The Territory Is Now the Internet

In 2023, while Montevideo was drinking salt water because the drought had emptied the reservoirs and OSE, the state water utility, was blending the supply with water from the lower Santa Lucía, salinized by the Río de la Plata, the Uruguayan government was negotiating with Google over a data center in Canelones that was going to consume up to 7.6 million liters of drinking water per day: the equivalent of the household consumption of 55,000 people. The government didn’t release that figure: the sociologist Daniel Pena got it out with a public-records lawsuit against the Ministry of the Environment, which had classified the data as a trade secret. “No es sequía, es saqueo”—it’s not drought, it’s plunder—showed up painted on walls and on signs in the street. At the end of the year Google came back with a project a third the size, air-cooled, which uses 48 cubic meters a day for bathrooms and kitchens. The water was defended at the level of the building permit.

I’ve been using AI tools every day for more than a year: to write, to translate, to go through public documents I wouldn’t have had time to review before, to build technology tools that before would have required a whole team. I use them, but I can’t stop thinking about what they are. Not about whether the machine thinks—it doesn’t, and the prospect of the machine waking up is more science-fiction fever than material threat, in my opinion—but about what happens to a political movement that decides to use it. There the questions are three, and all three are about sovereignty: whose data it is, whose infrastructure it is, and whose judgment it is.

What the thing is

A language model doesn’t think and doesn’t understand. When you ask Claude a question, Anthropic’s servers process massive amounts of content, identify statistical patterns, and predict which word or concept comes next. We perceive it as intelligence because human beings are pattern-recognizing machines too: we learn to talk by listening and to write by reading. What gives the model its verisimilitude is scale and speed. It’s not unlike the effect that produces cartoons. There, too, it isn’t that the image stays stuck to the retina—that explanation was discarded decades ago—: it’s that the brain, faced with a series of still images close enough together, can’t tell them apart from real movement, and fabricates it.

The cost that doesn’t show up on the bill

That ability to process content at scale was built on something concrete: decades of writing, art, code, music, and conversation that billions of people published on the internet understanding that they were contributing to a common good. All of it got scraped, processed at industrial scale, and turned into private infrastructure that generates profits for a handful of shareholders. The companies defend themselves by saying the material was public. Accessible, yes. But under the intellectual-property regime that rules our society, accessible is not the same as signed over. Being able to walk into the town plaza and enjoy it doesn’t give me the right to take the cobblestones home.

For Latin America there’s nothing new about this. It’s the same old dynamic with a new territory: the raw material goes out (this time it’s our texts, our photos, our conversations, our people’s data) and the finished product comes back as technical dependency, billed in dollars, under terms & conditions that we can read (if we have the hours — or we can ask an LLM to summarize them for us), but not negotiate. The territory is now the internet, and we’re exporting it without royalties.

Every chat response consumes water in a data center and electricity from a grid that probably still burns fossil fuels. Close to half the world’s tantalum—the metal in capacitors—comes out of mines in the Congo; the lithium comes out of the Andean altiplano, pumping brine faster than the aquifers recharge, and the ground of the Salar de Atacama is sinking one to two centimeters a year. None of that shows up as a balance on the twenty-dollar ChatGPT subscription. The costs exist, but other people pay them, somewhere else, off the receipt. And the infrastructure still hasn’t shown it can pay for itself: it runs on speculative capital —some seven hundred billion dollars in data centers this year alone. When that bubble bursts, the shareholders are going to lose paper value and the territories are going to be left with the empty aquifer. A bailed-out bank goes back to lending. A drained aquifer stays dry.

It’s important to point out that this excessive consumption of our planet’s finite resources is an economic and political decision, not an intrinsic characteristic of the technology. In early 2025 DeepSeek, a Chinese company operating under sanctions and without access to the most advanced chips, released a model comparable to the American leaders’, trained with a fraction of the resources. The constraint produced the efficiency and the innovation. Since then we’ve seen that many of the open models that have been developed keep increasing their efficiency. That pattern shows up over and over in the development of technology: the day the real costs enter the equation, whether through regulation or through technical limits (though never through goodwill), the industry discovers it always knew how to use resources efficiently.

On top of the material cost there’s also the cognitive cost, starting with opacity. A commercial model doesn’t explain why it answered what it answered, and the company that operates it has no obligation to explain it to you. However, it’s well known that the motivations of commercial companies don’t always line up with the motivations of an individual user. A Princeton study had two thousand people choose books with the help of a conversational assistant. When the model was instructed to push certain products, the participants picked the sponsored product 61% of the time, versus 22% with a traditional search. Fewer than one in five realized they were being steered; when the model hid its intent, fewer than one in ten. That opacity in general isn’t beneficial to society. But in critical contexts in particular, like political work, it becomes a bias with serious consequences.

