There is a scene that repeats itself in litters raised under scarcity.
The food comes down. The largest puppy gets there first.
At the beginning the difference is not dramatic. A little more weight. A little more reach. A little more confidence in the shove of the shoulder. But meal after meal the advantage compounds. The biggest pup eats first and longest. The smaller ones do not simply get less. They get what is left after the hierarchy has already done its work.
Soon the gap stops being a matter of temperament and becomes a matter of structure. The dominant puppy grows stronger because it eats first. It eats first because it is stronger. The weaker ones do not lose because they are lazy or stupid or morally inferior. They lose because in a scarce environment, access itself becomes destiny. By the time they are old enough to fight harder, their bodies have already been written by deprivation.
That is beginning to happen in artificial intelligence.
For years, many people liked to tell a more democratic story about AI. The technology would lower barriers. A student with a laptop would gain leverage once reserved for large firms. Smaller countries could leapfrog older forms of industrial dependency. A smart founder in Nairobi, São Paulo, or Dhaka could use the same models as a giant company in San Francisco. Intelligence would become more abundant, and abundance would flatten hierarchy.
Some of that story was real. Cheap and widely available models did widen access. Open tools made experimentation possible far from the richest labs. A small team could suddenly write software faster, analyze documents more deeply, and automate work that once required a department. For a moment it felt possible that AI might behave less like oil and more like literacy.
But that was the early bowl, before the real scarcity asserted itself.
The scarce thing was never intelligence in the abstract. It was always the calories that turn intelligence into durable advantage: the compute, the power, the training runs, the proprietary data, the research talent, the fabs, the cloud contracts, the transmission lines, the political cover, the export licenses. Those are not evenly distributed. They sit in a very small number of hands, in a very small number of places, backed by a very small number of states.
Now the biggest puppies are not only eating first. They are using their extra calories to redesign the feeding system itself.
This is not abstract. When the United States moved in 2022 and again in 2023 to restrict the most advanced chips from reaching China, the leading chipmaker designed a slightly slower part, the H800, to sit just under the threshold, and when the threshold moved, that part was blocked too. The lesson was not who won a single round. It was that the bowl now has gatekeepers, that the lines are drawn by a handful of states and firms, and that access to the highest tiers of compute has become an instrument of policy rather than a feature of an open market.
That is what makes this moment more dangerous than a standard technology race. If frontier models can materially accelerate research, coding, experimentation, and even the design of their own successors, then the lead is no longer linear. It begins to compound. The lab with the best model gets more from every unit of compute, hires more of the scarce talent, builds better internal tools, improves faster, and then uses that improvement to pull further away from everyone else. A lead becomes a gulf. The gulf becomes an order.
This is the part that should worry countries that once imagined AI as a great equalizer.
A small or middle power can live with a world in which others are ahead. Most countries already do. What is much harder to live with is a world in which the leading systems sit behind alliances, export controls, trusted-access programs, and compute bottlenecks that no amount of ordinary industrial policy can quickly overcome. In that world, the promise of AI is still global, but the command of AI is not. The rest of the world is invited to build on top of someone else’s cognition, under terms that can change without warning.
Dependence is only the visible layer. The deeper problem is developmental lockout. If the most capable models are used to build still more capable models, then latecomers are not merely behind in output. They are behind in the machinery that determines the next round of output. They are the underfed pups trying to catch up after the hierarchy has already shaped their bones.
The economics of the bowl
This is not only a story about technology. It is a story about political economy.
One useful framework is cumulative advantage, what the sociologist Robert Merton called the Matthew effect, sometimes summarized in plainer language as the rich getting richer. In markets with strong increasing returns, an early lead does not fade as competition arrives. It widens. Better products attract more users. More users generate more data, more revenue, and more complementary investment. Those improvements feed back into the product and reinforce the lead.
AI has several layers of increasing returns stacked on top of one another.
- The best models attract the most users.
- The most users generate the most commercial integration.
- The most commercial integration funds the most compute.
- The most compute supports the most ambitious training and inference cycles.
- The most capable systems then assist in producing the next generation of capability.
That is cumulative advantage with a laboratory attached to it.
The numbers track the logic. By the most careful independent estimates, the cost of the largest training runs has grown by roughly two and a half times a year, climbing from a few million dollars to tens and then hundreds of millions in the span of a few model generations, with credible projections that the leading runs will pass a billion dollars before the decade is out. At those levels the frontier is not a place a clever newcomer wanders into. It is a place a balance sheet grants entry to.
