The One Who Holds the Valve Asks: From the US-Iran War to Fable 5

We've Seen This Movie Before

In March 2026, the Strait of Hormuz closed — the route a fifth of the world's oil once passed through. Crude topped $120 a barrel, and QatarEnergy declared force majeure on all exports. The International Energy Agency (IEA) called it "the largest supply disruption in the history of the global oil market." Governments reached for subsidies they had never used, even export controls, and held on until the war ended.

Something similar happened 53 years earlier, in 1973. Right after the Yom Kippur War, the Arab members of OPEC cut crude production and shut off oil to countries supporting Israel. Crude that had cost around $3 a barrel quadrupled to $12 within months. Korea, which doesn't produce a single drop of oil, took a direct hit. A growth rate that had boasted double digits collapsed to the 2% range the following year, and prices spiked over 40%. The same thing repeated during the second oil shock, triggered by the 1979 Iranian Revolution. Oil prices more than doubled again, and a growth rate that had exceeded 10% in the first half was nearly halved in the second.

Half a century apart, a crisis with the same structure has come again.

A crisis of the same structure across half a century

What the oil shocks taught wasn't a simple economics lesson. It was how power works, in the political-science sense. The moment the side that holds the resource shuts the valve, the side that doesn't can do nothing. Technology, diplomacy, market logic — all were powerless before it. Price wasn't set by the market; it was set by whoever could turn it off.

And now, the US-Iran war shows that lesson still holds. On news that the US and Iran had agreed to a memorandum of understanding (MOU) extending a 60-day ceasefire, oil fell 20% from its peak. But the path to a permanent peace remains opaque. There's no telling when or what Trump will say, and no telling what extreme choice Netanyahu might make if it keeps him in power. The daily lives of hundreds of millions hang on the decisions of a few.

We've already seen this scenario. The interesting part is that a crisis of exactly the same structure is starting again over an entirely different resource. This time it isn't oil.

The Valve Moved from the Oil Field to the Fab

What's being shut off this time isn't oil — it's compute.

In May 2019, the US placed Huawei on its export-control list (the Entity List). The world's largest telecom-equipment maker could no longer buy parts containing US technology. In December 2020, SMIC, China's largest foundry, was added to the same list. Then in August 2022, the Biden administration signed the CHIPS and Science Act, pouring $52.7 billion into its domestic semiconductor industry. That October, it blocked exports of advanced sub-14-nanometer chips and manufacturing equipment to China.

Up to here it looked like a solo play by a US worried about China's rise — the desperate flailing of a single hegemon trying to block a power transition tends to be fierce. But control soon spread to the alliance. In July 2023, Japan joined in controlling exports of high-performance chip equipment. The Netherlands' ASML fell in line too. As the encirclement tightened, China struck back, restricting exports of gallium and germanium — minerals critical to semiconductors and electronics. China produces 90% of the world's gallium and 60% of its germanium. Both sides weaponized their resources.

The edge of control kept turning toward AI. In October 2023 and December 2024, the US raised the level of export controls, sweeping in Nvidia's A100 and H100, and even the next-generation Blackwell GPUs — the chips used to train and run AI models. High-bandwidth memory (HBM), essential to AI, came under control too. In March 2025, the Trump administration added dozens more Chinese firms to the blacklist.

The descent of tech control — from chip to identity

Step back and the structure becomes clear. Control started with chips, extended to equipment, to memory, and to the AI model itself. If the oil shocks — or Iran's recent tactics — took the simple form of "shut off the crude," today's tech war tightens the valve at every layer of the stack at once: chip design, manufacturing equipment, memory, training chips, and even the right to access the model. Block any single point and everything stacked above it stops.

If the oil shock was a one-off blow 50 years ago, the tech-hegemony war is a permanent war, adding new rules every year. And the front line of that war is precisely AI.

The US government's block on Fable 5 shows what has changed on that front.

Fable 5 wasn't simply a smarter model. It could follow a long codebase to the end, hold onto complex tasks for a long stretch, and markedly reduce the segments where a human had to keep waking it back up. It was less a tool than a colleague you could hand work to. And that model was locked three days after release. The stated reasons were countering jailbreaks and blocking distillation by Chinese models — but the criterion for the block wasn't technology, it was nationality. The target wasn't a specific country either, but "all foreign nationals." You could have a payment method, a company email, an API key — and in the end they asked for your passport. Foreigners inside the US: blocked too. Even foreign-national employees inside Anthropic itself got caught.

Here you can see the nature of control evolving one more step. Oil didn't care who was buying. Pay, and it went on the ship. A chip is, after all, a thing, so control was about where it went. But access to an AI model began to ask who was using it. The unit of control came down from the nation to the product, and then to the individual. The moment a passport sat on top of an API key, access to intelligence became subordinate to that modern invention called nationality.

