Over the last few years, we have talked endlessly about the risks of artificial intelligence.
We talk about jobs disappearing, errors nobody can explain, manipulation, and even systems that may one day become too intelligent for us to control.
I am not saying those conversations do not matter. But another, much less cinematic risk has been on my mind for a while. One that does not require waiting for superintelligence.
What if AI does not rebel? What if it works extraordinarily well, becomes indispensable, and ends up controlled by a very small group of companies and governments?
Perhaps the most urgent question is not when a machine will surpass us. Perhaps it is a much older one:
Who will hold the power?
Looking beneath the chatbot
From the surface, the AI market seems full of choices. We have ChatGPT, Claude, Gemini, Grok, open models, Chinese labs, and thousands of applications promising to do something different.
It is easy to look at that variety and assume power is equally distributed. It is not necessarily so.
Imagine a city with thousands of restaurants. You can choose Italian, Japanese, Mexican, or Uruguayan food. It looks like a diverse market. Now imagine they all buy ingredients from the same supplier, process payments through the same company, and rent their premises from the same owner.
The menus remain different. The point of control does not.
Something similar happens with AI. Beneath an application are models. Beneath the models are cloud platforms and data centers. Further down are energy, advanced chips, and a manufacturing chain dominated by very few companies.
We see the final layer because that is the one we use. Yet ten thousand products may depend on three or four decisions made far below. A price change, a new restriction, or a retired model can propagate throughout the ecosystem without most users knowing where it came from.
The World Bank describes a concentrated supply of compute and a highly concentrated AI chip market. It also reports that high-income countries held 77% of global colocation data-center capacity in June 2025. Talent is far more distributed than the infrastructure required to turn an idea into a frontier model.
That is where the problem begins.
The tool that quietly becomes a dependency
A company starts by using an API to summarize documents. Then it connects the model to customer service. Later it uses it to analyze internal information. When agents arrive, it begins delegating entire processes.
At first, it looks like one more tool. Until replacing it stops being simple.
The new model does not behave the same way. Its pricing is different. Its data rules change. Processes have been tuned around particular response times and behaviors. People have learned to work around its peculiarities.
Then we discover that replacing the provider was never just a one-line code change.
An IBM and Oxford Economics survey of one thousand executives found that 71% considered switching their main AI provider or model difficult. More strikingly, 91% acknowledged that they did not fully understand their organization’s dependencies across providers, models, and infrastructure.
Those figures deserve care. This was corporate research based on executive responses, not an independent measurement of every system. Even so, it describes something anyone who has integrated external services will recognize: adding a dependency is easy; recovering the freedom to choose is much harder.
This has happened with operating systems, clouds, social networks, and app stores. The difference is that AI can occupy a deeper position. It does not merely store information or distribute content. It starts writing software, serving people, recommending decisions, and acting on behalf of organizations.
The more capable the tool becomes, the more painful it may be to leave.
Power also lives in what we cannot see
When we discuss concentration, we usually name the labs building models. Much of the power, however, sits in less visible layers.
It lies with those who can buy chips, secure energy and land for data centers, and access enough capital to sustain months of training with no guarantee that an experiment will work.
It also lies in accumulated knowledge.
A paper can explain an architecture. It does not contain every failed attempt, the details of cleaning data at that scale, behaviors discovered in private evaluations, or the operational knowledge required to serve millions of people reliably.
That knowledge lives in teams. Every training run, incident, and user interaction produces experience that returns to the same organizations. More infrastructure enables better models; better models attract users; users produce operational information; that information improves the product and strengthens whoever was already ahead.
No conspiracy is required. The dynamic reinforces itself.
Something else worries me. More people now ask an assistant for a synthesis instead of visiting several sources and building their own answer. I do it too. It is convenient and often useful.
That convenience turns a small number of platforms into mediators of everyday knowledge. Somebody decides which sources are retrieved, which policies filter the answer, which nuances disappear, and what appears first.
Not because the model contains an organized truth. Precisely because it does not: it produces statistical answers, can be wrong, and reflects choices made during its construction. That is why it matters who trains it, with which data, and with what possibility of challenge.