Another useful framework comes from industrial organization, the branch of economics concerned with how markets concentrate. When a market depends on huge fixed costs, scarce inputs, and infrastructure that only a few firms can finance, competition often narrows toward oligopoly. In ordinary language, that means a small club ends up controlling the field because the cost of entry becomes too high for everyone else. Advanced AI increasingly looks like that kind of market. A narrow set of hyperscalers, chip firms, frontier labs, and states now sit near the choke points of progress.
A third framework is dependency theory, developed by economists and political theorists such as Raúl Prebisch, Fernando Henrique Cardoso, and Immanuel Wallerstein, who observed that peripheral countries often do not simply fail to industrialize on their own timetable. They become structurally reliant on the core for the terms of development itself. They buy tools made elsewhere, run systems governed elsewhere, and adapt their own institutions around choices made in distant capitals. AI now threatens to reproduce that pattern at the level of cognition. It offers participation without sovereignty.
That is the hidden cruelty of the equalizer story. A technology can be broadly available and still intensify hierarchy if the layers that matter most remain scarce and centrally controlled.
What monopoly puppies look like in the future
It is tempting to imagine monopoly in the old way: one company controls one market. That is too small for what may be forming.
The future monopoly puppy may not look like a search engine or a software suite. It may look like a vertically integrated stack that controls the model, the chips, the cloud, the distribution rails, the safety regime, and the orbital or terrestrial infrastructure that keeps the entire system fed.
That is why the thought experiment about space infrastructure matters. The world has already seen what happens when private infrastructure becomes geopolitically decisive. In 2022 a single satellite internet network, built and owned by one company, became essential to a country at war, and its owner could decline a request to extend coverage over contested waters because activating it might have enabled a strike he did not wish to be party to. A commercial marvel had become a battlefield lever, and the decision sat with a private individual rather than any government.
This is already the direction of travel, and it is becoming more physical, not less. The contest at the frontier is shifting from cleverness to procurement, from who holds the best algorithm to who has secured the most electricity, the firmest long-term power contracts, and the most reliable supply of chips. Large technology firms have become among the largest buyers of long-term power in the world. They sign multi-gigawatt agreements, revive retired nuclear plants, and finance new reactors to feed clusters whose appetite is measured not in processors but in the output of power stations. The binding constraint is no longer ingenuity. It is megawatts, transmission, cooling, and the supply chain that delivers them. The bowl, in other words, is becoming a question of grids and substations as much as model weights.
Extend that logic one layer upward. Imagine a company that becomes the de facto highway to AI compute itself, not only in data centers on the ground but through launch, satellite relays, orbital power, cooling, or space-linked compute infrastructure. The firm that owns the road does not have to own every truck to govern the movement of goods.
Applied to AI, that means the company or alliance that controls the delivery system for compute could become as important as the lab that trains the model. The bowl is not only the model weights. It is the pipeline that determines who eats, how often, at what price, and under whose permission.
That is one possible future shape of monopoly. Another is subtler. A small number of model families become so good, so integrated, and so embedded in business workflows that switching away from them starts to feel impossible. Not because alternatives do not exist, but because the cost of retraining people, tools, processes, and institutional habits becomes too high. At that point monopoly is no longer only about price. It is about memory. The winning systems become the place where the world’s operational habits are stored.
Who loses control
The obvious answer is that smaller countries lose control. That is true, but not sufficient.
Citizens lose control when a handful of opaque systems mediate their access to knowledge, work, insurance, credit, and public services. Regulators lose control when the frontier advances inside private labs faster than public expertise or public compute can keep pace. Companies lose control when their workflows become dependent on upstream model providers who can change access, pricing, retention, or policy overnight. Even labs lose a measure of control once they become strategic infrastructure for states that will not tolerate purely private discretion in moments of crisis.
There is also a quieter loss of control at the level of imagination. When one or two model families become the default interface to reason, summarize, code, and decide, they begin to shape not only what people can do, but what people think is normal to do. They become rating agencies of reality. Their biases are not merely ideological in the narrow sense. They are architectural. They prioritize some forms of knowledge, some kinds of evidence, some tempos of decision, some norms of safety, and some distributions of access over others.
