minds, a large language model produces the reasoning directly. It sits inside the loop where judgment used to live.
This is why the standard antitrust playbook, which focuses on consumer prices and market share, misses the point. The ai regulation monopoly issue is not about whether prices are fair. It is about whether one entity holds unprecedented influence over the thinking of an entire civilization.
The Silent Compounding of Model Power
There is a further complication Montesquieu would have appreciated. Each successful generation of a model creates the data, the revenue, and the reputation that fund the next generation. Winners in this field do not just win a market. They accumulate the raw ingredients required to keep winning. Compute begets capability, capability begets customers, customers beget capital, capital begets compute.
Historically, monarchs consolidated in similar loops. Land yielded taxes, taxes bought soldiers, soldiers seized more land. Montesquieu watched that spiral in France and did not like what he saw. The modern version runs on GPUs instead of grain, but the geometry is identical.
The Four Fusions That Should Worry Us
To make the analysis concrete, consider four fusions of power happening quietly inside the current AI landscape. Each one, taken alone, is defensible. Taken together, they describe the exact configuration Montesquieu spent his life warning against.
Fusion 1: Builder and Regulator
Frontier labs increasingly write the safety standards that governments then adopt as templates. This is not conspiracy. It is inevitable. Legislators do not know how to define a dangerous capability threshold, so they call the people who do, who happen to be the people selling the models. The result is a regulatory environment where the regulated entity drafts the rules of its own regulation.
Compare this to a courtroom where the defendant writes the statute, hires the judge, and interprets the verdict. Everyone would recognize the problem immediately. Yet in AI policy, we treat the arrangement as pragmatic rather than corrosive.
Fusion 2: Model Provider and Model Evaluator
The same lab that builds a model publishes the benchmarks it performs well on, and often the safety card explaining why the model is safe enough to release. Independent evaluators exist, but they run on smaller budgets, use older benchmarks, and rarely have access to the training data. The scoreboard is being kept by one of the teams playing the game.
There is no liberty when the judicial power is not separated from the legislative and executive. If joined with the legislative, the life and liberty of the subject would be exposed to arbitrary control. If joined with the executive, the judge might behave with violence and oppression.
Substitute model release for legislation, and safety evaluation for judgment, and the paragraph reads like commentary on the 2025 AI market.
Fusion 3: Platform and Application Layer
A frontier lab that sells API access also sells the flagship consumer product, the enterprise assistant, the coding tool, and the agent framework. Third party developers build on the platform, only to discover next quarter that the platform itself now offers a competing feature, often better integrated and often free with the base subscription. This is a familiar move from the operating system wars of the 1990s. What is new is that the platform in question also decides which applications are allowed to run at all, since usage policies are enforced at the model level.
Fusion 4: Corporate and National Interest
The fourth fusion is perhaps the most delicate. Governments have begun to treat frontier labs as strategic national assets, roughly the way they treated nuclear weapons programs in the 1950s. In exchange for compute contracts, security clearances, and export protections, labs align themselves with the geopolitical interests of their host state. This can look reassuring from inside the state in question. From every other vantage point, it looks like the emergence of a state industry complex with global reach and no international check on its behavior.
What Montesquieu Would Actually Recommend
The point of invoking a philosopher is not to score rhetorical points. It is to steal a working framework. Montesquieu did not just diagnose consolidation. He proposed structural remedies. Applied to AI, those remedies suggest at least five concrete moves.
Separate Training from Evaluation by Law
The lab that trains a frontier model should not be allowed to publish the primary safety evaluation of that model. Evaluation should be conducted by an entity that is legally, financially, and reputationally independent, with binding access to weights and training data under confidentiality. This is analogous to how public companies cannot audit their own books. We did not always require external auditors. We required them after Enron, WorldCom, and a long list of quieter failures made clear that self reporting on high stakes matters is a fantasy.
Break the Platform Application Loop
Any lab above a defined capability threshold should be required to choose. Either you provide the foundational model to the market and refrain from launching competing applications on top of it, or you build applications and license someone else’s foundation. This is the AI equivalent of Glass Steagall for banks. It is not a perfect analogy, but it targets the same pathology, which is the exploitation of privileged information across roles that ought to be separate.
Multiply the Number of Frontier Actors
Montesquieu believed liberty required competing centers of power, not the abolition of power itself. Applied to AI, this means public policy should actively cultivate at least 5 to 7 independent frontier scale actors across different jurisdictions, ideological orientations, and funding sources. A world with 2 labs is a duopoly one email away from collusion. A world with 6 labs is an ecosystem, and ecosystems tolerate error better than duopolies.