The safety paradox
This is where the issue becomes genuinely uncomfortable.
We all want safer systems. We want evaluations, data protection, audits, and controls before dangerous capabilities are released. That is reasonable.
Doing all of it costs an enormous amount of money.
The higher the requirements, the fewer organizations can meet them. And the fewer organizations remain, the more power goes to those promising to keep the technology safe.
In trying to reduce one risk, we may increase another.
Making models completely open does not dissolve the tension either. Openness enables research, adaptation to other languages, and construction without permission from a central provider. It also makes some harmful uses much harder to control once the capability is available.
Closed models allow a company to enforce policies and respond to abuse, but leave those policies in private hands. The state can protect competition and rights, yet it can also regulate to surveil, favor allies, or close the market. And simply asking the market to compete is insufficient when entering the frontier costs billions.
I do not see a perfect answer. I am suspicious of anyone who claims to have found one.
Sovereignty does not mean building everything alone
This is why technological sovereignty is becoming a common phrase. The European Commission connects it with resilience, strategic autonomy, and reducing dependencies on key digital products, services, and infrastructure.
The word can sound isolationist, but it should not mean that. It would be absurd to expect every country to manufacture its own chips, build its own cloud, and train a frontier model from scratch.
To me, sovereignty looks more like preserving options.
Being able to switch providers without rebuilding the business. Running locally what should not leave the organization. Adapting a model to our language, laws, and culture. Keeping an essential service alive if a foreign company changes its terms or a geopolitical crisis cuts access.
It also means knowing where data travels, who can access it, and which jurisdiction governs it. A server may be physically located inside a country and still depend on software, keys, and staff controlled elsewhere. Location matters, but effective control matters more.
This feels especially relevant from Uruguay and Latin America. If we rely exclusively on systems built in other contexts, we do not import only technology. We import priorities, categories, and ways of interpreting the world that were rarely designed with us in mind first.
Cooperation is necessary. Complete dependence is something else.
What we can demand without fantasy
I do not believe we can eliminate all concentration. Some infrastructure requires enormous scale, and that scale can produce real innovation.
Nor do I think a company becomes guilty simply by growing or integrating several layers. The point is not to moralize every actor. It is to design a system that does not require powerful actors to remain benevolent forever.
We can demand portability and contracts that preserve an exit. We can design products capable of working with more than one model. We can use smaller, local systems when a task does not require a gigantic machine. We can support public infrastructure for universities and research, defend open standards, and prevent artificial barriers to competition.
And we can do something simpler: stop confusing convenience with control.
An easy-to-use platform is not necessarily ours. Many chatbots do not necessarily mean many centers of power. A company making good decisions today is not a guarantee that it always will.
Real diversity exists when no single decision can shut down, reprice, or redirect the entire system.
A conversation about power
For years, we imagined AI danger as a machine gaining too much power. Before that scenario arrives, we may have to face a far more familiar one: human beings using an extraordinary technology to concentrate economic, political, and cultural power.
AI does not need consciousness to change who is in charge. It needs no intentions and no rebellion. It only needs to become infrastructure on which everyone depends.
So, besides asking how intelligent it will become, we need to ask who can build it, who defines its limits, who can turn it off, and what chance the rest of us have to contest those decisions.
I do not have a closed answer. I worry about anyone being able to use very powerful systems to cause harm. I also worry about handing the authority to decide which uses are legitimate to too small a group.
That tension will not be resolved by a slogan. But it becomes visible when we stop looking only at the chatbot and follow power through the entire chain.
Artificial intelligence may become one of the most important tools we have created. Every time humanity has built a tool of that scale, the conversation has never remained purely technical.
It always becomes a conversation about who decides.
References
- IBM Institute for Business Value and Oxford Economics, The Calculus of AI Sovereignty, 2026.
- World Bank, Digital Progress and Trends Report 2025: Strengthening AI Foundations.
- European Commission, Strengthening Europe’s Tech Sovereignty.
This piece grew out of a script I prepared for El Rincón de Sarreche, where I talk about artificial intelligence, technology, and work without pretending there are easy answers. You can watch the original video on YouTube.