This is one reason the status quo deserves critique even from those who admire the technical achievements. The system is not only producing better tools. It is producing a world in which the right to scale those tools, refine those tools, and govern those tools is narrowing around a few actors whose incentives are not automatically aligned with global flourishing.
The question of jailbreaks
Most people hear the word jailbreak and think of a clever prompt that tricks a model into answering a forbidden question. That matters, but it is not the most interesting layer.
The more consequential jailbreaks are institutional. They happen when states, firms, labs, or brokers decide that the formal rules around access are too constraining to live with. A government cut out of a trusted-access arrangement starts building procurement workarounds. A sanctioned actor routes through friendlier jurisdictions. A company denied dangerous capabilities seeks them through an intermediary. An excluded region turns to less aligned systems because those are the only systems it can fully control.
In that world, jailbreaks are not only about bypassing a safety filter. They are about bypassing a political and economic order.
That raises a difficult ethical question. Is there such a thing as ethical sneaky jailbreaking?
The answer should not be an easy yes. A world of unbounded circumvention is not a world of justice. It is often a world in which the ruthless win. But the answer cannot be an easy no either. History is full of moments when rigid control over foundational infrastructure protected incumbents more than humanity. If a small country, a public university, or a critical humanitarian institution is structurally excluded from capabilities needed for defense, resilience, or development, then some forms of controlled circumvention begin to look less like criminality and more like self-preservation.
It is worth being precise about what is and is not being defended here. There is a real difference between circumventing a safety constraint, which removes a guardrail meant to prevent harm, and circumventing an exclusion constraint, which removes a barrier meant only to preserve advantage. The first is rarely justifiable. The second becomes harder to condemn the more arbitrary the exclusion is. The case for circumvention is a case about access and legitimacy, not about bypassing the protections that keep dangerous capability in check.
That does not justify recklessness. It does suggest that a legitimacy gap opens whenever the official system offers dependence without recourse. The tighter the bowl is guarded, the more morally ambiguous the scramble around it becomes.
Who holds the bowl
A responsible owner looking at a hungry litter does not congratulate the biggest puppy for winning. A responsible owner notices that scarcity is teaching the wrong lesson and steps in. The bowls are separated. The portions are measured. The smaller pups are protected long enough to grow strong enough that feeding time stops being a destiny machine.
That is the part of the metaphor that matters most. Hierarchy is not inevitable in the puppy world if someone legitimate intervenes early enough, firmly enough, and consistently enough to prevent appetite from becoming permanent structure.
The question is who plays that role in this world.
Markets will not do it by themselves. Markets reward the dominant puppy for efficiency, speed, and scale. They do not naturally reward restraint. The company that can build the strongest model, secure the most compute, and capture the most demand has very little internal incentive to slow down just because its dominance is becoming self-reinforcing. Shareholders do not usually fund voluntary surrender.
National governments are powerful enough to intervene, but they are also tempted to become partisan owners of their own preferred puppy. A state that believes frontier AI is decisive for growth, military advantage, and geopolitical leverage will not approach the bowl as a neutral caretaker. It will approach it as a strategic patron. It may talk in the language of safety and public good while quietly making sure its own champion keeps eating first.
Global bodies sound like the obvious answer because the problem is global, but their weaknesses are familiar. They are slow where technology is fast. They are fragile where power is hard. They are often dependent on the very states and companies they are supposed to coordinate or constrain. A global body may be able to name the danger, publish norms, and create forums for negotiation. It is much less clear that it can force the strongest actors to accept near-term limits when those actors believe they are racing for civilizational advantage.
Religious traditions and moral authorities can contribute something no market and no regulator can fully manufacture. They can remind societies that power is not self-justifying, that dignity matters, that the weak are not expendable, and that a system built only around hunger and dominance is not wise simply because it is efficient. History suggests this is more than sentiment. Moral movements have repeatedly altered what powerful actors could afford to do. Abolition reshaped what economies were permitted to run on. The campaigns against apartheid moved capital and governments through sanction and divestment. The pressure of public conscience has forced firms to abandon practices that remained perfectly legal and perfectly profitable. Norms are a slow form of power, but they are power, and they work by changing the cost of legitimacy rather than the price of compute. What they cannot do is allocate compute, inspect model training runs, or redesign procurement regimes by themselves. They can raise the price of behaving badly in front of the bowl. They cannot become the owner.