Create a Genuinely Independent AI Judiciary
When a model refuses a request, deplatforms a user, or flags content as unsafe, there is currently no appeal beyond the customer service inbox of the lab. This is closer to the Star Chamber than to any modern legal system. A serious answer to the ai regulation monopoly problem includes an appeals process, external to the lab, where consequential decisions can be reviewed by parties with no financial stake in the outcome.
Constitutionalize Model Behavior
Rules that determine what models will and will not do should be published, debated, and amended through processes visible to the public. At present, the specs are written internally, updated silently, and enforced at the pace of a software deployment. Compare this with legislative change in any functioning democracy, which is slow, contested, and traceable. Montesquieu would find the current arrangement not just undemocratic but pre democratic.
Objections Worth Taking Seriously
The strongest objection to this framework is that AI is a race, and races are lost by the side that pauses to design committees. If you fragment power now, the argument goes, someone else consolidates it elsewhere and wins the future.
This is a serious concern. It deserves a serious answer. The answer is that Montesquieu was writing at a moment when France was locked in exactly this kind of race, against Britain, Prussia, and the Habsburgs. He argued, and history vindicated him, that societies with distributed power outperform societies with concentrated power over any sufficiently long horizon. Britain won because Parliament was strong, not because the king was clever. The United States surpassed European rivals because federalism prevented any single faction from freezing the system in its own image.
The race argument also assumes that speed and safety are opposed. In practice, they compound. Systems audited by independent parties fail less often, get trusted more, and reach broader deployment faster than systems whose safety is asserted by their makers. A frontier lab that welcomes external evaluation will build a moat of credibility that a self certifying competitor cannot cross.
The Other Objection: Regulation Kills Innovation
The second common objection is that antitrust action on AI would strangle a young industry. This misreads both the history of antitrust and the current state of the industry. AT&T’s breakup in 1984 did not kill telecommunications. It gave us the internet. The consent decree on IBM in the 1950s did not kill computing. It gave us the software industry. The Microsoft settlement in 2001 did not kill the personal computer. It gave us the web platforms that came after. Every one of those interventions was denounced at the time as ruinous. Every one of them, in retrospect, unlocked the next wave of value.
A well designed intervention in AI would not attack scale. It would attack fusion. The two are not the same. You can have very large firms and still separate their functions. Boeing is large. It cannot certify its own aircraft, at least not without the consequences we saw with the 737 MAX, which itself is a case study in what happens when the builder captures the regulator.
Reading the Present Through an 18th Century Lens
Return to the opening image. A French aristocrat, writing by candlelight, arguing that liberty depends on the location of power rather than the character of rulers. It is easy to think of Montesquieu as historical furniture, useful for citation in graduation speeches and nothing else.
But his framework predicts the current AI situation with unnerving accuracy. He predicted that any group holding executive, legislative, and judicial functions simultaneously would slide toward abuse regardless of intent. Frontier labs today hold all three. He predicted that the antidote was structural, not moral. Founder letters and mission statements, however sincere, cannot substitute for competing centers of power.
To become truly great, one has to stand with people, not above them.
Montesquieu would say the current architecture places a small number of firms above the users, above the regulators, and above the public that increasingly relies on their models to think. The correction is not to demonize those firms. The correction is to redesign the architecture so that no firm, however well intentioned, can occupy that position again.
The Choice in Front of Us
Every generation encounters a technology that concentrates power faster than the surrounding society can respond. The printing press did it. The railroad did it. The radio did it. The response was never to smash the technology. The response was to invent institutions that spread its benefits and constrained its concentration.
Artificial intelligence is now that technology. The question is whether the response arrives in time. The ai regulation monopoly debate will define the next 20 years of political economy, the same way antitrust defined the early 20th century and telecommunications policy defined the late 20th.
Montesquieu’s contribution, if we choose to accept it, is this. Do not wait for the current holders of power to volunteer their own restraint. Do not assume the race can be won only by consolidating faster than rivals. Build the checks now, while the field is still forming, because retrofitting them onto entrenched incumbents is an order of magnitude harder than designing them in from the start.
He would recognize the moment. He would tell us to act like founders, not subjects. And he would remind us, with the dry precision of a man who watched a monarchy calcify, that the location of power is destiny. Where we place it now determines what our grandchildren will be free to say, think, and build.