The power to lock now asks, "What country are you from?" The valve that began at the oil field, passing through the GPUs of the data center, has finally arrived at the individual's identity.

Alike — But Decisively Different

Up to here, AI looks like the new oil. There's the one who locks, the one who gets locked out, and between them a country like Korea is left agonizing. But there's a decisive difference between oil and AI.

Oil vs. AI — same structure, different variables

Oil had no substitute. When the Middle East shut the valve, all Korea could do was release its strategic reserves, cut consumption, and hold on. There was no way to suddenly conjure a resource that doesn't come out of the ground. Today the US and other producers are trying to spread the risk by raising output, but the set of those who hold the valve is still limited.

AI is different. Next to a blocked model sits an open one. Even if the US locks its top-tier model, an open-weight model can be downloaded and run on my own server. A model with published weights doesn't disappear even if the company goes under. Just as cryptographic technology broke through export controls, spread through code and papers, and became the default of the internet — just as the local-inference ecosystem exploded after LLaMA's weights leaked — a model, once released, can't be gathered back up. The speed of locking can't beat the speed of spreading. This looks like AI's own way out, one oil never had.

Many people conclude from this that the diffusion of technology is democratic. But there's a blind spot here too.

Even with an open model in hand, you need a GPU to run it. That GPU is designed by Nvidia, made by TSMC, and printed by ASML's equipment. To train and run a model you need enormous power and data centers. In other words, even if the door to the model layer is open, the door to the infrastructure layer beneath it is still monopolized by a few. This is exactly why export controls targeted chips, equipment, and memory. The US doesn't just lock the model — it locks the foundation that runs the model.

So open source is a way out and another dependence at the same time. The model is released for free, but the compute that makes the model useful is not. It's like the era of resource nationalism, when we said "we'll just build our own refinery," yet still had to import the crude itself.

The real lesson is here. Technology seems to flow without borders, but power always operates on the basis of borders. Most of the time, open source and global communities and distributed infrastructure make it feel as if borders have been erased. Then, at the decisive moment, when a particular government steps in, that illusion shatters in an instant. You have the chip but can't buy it; you have the model but can't run it; you have the key but get stopped at the passport. Twentieth-century sovereignty revives at the very heart of the technology that looks most borderless.

Sovereign AI.

The term has been used like a slogan of techno-nationalism: a declaration that "we too should have our own LLM, our own AI." But the story running from the oil shock to Fable 5 shows we should understand "sovereign AI" in a somewhat different context. The core isn't owning the model — it's not stopping when it's locked. Just as resource sovereignty wasn't about owning oil fields but about keeping the economy running even when supply is cut, AI sovereignty isn't about building the best model but about keeping core work and technological progress going even when one source shuts the valve.

The question we should be asking isn't "should we have our own model." It's: the moment the valve is shut, do we have another valve? Sovereign AI may be less a matter of self-sufficiency than of "the cunning hare's three burrows."

From Hydrogen Fluoride to GPUs

There's a reason this scenario doesn't feel unfamiliar. Korea has already experienced the valve being shut — twice.

In July 2019, Japan choked off exports to Korea of three materials critical to semiconductor and display processes — hydrogen fluoride, photoresist, and fluorinated polyimide. It was retaliation for a court ruling on forced-labor reparations. Politics had shut the valve of trade. At the time, Korea's dependence on Japan was 44% for hydrogen fluoride and a staggering 92% for photoresist. Memory-chip production itself could have stopped. Only then did Korea start running. It enacted a special law for materials, parts, and equipment, and poured a trillion won a year into localizing 100 key items. The results were decent. Imports of hydrogen fluoride from Japan dropped to a sixth, and some items were successfully localized.

It sounds like a success story, but not entirely. Once the export restrictions were lifted, even firms that had succeeded in localizing went back to the cheaper, more familiar Japanese product. We cry self-reliance only in a crisis, and the moment the valve reopens we go right back to dependence. (In a market economy, where trade is guaranteed, it's natural to gravitate toward the optimum.)

Another case. In 2017, in retaliation for the THAAD deployment, China issued an unofficial ban on Korean cultural content. The game industry took a direct hit. From 2014 to 2016, Korea had exported some 50 games to China, but after March 2017 not a single one got through. The door to the world's largest gaming market had closed. It took about three years and nine months for Korean games to be cleared for China again. In the meantime, Chinese games freely found success in the Korean market. The valve was locked in only one direction.

Japan shut the valve of materials, China the valve of the market. And both were political decisions, not technical or economic ones. Forced-labor rulings, the THAAD deployment — diplomatic variables Korea couldn't control decided the life and death of an industry. What the US and China are now doing over AI is the same kind of thing. The only difference is that the valve to be shut this time is neither materials nor a market, but AI intelligence itself.

