Washington’s Fight with Anthropic: What it Means for the Rest of Us

Colonialism had a clear method. Take the land. Extract the raw material. Send the value home. Govern just enough to keep that value-extraction pipeline open.

The data economy works the same way: Take something from people before they understand they are giving it away. Turn it into power. Sell it back to them as a “service.” The raw material used to be cotton or oil. Now it is attention, language, behaviour, and with AI, the ability to predict how people and governments will act. It moves through an “API” and almost nobody sees it happen.

What is new is who needs whom. Social media companies extracted attention from entire populations without asking permission from any government. They simply “built around” the state. Frontier AI companies are harder to build around. Anthropic and a handful of others have built something governments cannot easily replace: the ability to find security flaws no human team could find in time, and to defend critical systems before they are hit. A government cannot regulate this from a safe distance. It needs the company and it needs to control it.

This flips the old colonial question. It used to be: who controls the territory that has what we need? Now it is closer to:

Who controls the company that has built what we need, when we cannot build it ourselves?

Europe’s answer is what officials now call digital sovereignty. In June, the European Commission picked a European consortium, EUROPA, to build an open-source frontier AI model covering all 24 EU languages with a scale associated with the world’s most advanced AI systems. The goal is not just a product. It is independence. As the Commission’s own tech sovereignty chief put it, “Europe can lead in advanced AI on its own terms,” building a model that can match the best while staying true to its own values. The message is clear: do not depend on someone else’s model, build your own, on your own infrastructure, under your own rules.

But sovereignty like this needs money, computing power, and talent at a scale almost nobody else has. Right now only two places clearly do: the United States and China. Everyone else is choosing between three positions: buyer, regulator, or battleground. This is what the AI race actually means underneath the headlines. Not just who builds the smartest model, but who controls the infrastructure the rest of the world has to plug into.

That is the choice quietly forming for most of the world, including this region: align with the American AI stack, or the Chinese one. Europe seems to be the only actor seriously attempting a third path. Whether it can pull it off is still an open question.

In 2025, Anthropic became the first AI company allowed onto America’s classified military networks. It agreed to this on one condition: its AI could not be used for mass surveillance of Americans, and it could not be used to let weapons choose and kill human targets on their own. The Pentagon agreed.

By early 2026, the Pentagon wanted those limits gone. It asked Anthropic to allow its AI to be used for “any lawful use,” with no exceptions. Anthropic refused. Hours later, OpenAI signed its own deal to take Anthropic’s place;  a deal its own CEO later admitted looked rushed.

The White House response was harsh and personal. Officials called Anthropic “sanctimonious” and its CEO a man with a “God-complex.” Trump called the company “radical left” and “woke.” He then ordered every federal agency to stop using Anthropic’s technology. Days later, the Pentagon gave Anthropic a label that had never been used on an American company before: a national security supply chain risk;  the same label normally used for companies tied to China or Russia.

Anthropic said mass surveillance was simply incompatible with democratic values, and that today’s AI was not reliable enough to be trusted with life-and-death decisions on its own. Ordinary users responded by signing up for Claude in record numbers that same week, pushing it past ChatGPT and Gemini as the most downloaded AI app in many countries. The public, in effect, rewarded the company the state was punishing.

This is the backdrop that matters for what came three months later. When the US government ordered Anthropic to cut off its newest models over an alleged security flaw, it was already dealing with a company it had just tried to publicly shame and financially starve for refusing to hand over the one thing that made it valuable to begin with: real control over a powerful technology.

What changed was not the company, it was the politics

Once Anthropic complied quickly and visibly with the new restrictions, and once its CEO met Trump in person at the G7 global summit, the same president who had called the company “radical left” said he no longer saw it as a threat, and praised it for acting “responsibly.”

Nothing major about the technology had changed between the two moments- only how useful the company looked, and how willing it was to be “seen” obeying.

The US Administration did not take anything from Anthropic by “force” till now. It tried to take something subtler: control over how Anthropic’s technology could be used, and access to a capability the state itself could not build. When Anthropic refused, the state punished it the way it once punished foreign adversaries. When Anthropic complied, the state called it a partner again, in what is now a classic Trump approach.

This is what extraction looks like now. It does not necessarily need ships or soldiers. It needs dependency. The state does not own the resource. It needs the company that does, and it will reward or punish that company depending on how much access and control it gets in return.

Anthropic is a useful example because it is not extracting in the old, simple sense either. It says, often and publicly, that it wants to build AI safely, and that some uses of its own technology are lines it will not cross. That claim was tested twice in three months. This is the tension sitting underneath every major AI company today: the pressure to extract maximum value and growth, against a “stated commitment” to limit how that power gets used. Anthropic is not unique in claiming both. It is “unusual” in how publicly it has been forced to choose.

Out of this tension, a new kind of elite is forming. It is not built on land, factories, or oil, the way older elites were. It is built on being the only ones who can do something the state, the market, and the public all together depend on, and on deciding, even if partly for now, where that power’s limits sit. That is a “strange” kind of power. It is contested by governments and courted by them at the same time, and ultimately accountable mostly to itself. The people running these companies are not colonial governors per se. But they are starting to occupy a similar role: the ones who decide how a valuable, half-understood resource gets extracted, who gets to use it, and on what terms.

This is the part of the story that should worry more than just the United States. The dependency that played out between Washington and Anthropic this year is the same dependency every other state will eventually face, just from a weaker position. Washington, at least, had something to bargain with. The next question is what is left for a state that does not.

No country builds frontier AI alone anymore, not even most of the world’s richest ones. It takes computing power and energy on a scale only a few places have, enormous amounts of capital, and research talent. So the real question is not whether a country like  Jordan can build its own Anthropic. It cannot, not yet, maybe not ever. The real question is where it “plugs-in” to someone else’s.

This question has a national security dimension: The dependency at the heart of the Anthropic story; a government needing a capability it cannot build, and a company holding leverage because of it, is not unique to Washington. Every government that adopts foreign AI into its security, defense, or critical infrastructure inherits a version of the same problem, only from the weaker side of it. If a state cannot fully see inside the model it depends on, cannot verify what it actually does with sensitive data, and cannot replace it quickly if the relationship deteriorates, how much sovereignty does it really have over its own security decisions, and to what level will foreign interference and control trickle down?

There is also a domestic dimension, separate from who Jordan aligns with. The companies and governments setting the terms of this new economy are still working out, in public and often badly, what privacy, accountability, and oversight should mean once AI is involved. Jordan will have to answer the same questions, at some point, for its own citizens: who is accountable when an AI system makes or shapes a consequential decision, what happens to the data Jordanians generate every day, and who, if anyone, is watching closely enough to know.

None of this has obvious answers yet, for Jordan or for most of the region. But the choice underneath all of it is the same one the rest of the world is quietly being asked to make: align with one bloc’s infrastructure, the other’s, or try to carve out independence neither may be willing to grant.

Jordan does not have Europe’s leverage to attempt the third path alone.

The question, then, is who it builds that leverage with, and whether that choice gets made deliberately, or simply by default.