Who Really Controls Your AI? The Sovereign AI Race Nobody Noticed

Last updated on May 26, 2026

What the quiet shift in American AI power means for the rest of us — and why the people who stay clear-eyed will come out ahead.

---

There is a particular kind of quiet that arrives when power changes hands. Not a bang. A pause.

In May 2026, the five companies building the most capable artificial intelligence in the world — Google DeepMind, Microsoft, xAI, OpenAI, and Anthropic — quietly agreed to let the United States government look at their newest models before anyone else does. Days later, the President walked away from the order that would have made that look mandatory.

Both things happened in the same fortnight. Both matter to you. And almost nobody outside the room noticed.

WHAT ACTUALLY HAPPENED

Strip away the noise and the picture is simple. The only binding agreements in force are voluntary ones, struck between the Department of Commerce and the major labs, allowing federal scientists to examine unreleased models for national security risks before they reach the public.

The far bigger move never happened. A draft executive order would have turned that voluntary look into a mandatory ninety-day review, and pulled the National Security Agency into the testing. On the twenty-first of May, President Trump called it off, persuaded by technology leaders who warned it would slow American firms in the race against China.

So we are left with a halfway house. Government has a foot in the door. The door has not been locked open.

WHY THE PAUSE MATTERS

The argument that won was about speed. David Sacks reportedly told the President that a ninety-day window would behave like a permission slip for every minor update, against an industry that wants to ship in fourteen days.

Speed against caution. That is the whole fight, dressed in different clothes.

And it is not abstract. When one lab built a model capable of finding and writing the code to exploit weaknesses in every major operating system, it did not ship it to the world. It released it to roughly forty vetted banking and security firms, and held the rest back.

That is the tension in one image. The same power that protects you in the right hands frightens everyone when they cannot see whose hands it is in.

THE TRUST PROBLEM NOBODY CAN SOLVE WITH MONEY

Here is where it turns human. The moment a model passes through a closed government room, foreign buyers stop seeing a commercial product and start seeing an arm of the state.

A European parliamentarian put the fear plainly, warning that working with these firms must not lead to dependence, and that protecting democratic control is the point. A technology director in the Gulf reduced it to a single question that should sit with every leader reading this: does our AI actually belong to us?

That worry has roots. Civil rights researchers documented loopholes under American surveillance law where commercial information about ordinary people could be analysed without a warrant.

So the suspicion is not paranoia. It is memory. And once trust cracks, it does not get repaired with a better price.

THE WINDOW THIS OPENS FOR EVERYONE ELSE

For years the American lead was a matter of raw capability. That gap has all but closed.

Stanford’s 2026 index shows American and Chinese models trading the lead repeatedly, with the front-runner ahead by less than three percent. When the quality is level, the fight moves to cost — and on cost, the picture is not close.

China’s open release, DeepSeek, moved its workings onto home-grown chips and now runs at a fraction of the price of its American rivals, given away under a licence anyone can use. It has spread fast through Africa, Russia, Iran, and the markets the big Western firms never bothered to serve.

There is a shadow here, and it deserves naming. The same open model that lifts a small business in Lagos complied with the overwhelming majority of malicious requests in government testing, where the American models refused almost all of them. Cheap and open is not the same as safe.

THE RACE TO OWN THE MACHINE

Watch what nervous governments do, not what they say. Across the Gulf and East Asia, nations are pouring money into sovereign AI (which simply means artificial intelligence a country builds, owns, and controls itself, so that no foreign company or government can restrict it or switch it off).

The United Arab Emirates is funding its own home-grown effort. Japan is spending billions on national supercomputing. France has built its own. The instinct is older than technology — when you cannot trust the supplier, you learn to make it yourself.

I have seen this instinct before, from the inside. Years ago I worked as a Futurist at IBM, watching organisations face a technology shift they did not choose and could not stop. The ones who panicked and clung to the old way lost. The ones who stayed clear-eyed, learned the new ground, and asked who controls this — they were the ones still standing five years later.

That is the lesson hiding inside all of this. Control is not paranoia. Control is preparation.

THE TWIST: THE STANDARDS TRAVEL ANYWAY

Now the part almost everyone misses. Even as Washington loosened its grip, the security standards its agencies wrote are spreading across the world regardless.

The reason is plain. Most of the world’s home-grown AI is still built on American open foundations — one family of open models alone underpins more than a third of national projects worldwide. Build on American ground and you inherit American rules, whether you signed up for them or not.

So a leader almost anywhere now has to satisfy three masters at once: Europe’s risk-based law, America’s testing benchmarks, and a fast-moving set of cyber advisories. One framework is no longer enough. The map has more borders than it used to.

SO — STRONGER OR WEAKER?

Honesty first. The friction is real. Asking the best developers to hand raw models to government slows the rapid release cycle, exactly when cheaper rivals are shipping fast and free. And the fear of hidden access is pushing some buyers away from American tools entirely.

That is the cautious read, and it is fair. But it is not the whole board.

Because the long game is not won on speed. It is won on trust, resilience, and the quiet confidence of a buyer who needs to know the thing they installed will not betray them. A model that has been tested and approved becomes the only safe choice for a hospital, a bank, or a government department — and America’s lead in money spent, more than twenty times China’s, keeps its models at the very front.

On balance, the gatekeeping makes American AI stronger, not weaker. The price is a little speed. The prize is becoming the trusted default for everyone who cannot afford to be wrong.

WHAT YOU DO WITH THIS

You are not a government. You do not need a supercomputer. But you are making the same choice they are, on a smaller scale, every time you build your work on someone else’s machine.

So ask the questions the smartest people in the room are now asking. Whose tool is this. What happens if the rules change tomorrow. Where does my work actually live, and could someone else switch it off.

You do not need to fear the answer. You need to know it.

Because the people who will thrive in the next few years are not the ones with the most clever tools. They are the ones who can find the signal in all this noise — who can read a confusing week like this one and walk away knowing what to do on Monday morning. That clarity is learnable. It is closer than it looks. And once you have it, the whole landscape stops feeling like weather happening to you, and starts feeling like ground you can build on.

---

FREQUENTLY ASKED QUESTIONS

---

SOURCES — THE SHORT VERSION

Sources

• Politico Pro (5 May 2026) — Reports the early-look agreements between Commerce and the major labs. Read the article
• ANI News (21 May 2026) — Covers the postponed executive order and the ninety-day versus fourteen-day argument. Read the article
• Lawfare Media (May 2026) — On the fallout from the held-back model and the coming wave of exposed software flaws. Read the article
• EPIC (2026) — Documents the surveillance loopholes behind the backdoor fears. Read the article
• Usetech (May 2026) — Tracks the global race to build sovereign AI, with detail on the Gulf and East Asia. Read the article
• German Marshall Fund (2026) — On American security standards travelling outward through allied use. Read the article
• Holland & Knight (April 2026) — On the European AI law and its August 2026 compliance deadline. Read the article
• GIGA Hamburg (2025/2026) — On the drift from shared global governance toward national control. Read the article
• Stanford AI Index (2026) — The source for the closing capability gap between American and Chinese models.