Purely democratized decision-making also has appeal. If these systems are reshaping everyone’s future, then ordinary people should have a meaningful say in how power is distributed and constrained. But democracy at planetary scale is not a magic solvent. Publics are vulnerable to manipulation, fatigue, disinformation, and the seduction of short-term national advantage. Majorities can be unjust. They can also be technically outmatched by the systems they are trying to supervise.
None of these candidates is sufficient on its own.
That is frustrating, but it is also clarifying. The world is unlikely to produce one benevolent owner with a scoop and a fair heart. What it may be able to produce is a layered regime of partial restraint: states checked by other states, firms checked by regulation, regulation checked by transparency, transparency strengthened by independent researchers, and all of it pressured by a moral culture that refuses to call domination progress.
That is not a clean answer. It is probably the only honest one.
What responsible intervention looks like
If the metaphor is right, then the answer is not to pretend scarcity does not exist. It is to change the feeding regime before hierarchy hardens into permanent biology.
Responsible intervention would start with public and allied compute capacity. Not symbolic research grants, but durable access to the infrastructure that allows universities, midsized firms, and middle-power states to train, fine-tune, and evaluate serious systems without total dependency on a single provider.
It would continue with anti-concentration policy that understands AI as a stack. Competition policy cannot look only at the chatbot someone sees on a screen. It has to examine chips, cloud, model distribution, safety tooling, data partnerships, energy access, and communications infrastructure as one interlocking system.
It would also require model accountability rules that scale with power. The stronger the system, the stronger the duties around transparency, auditability, incident reporting, and controlled access. Not because powerful models are evil, but because concentration without explanation is politically corrosive.
And it would require a serious sovereignty agenda for countries outside the inner ring. That does not mean every country must build a frontier model from scratch. It means every country should have a strategy for what it must be able to understand, host, test, govern, and refuse. Sovereignty is not the fantasy of total independence. It is the refusal to be permanently infantilized.
There should also be areas where slowdown is not treated as cowardice but as stewardship. If a handful of actors are approaching systems that meaningfully accelerate their own improvement, then society has a legitimate interest in delaying deployment until governance catches up. The usual objection is that someone else will move first. That may be true. It is not an argument for blind acceleration. It is an argument for building political coalitions strong enough to make restraint real.
The highway to the worst outcome
The most dangerous feature of the current moment is not simply that powerful actors are behaving as powerful actors usually do. It is that the world still has time to see where this road leads and is nonetheless drifting toward it with remarkable passivity.
Everyone can describe pieces of the danger. Some talk about monopoly. Some talk about alignment. Some talk about sovereignty. Some talk about jobs, war, propaganda, or electricity. All of them are looking at the same highway from different exits.
At the end of that highway is a world in which a small number of model families, firms, and states control the most consequential layers of machine intelligence; in which most countries rent cognition they cannot truly govern; in which public institutions supervise systems they cannot independently reproduce; in which dependence is marketed as participation; and in which the weaker puppies are told that the bowl was open to everyone when what was really open was the struggle for scraps.
That is the worst outcome because it combines fragility with domination. The world becomes more dependent on a narrow technical core at the same time that the power to shape that core becomes harder to challenge. It is not only unfair. It is dangerous. A civilization that centralizes too much cognition in too few places creates failure modes that are as political as they are technical.
And yet this outcome is not inevitable.
A great equalizer can still become something closer to what its early believers hoped if enough people are willing to contest the feeding order before it hardens into common sense. That means confronting monopoly even when it is wrapped in brilliance. It means refusing to confuse access with sovereignty. It means building institutions that can slow, inspect, share, and sometimes say no. It means accepting that no single owner will save the litter, and that fairness in this domain will have to be built through awkward coalitions, imperfect rules, distributed oversight, and moral seriousness.
The point is not to make every puppy identical. It is to prevent a temporary advantage from becoming an irreversible caste system.
A responsible society would not stand beside the bowl, watch the same strongest body grow larger on every feeding, and call that natural. It would intervene before hunger hardened into destiny.
This is where the AI question has arrived.
Not at the moment of invention.
At the moment when the feeding order is beginning to decide who will remain strong enough to shape the future, and who will spend the rest of the century trying to recover from an early deprivation that nobody chose to stop.