There's a different story too.

In the early 2000s, when Korea tried to protect a homegrown search portal called Naver, many were puzzled. What's search sovereignty, they asked, in an age when Google was conquering the world? But as it turned out, Korea became one of the very few countries — apart from the US — where a domestic search engine holds the top market share. The Czech Republic, Russia, Japan, China — those are roughly the countries with their own search engine. That decision came back, twenty years later, as a crucial national asset.

While Naver held the top spot in search, a vast trove of Korean-language data piled up inside it — decades' worth of Korean corpus accumulated through news, blogs, cafés, Knowledge-iN, and more. And that accumulation became the training foundation for Korean-specialized large language models like HyperCLOVA X. Had Korea's digital life run entirely on top of Google, we wouldn't have had this data. Sovereignty was an asset for the future.

Materials, markets, data — whatever can be locked is, at some point, eventually locked. The side that didn't collapse was the one that had prepared another valve in advance. Korea mostly moved only after a crisis broke, but on search sovereignty it made a pretty good choice. So what about this time, with AI? Can we lay down another path before the lock comes?

The Country That Becomes a Hub

So what should we do? Build our own giant model? That just means building, at enormous cost, the single most expensive valve by ourselves. The real goal may not be one model, but a structure in which work doesn't stop no matter which valve is shut.

And this has to be prepared at three layers at once.

The three-layer structure of sovereign AI

First, the model layer. There's no need to stake everything on a single in-house model. We should be able to swap between a US model, a Chinese open-weight model, and a Korean model depending on the situation. What matters isn't owning a model but the capability to measure quality and swap models even as they change — things like evaluation systems (evals) and fine-tuning. When you can instantly switch to another model the moment one is blocked, you can build alternatives without being held hostage by the threat of a cutoff.

Second, the infrastructure layer. This is where Korea has the most to say. Running AI ultimately happens on top of memory and inference chips, and Korea sits at the center of the world memory market. HBM is already a valve Korea holds. Add domestic AI chips specialized for inference (like FuriosaAI or Rebellions), and Korea rises from "a country that borrows models" to "a country that owns the foundation for running them." Even if the intelligence was built by someone else, if it runs on my infrastructure, that's an entirely different bargaining chip.

Third, the operations and governance layer. Which model to use for which task, where sensitive data should run, how to reflect regulations and industry-specific practices — judgments like these. The accumulation of such judgment is the asset hardest to replicate. In particular, a field's own ontology — structuring the concepts, terms, regulations, and relationships of a specific industry — doesn't come along even if you copy a model by distillation. You can imitate a model, but the domain data piling up daily, the interfaces embedded in the work, the loop in which human corrections feed back into the system, and the knowledge system for understanding that industry — these can't be copied easily. (Satya Nadella's recent point about human capital vs. token capital seems to be in a similar vein.) The moat of the future comes not from what you lock, but from what you accumulate.

A country with these three layers grows stronger not by closing but by opening. The moment the US drew a line — "you're a foreigner, you can't use this," as it did with Fable 5 — is, paradoxically, an opportunity for Korea. That's when Korea can say, "here, anyone can build." Creating participation based not on nationality but on contribution and trust. There aren't many countries with memory and chips, fast networks, a stable democracy, and the data of real industries actually running.

In a network, power comes from the number of connections. A node connected to no one, however strong on its own, is just an isolated point; the hub through which the most connections pass makes its position irreplaceable. Control can turn Korea into an isolated node, or into a hub everyone passes through. The country that locks the door gains a brief edge but, as connections sever, gradually becomes an outlying node; the country that opens the door becomes, without owning anything, the hub where the most flow gathers. If developers and companies worldwide could build on Korea's infrastructure, and stack still more on top of that — because the more connections grow, the harder it is to leave, what's needed is not the power to lock but the power to draw people in. That, I suspect, is the path to becoming the center in the age of AI.

Of course, no one knows whether Korea can become one of the world's top three AI powers. But one thing is clear: declaring "we can't" without even trying is a foolish thing to do. If we narrow the front to where we're genuinely good — inference and memory — rather than a "K-AI" inflated by politicians' rhetoric, there's a real chance.

Fifty years ago we were a country without a single drop of oil. So we built refineries, brought in crude, and created a refining capability the world recognized. We had no resources, but we had the ability to handle them. AI is the same. The best model may not be ours. But the ability to run it, swap it, stack on it, and make it trustworthy can be ours.

The one who locks asks: what country are you from? The hub doesn't ask that. It asks what you came to build. Whether Korea becomes an outlying node or a hub everyone passes through — I'm so curious to find out.