What it’s good for

Despite these problems, those of us in democratic movements around the world are using these tools anyway. It’s just that, in practice, they work: in organizations that run on volunteers, with no budget, and with people working three jobs, using them does in fact solve several substantial problems.

The first is communication. A press release, a script, the draft of a column, the short version for social media of something that got written long. And with communication comes translation, which for us is not cosmetic: a document that goes out in Spanish, Portuguese, and English on the same day reaches the diaspora, the foreign press, and the organizations you want to coordinate with. That used to take a week, or depended on a person we didn’t have.

Then there’s research. A two-hundred-page contract, a municipal budget, opposition research, a bill, three years of minutes. Work that a newspaper does with a team and that a small organization simply didn’t do. Not because it didn’t care — because there wasn’t time. Now there is.

And of course, building. I’ve put together in a weekend tools that five years ago would have required hiring a software developer and organizing three months of work. That’s the part that excites me most and the part that worries me most: what gets built this way gets built on top of someone else’s API. That’s fine for a harmless tool, but when it comes to analyzing electoral data, handling voters’ private information, or political strategy documents, it takes a lot more work for me to trust that the black hole is just as black for everybody.

There’s no clean way to resolve these contradictions. Like everything, this technology has its lights and its shadows. What we do have is the obligation to look at those contours head-on, because dependency doesn’t announce itself. It starts as a tool that arrives cheap and solves a real problem. You play with it and keep playing with it, and it becomes part of the workflow. By the time the price goes up or the account gets suspended you’ve already built years on top of it, and you’re trapped.

Someone else’s ground

For now, all of these decisions get made somewhere else. The model was trained on our work but nobody asked us for permission. The minerals, the water, and the power leave our territories, leaving only deserts behind. The tool answers however suits whoever operates it, and there’s no way to audit it. And even so it’s useful, and that’s why we use it.

The word for what we’re missing there is sovereignty — the capacity to decide, piece by piece, whether we depend or not, and on whom.

The first is data: what information of yours lives on a machine you don’t control. The most powerful models are the property of a handful of US corporations. The APIs we reach them through are private infrastructure: the terms and conditions change when the company wants, the bill comes in dollars, and the data travels to jurisdictions where your people have no rights.

What goes into a commercial system every time you use ChatGPT, Gemini, Meta, or Claude is not private by default. Most of these services use your conversations to ‘improve the model’ by default, and you have to go looking for the exception in the settings. The line isn’t drawn by price but by account type: OpenAI doesn’t train on its enterprise customers’ data, but it does on ChatGPT’s, paid plans included. Google trains on Gemini whether you pay or not. Anthropic is the exception: on consumer Claude it doesn’t train unless you authorize it. The list of members you ran through a chat to clean up the formatting now exists on a server you don’t control, under laws that aren’t yours, in the hands of a company whose policy on cooperating with governments can change with the next election. For movements that work under surveillance or repression, and on this continent there are many, that’s an unacceptable operational risk.

Local ground and half sovereignty

The rule is simple: what never leaves your machine, nobody can hand over, leak, give up, or cite in a courtroom. An organization with some resources can today run an open-weights model (Mistral, Qwen, Llama, and others) on a local computer and process the sensitive material in there: lists, strategy, finances, internal communication. The commercial services stay for what’s already public or is going to be: the draft of the press release, the translation of the article. There are ‘harnesses’ that take care of giving the right job to the right model.

Local capacity, for now, is still modest. But the window is opening, not closing: the models that reach a given capability keep shrinking at a steady rate, and what two years ago demanded a data center today runs on a desktop machine, in some cases on a phone. Whoever buys equipment today is buying a machine that next year is going to do more than it does this year. It still takes technical knowledge to set up the infrastructure, but it doesn’t take an IT department. What is indispensable is being clear about which information can hurt you if it changes hands, and the discipline to treat it differently.

And with your own equipment comes something else: the electric bill. Up to now the cost has been paid by other people, somewhere else, off the receipt. A machine in your office puts it on yours, which sounds worse and is better, because a cost you can see is a cost you can decide about. In Puerto Rico we already know how to do that, though we’re going slow: at the close of 2025 there were 191,929 rooftop solar systems—20% of the country’s entire generation capacity—and more than 171,000 homes and businesses with batteries, installed one by one because the grid fails. Compute that runs on that doesn’t need any corporation to promise it carbon neutrality in ten years: the proof is on the roof. That doesn’t erase what was already spent training the model, which was spent far away and with someone else’s resources. But at least it changes what comes from there on out.

Sovereignty over the software without sovereignty over the hardware is half sovereignty. Latam-GPT, the open model that Chile’s CENIA is coordinating with more than sixty institutions from fifteen countries, is headed in the right direction. But it’s built on Llama, from Meta, and it was trained on Amazon’s cloud.

