Imagine a login screen that checks your passport instead of your API key. The core of the Fable 5 block was nationality. Your payment works. You have a company email. You have an API key. But at the end, it asks: What country are you from?
Fable 5 was not just a new Claude. It evolved into a model that does not let go of a task, rather than simply one that answered better than its predecessors. It could follow a long codebase, hold onto complex tasks for a long time, and clearly reduce the parts where a person had to keep waking it back up. That model was blocked within three days. The target was not a specific country but “all foreign nationals.” This included foreigners in the United States and foreign-national employees inside Anthropic. A passport had been placed on top of the API key.
The surface-level reason was jailbreaks. But the larger backdrop was the distillation attacks by Chinese models. A model like Fable 5 is both a product and a teacher. Once it is open, people do not stop at simply using it. They query it at scale, collect its answers, imitate its reasoning style, and train smaller models. With Chinese open-weight models rapidly catching up, the United States likely felt that the moment it released its top model, it would become a competitor’s training data.
But this defense also reveals a weakness. The moat around the frontier model itself is thinner than we thought. If it were a truly durable moat, there would be no need to lock down access this tightly. A good model is observed the moment it is released. Once observed, it can be imitated. Even if it cannot be perfectly replicated, it can be followed closely enough. Today’s frontier becomes tomorrow’s distillation target.
Let’s look back at the history of software controls. The United States once placed powerful cryptographic technology under export controls, but cryptography ultimately spread through code and academic papers and became a default part of the internet. LLaMA was similar. After its weights were leaked, llama.cpp and the local inference ecosystem took hold, moving the center of gravity for running, compressing, and tuning models outside the company. Stable Diffusion showed the same direction. People built ecosystems on models they could hold, change, and build on, rather than on the most closed, highest-performing models.
Nationality is too crude a filter for blocking access to AI. People do not belong to just one country. A Korean national may work for a U.S. corporation, pay through a Singaporean corporation, live in Dubai, and handle European customer data. Even if a foreigner does not use the model directly, countless companies and people with access can create the output. They send a problem, not a prompt, and receive completed code, not an answer. So who used the model? The person who pressed the button, the person who posed the problem, the person who received the output, or the person who paid for it?
This is a strategic misstep by the United States. How long Fable 5 remains blocked is not very important. A powerful signal has already gone out. From the moment news of the block spread, every country and major company outside the United States would have begun emergency preparations to reduce their dependence on U.S. frontier labs. How much are our core operations tied to a single American model? If a stronger model is blocked tomorrow based on nationality, use case, or government guidance, what stops working? Do we have an alternative model? Do we have an open-weight fallback? Can we tune it ourselves? If these questions have come up even once in a boardroom, a national security ministry, or a CTO organization, then the United States has already lost future trust capital.
Whether Fable 5 opens again is no longer the point. Before this debate ends, another, stronger model will appear. There is no way to perfectly lock down every stronger model each time one emerges. In the meantime, models grow stronger, the cost of distillation falls, and open models evolve. The speed at which doors are locked cannot outpace the speed at which learning spreads.
In the past, nationality was something given to you. When you were born, you became a person of a particular country and entered that country’s laws, taxes, education system, and passport. But now people choose where to live, where to establish a corporation, where to pay taxes, which communities to join, which payment networks to use, and which networks to belong to. Nationality is moving from fate to membership.
This is also where Balaji’s network state thesis grows stronger. People increasingly feel their identity through the networks they belong to, the infrastructure they use, the values they share, and the opportunities they can access, rather than through physical borders. The more states use nationality to block access to AI, the more people will ask: Why should the one passport I was born with determine the intelligence I can use? Why are the networks I contributed to, the reputation I built, and the data I created less important?
Nationality is becoming less a source of loyalty and more a package of access rights. Which country provides better healthcare? Which country offers lower tax rates? Which country offers better access to finance? And now, which country offers better access to AI and a better AI ecosystem?
A country can buy a little time by keeping its models closed, but it will completely lose ecosystem leadership. What matters now is making people around the world want to build on that country’s infrastructure. Good open models, low inference costs, sufficient power and data centers, clear rules, and an environment where researchers and startups gather. A country should be a harbor for an ecosystem, not a vault for its models. Not a place that locks people out, but a place they keep coming back to.
Korea’s sovereign AI strategy should take a lesson from this. If it says, “We cannot use American models, so we will build our own to protect ourselves,” it will be too narrow. This moment is a rare opportunity for Korea to establish an open position in the AI ecosystem before the entire world. The moment the United States says, “Some people cannot use it,” Korea should say, “Anyone can build here.” We open our doors. We look at contribution, use, and trust—not nationality. Korea is a country with memory and chip technology, a manufacturing supply chain, ultrafast networks, real-world enterprise data from major companies, a stable democracy, and experience with global companies.
Open Korea’s homegrown models and tuning ecosystem to developers around the world. Become a hub for open inference infrastructure centered on memory and inference chips. Lead the routing, eval, and compliance layers that combine American models, Chinese open-weight models, and domestic models. What companies will truly need going forward is not a single model, but a structure in which work does not stop even when the model is swapped out. Korea does not need to become a small America of closed models. It is bigger to become a harbor for an open AI ecosystem.
The moats of the future will come not from closure but from accumulation. Domain data arriving every day. Interfaces embedded in customer workflows. Evals that can measure quality even when the model changes. Loops in which the traces of human corrections flow back into the system. Datasets that understand the regulations, language, and practices of a specific industry. Operational experience that tells you which model to use and when. These things cannot be replicated immediately by distilling another model.
The Fable 5 block brought the question of AI nationality to the surface. States want to place passports on top of APIs. But people are starting to see nationality more and more as a membership they can choose. Models can be locked down, but ecosystems can grow somewhere else. The more access is blocked, the larger the market for access rights becomes, and the more borders are erected, the stronger the argument for network states becomes.
Who decides the nationality of AI? The government, the company, the data center, or the user network?
And in the age of AI, what country am I from? The country where I was born, the country whose model allows me to use it, or some combination of the infrastructures I can access?
A state can ask who is using its model. But an ecosystem does not ask. In the end, won’t people flow toward the place that opens itself most widely to them?
