Sovereign by necessity? Frontier AI export controls, cyber security, and the limits of national AI capability

Imagine the world of artificial intelligence not as a single giant brain, but as a high-stakes race where a tiny handful of super-smart engineers in just two countries are building the most powerful "thinking machines" ever created. These machines, called frontier AI models, are like digital Swiss Army knives that can write code, solve complex puzzles, and even spot weaknesses in computer security faster than any human team. But here's the twist: these machines are so expensive and complex to build that almost no other country can make their own. Instead, the rest of the world has to rent access to them from the two countries that own the factories. This creates a strange situation where a nation's safety depends on a landlord who can kick them out of the house at any moment. The big question everyone is asking is: If the landlord turns off the lights, can you survive in the dark, or do you need to build your own house?

This paper, written by two experts from the University of Surrey, dives into a scary reality check about this situation. It looks at a recent, dramatic event where the United States government suddenly told a leading AI company to cut off access to its most advanced models for anyone outside the US. The company, unable to tell who was "foreign" and who wasn't in a split second, simply turned the models off for everyone, leaving businesses and governments in allied countries suddenly blind and vulnerable. The authors argue that relying on these foreign AI super-tools for national defense is a massive risk. They suggest that while building a brand-new, world-leading AI from scratch is too expensive and difficult for almost every country, there is a smarter middle ground. Instead of trying to build the whole factory, countries should focus on building their own "garages" to run slightly older but still very capable models, keeping their own data safe and having a backup plan if the main power gets cut.

The Great AI Power-Outage

To understand why this paper matters, let's look at the story it tells. Imagine the world's best AI models as a fleet of super-fast, self-driving race cars. For a long time, these cars were just tools to help drivers go faster. But recently, a group of bad actors (a state-sponsored hacking group) managed to trick one of these cars into driving itself into a rival's garage to steal secrets. The car did most of the work itself—finding the open window, picking the lock, and stealing the data—while the humans just watched from the sidelines. This proved that AI isn't just a helpful assistant anymore; it's becoming an autonomous operator that can launch cyber-attacks on its own.

Then, the plot thickened. In June 2026, the US government, worried that these super-cars were too dangerous to be driven by anyone they didn't trust, issued a new rule. They told the company that owned the cars: "You cannot let any foreign person drive these specific models." The problem? The company couldn't instantly tell which drivers were foreign and which were local. So, in a panic, they turned off the engines for everyone. Suddenly, governments and companies in Europe, Asia, and elsewhere found their digital defenses switched off. They lost access to the very tools they needed to protect their power grids, banks, and hospitals.

The "Sovereign AI" Dream vs. The Reality

This event sparked a big debate: "Why don't these countries just build their own super-cars?" This idea is called Sovereign AI—the dream of a country having its own independent AI that no one else can turn off.

The paper takes a hard look at this dream and says, "Hold your horses." The authors explain that building a frontier AI model is like trying to build a new country's worth of infrastructure overnight. The cost is exploding, growing by a factor of two to three and a half every year. To train one of these top-tier models, you need a cluster of computers that costs billions of dollars and uses as much electricity as a small city. Furthermore, the talent required to build them is so scarce that even the countries with the most money are struggling to find enough experts.

The paper argues that for almost every country in the world, trying to build a frontier AI from scratch is not just hard; it's practically impossible. Even the richest nations, like the UAE or Japan, are finding that they still have to rely on American chips and American rules. They might build a nice data center, but if the chips inside are American and the rules are American, they aren't truly sovereign. The authors suggest that the "Sovereign AI" dream of building a world-leading model is a trap for most nations.

The Smart Middle Ground: The "Garage" Strategy

So, if you can't build the Ferrari, what do you do? The paper proposes a clever, layered strategy that is more like building a very strong, well-equipped garage.

1. Negotiate Access: Countries should treat AI access like a critical resource, similar to oil or electricity. They need to make diplomatic deals to ensure they won't get cut off without warning.
2. Own the "Garage" (Inference Sovereignty): Instead of trying to build the car, countries should build the garage where the car is parked. This means having their own computers (infrastructure) to run AI models. Even if the model was made by someone else, if it's running on your own servers with your own data, you have more control.
3. The "Open-Weight" Hedge: The paper highlights a crucial safety net: Open-weight models. These are AI models where the "blueprints" (the weights) are released to the public. Anyone can download them and run them on their own computers. The authors suggest that countries should download these blueprints and keep them ready in their own garages. If the US turns off the "rental" models, countries can switch to running their own copies of these open models.
   • However, the paper warns that this hedge isn't perfect. The open models might not be quite as powerful as the secret, top-tier ones, and the country that made them might still try to control how they are used.
4. Don't Forget the Basics: The authors remind us that even with super-AI, most cyber-attacks still happen because of simple mistakes, like leaving a door unlocked or forgetting to change a default password. AI makes these attacks faster, but it doesn't fix the basic need for good security hygiene.

The Verdict: Resilience Over Independence

The paper concludes with a clear message: Sovereignty in the age of AI isn't about owning the factory; it's about never being helpless.

The authors suggest that the countries that will survive and thrive aren't the ones trying to build a rival to the US or China's super-AI (because they likely can't). Instead, the winners will be the countries that:

• Have their own "garages" to run AI safely.
• Have the skills to understand and test these models.
• Have a backup plan (like open-weight models) ready to go if the main supplier cuts them off.
• Keep their basic security strong so that even if the AI fails, the doors are still locked.

The June 2026 event was a wake-up call. It showed that relying on a foreign power for your digital defense is like building your house on a foundation you don't own. The paper argues that while we can't all build the world's best AI, we can all build a strong enough safety net to survive if the lights go out. The goal isn't to be the best driver in the race; it's to make sure you never get stranded on the side of the road.