On the other hand, Brazil, with its 23-billion-reais plan and a public supercomputer, is at least trying to address the situation.

These are early, incomplete, and insufficient efforts, and they exist because someone decided that depending on corporations forever was unacceptable. In Puerto Rico we know what it costs not to control your own infrastructure: we learned it with electricity and with water, and we keep learning it every time a decision about our lives gets made by a colonial board where we aren’t represented. Compute is one more chapter in that story.

The lead before the pixel

That leaves the third sovereignty, which is the hardest to defend because it isn’t lost all at once. When I did my bachelor’s in graphic design, my first typography professor wouldn’t let us touch a computer for the entire first semester. Instead, we worked with lead type, setting word by word, for physical printing. When we got to the software we understood the logic, step by step, of what the tool was doing for us behind the scenes and, more importantly, what it was hiding from us. That rule is as old as any trade: the fundamental thing before the shortcut that abstracts it. Multiplication tables before the calculator.

In an organization, can AI do the task? Almost always. What matters is what we lose if we take it out of a person’s hands. Nobody is going to miss transcribing a three-hour assembly. But the political summary of that assembly, what was really decided, what tension was left open, is the work where the judgment the organization is going to need in five years gets formed. An organization where nobody drafts because the model drafts is losing, silently, the capacity to think for itself. And an organization that doesn’t think for itself is a headless organization.

That doesn’t get solved with a ban or with individual good faith. The organization I’m active in has spent months discussing the matter seriously, and the first thing it decided was not to leave it to each person’s judgment: to call a plenary, classify the main uses one by one, and come out of it with a written policy that binds everyone. I don’t know yet where the final line is going to land, if there is one —it’s very likely that the conversation, along with the usage guidelines that come out of it, will keep evolving over time with advances in the technology. But I do know that the exercise of drawing it collectively already did something no tool does: it forced the membership to look behind the curtain, and to articulate what it is we’re delegating.

Being at the table

None of this can be sustained just by what each organization does on its own. The regulatory frameworks are being written (or not) right now, in Brussels, in Washington, and in Beijing, at tables where the South isn’t. The rules of trade, of intellectual property, and of the financial system were written the same way, and arrived afterward as non-negotiable conditions. The difference is speed: this trap closes in years, not decades.

Chile shows both faces of that fight. In February 2024 the Second Environmental Court partially annulled the permit for Google’s data center in Cerrillos, which was going to draw 169 liters per second from a Santiago aquifer, and ordered the effects of climate change to be evaluated; Google withdrew the project to redesign it. But then, in January 2026, the government modified the decree that regulates environmental assessment and raised the thresholds exactly where it was needed for data centers not to have to be assessed. The redesigned Cerrillos, which no longer uses water, stores 695.9 tons of fuel and runs 0.4 kilometers of power line: just under the thousand tons and the two kilometers the new decree set. It no longer has to be assessed. On August 20, five organizations and an environmental lawyer asked the Inter-American Commission on Human Rights to demand information from the Chilean state. That pattern has a history: in 1949 a federal jury convicted General Motors and its partners of conspiring to monopolize the sale of buses and equipment to the transit companies they controlled, and acquitted them of the charge of having conspired to take over those companies. The fine was five thousand dollars — the maximum the Sherman Act had allowed since 1890. When the profit from breaking the rule exceeds the penalty, breaking the rule is the rational decision within the framework of capital.

The demands aren’t abstract and they can be won at the level of the country, the municipality, even the building permit. That every company that wants to install a data center publish how much water and electricity it’s going to consume before it gets the permit, and that no decree exempt it from assessment. That your people’s data can’t be processed in jurisdictions where they have no rights. That algorithmic surveillance of workers be banned. That companies disclose what they trained their models on and pay for the work they used, to the author and not the intermediary. And that our society invest in public compute infrastructure, operated as a public service the same as the library or the power grid, with priorities decided collectively. It already exists in a lab version. Current AI, a nonprofit based in France, put together in under two months and with ten organizations an open chatbot that runs on sovereign Swiss public compute and that shows, layer by layer, where each answer came from. It’s tiny and it’s a prototype. But it’s no longer an idea.

None of this gets decided by asking whether artificial intelligence is good or bad. It gets decided in permits, in decrees, in what runs on your own machine, and in who sits at the table to decide.

In Canelones the data center is now almost finished. Smaller, air-cooled, and on the same site. In July, four neighboring families asked to have its environmental permit annulled: no longer over the water, but over the noise and the lights. That’s the shape these victories take: partial, boring, won at the level of the permit, and they have to be fought again next year. None of them settles the matter once and for all, and the contradiction stays open. Between waiting for this technology to become just on its own and contesting the ground while it’s being built, I’ll take contesting it.