Why Countries Now Pay for Their Own Model, and Why It Is Hard

On 6 October 2026, South Korea's Ministry of Science and ICT announced a 4.7 trillion won programme, about $3.5 billion, to develop a frontier AI model starting in March 2027. The ministry will pick a lead developer through a competitive tender after parliament approves the 2027 budget in December, with a winner chosen as early as February.

$3.5 billion is a large number for one country outside the United States and China. The more interesting part is not the figure but the shape of the programme: the government combines state equity investment with private funding, and concentrates computing chips, data and talent on a single project.

What sovereign AI actually means

The phrase is used loosely, and the confusion is worth untangling. There are three levels that get mixed together.

The South Korean programme sits at the third level. The second, owning chips and data centres, is already held by most developed countries. The first, keeping data at home, applies almost everywhere because cloud computing forced the question by 2018.

What almost no country owns is a model of its own. And that is the expensive part.

Why programmes like this are appearing now

The reason is not technological optimism. There are three real triggers.

First, dependence on foreign suppliers. Most countries access frontier models through the API of three or four American companies. That dependence carries real risk: the price can change, the service can change, and the rules can change. The South Korean programme explicitly sets out to concentrate resources on one project, separate from its home grown foundation model initiative.

Second, language and data. Frontier models are trained on data dominated by a handful of languages. Countries with a specific language and specific industries often get lower quality than models trained on a far broader set.

Third, bargaining position. A country buying AI capacity from outside has little leverage. A country with its own model has an argument to negotiate with.

What Korea has and lacks

South Korea has three advantages not every country has. Samsung and SK Hynix together account for most of the world's memory production. Its domestic market includes high precision manufacturing that generates demand data. And its bureaucracy can move quickly.

The limitations are just as clear. A programme at the third level takes time, while a budget has a date. The tender chooses a developer in February 2027, and a frontier capable model takes longer than that.

A short history of national AI programmes

The first wave of these ran from about 2018 to 2020, and nearly all of them failed the same way. Countries announced programmes, then found that the models worth using were trained by private industry elsewhere. The lesson was consistent: a programme that stops at research loses relevance the moment frontier models are available commercially from the United States and China.

The programmes that survived were the ones that picked a specific use case early and let it grow.

South Korea is not repeating that pattern. Its frontier initiative is separated from its own domestic foundation model work, which is a sign the lesson was absorbed.

Numbers to read carefully

The 4.7 trillion won figure looks large, but the comparison matters. It is roughly 0.4 percent of South Korea's annual GDP. Set it against GPU spending at a single large company, which can reach a comparable amount within a year.

So the programme is large symbolically and relatively small next to the industry it is meant to matter in. It is big enough to be taken seriously and small enough that the money does not solve the hard part.

What is not yet clear is who gets picked in February 2027. The programme has no announced developer today, and every figure about the model it will produce is still a projection.

Four things that have to work

None of these are specific to South Korea, and any programme that succeeds has all four.

First, the chips have to be genuinely available. $3.5 billion disappears into GPU rentals if chip supply is not controlled. A country with its own fabrication holds a very different position from one that rents.

Second, the people have to exist. Training a frontier model needs several hundred people who can work on distributed systems. A country with an active AI research base has them. A country that has to import them will struggle, and recruitment is where it gets expensive.

Third, the data has to get used. A programme that builds a model but does not change how the country uses AI ends as a symbolic project.

Fourth, the results have to get used. A programme producing no real domestic benefit gets cut in the next budget, and the next budget is always smaller than the first.

What this means for readers in Indonesia

Indonesia sits in an interesting position in this conversation, and not quite the same one as South Korea. It has a large market, a language that frontier models largely do not serve well, and an economy large enough to have use cases of its own.

Three practical implications.

First, language is a translation layer. Frontier models are trained on data dominated by a few languages, and text outside those languages often performs worse on consistency, not just comprehension. For Indonesian language applications, a model trained with enough Indonesian data is cheaper to run over the long term, because it needs fewer corrections.

Second, CPUs and GPUs are not the same thing. Most office work in Indonesia is more text than heavy reasoning, and much of it runs on a small device. A bigger model is not automatically more useful.

Third, time to learn matters more than owning a model. A country with access to frontier models since 2024 has a capability it may not be using. What separates countries that benefit from ones that do not is usually not owning a model, but how fast organisations adopt it.

Four things can be done now, without waiting for any programme:

• Pick one clear use case. A programme with no specific use case becomes a symbol rather than infrastructure.
• Build your own evaluation. Vendor efficiency claims cannot be checked without internal measurement.
• Count token cost in rupiah, not dollars. The local exchange rate decides whether a frontier model makes economic sense.
• Train people first. Talent who can measure and evaluate is harder to find than GPUs.

There is a cost argument that gets missed. Running a frontier model costs money per
token, and that cost is usually quoted in dollars. At a local exchange rate that
moves, the same workload can double or halve in local terms within a year. Building
your own model does not fix that, because your own model also consumes tokens. What
changes the outcome is volume: the more internal usage exists, the more a locally
hosted model can spread its fixed cost over it.

That is the real reason a narrow use case matters. A programme with a small but
constant internal workload can justify a smaller model. A programme that has to
serve everything cannot.

Analysis: where this is most likely to fail

The riskiest line is the last one. The release cycle of frontier models is measured in months; a national budget cycle is measured in years. A programme that waits two budget cycles will always trail whatever already exists in the market by the time it finishes.

One thing reduces that risk, and it is not technical: design the programme to produce something useful in eighteen months rather than four. A smaller model that is already in use is worth more than a large one that is not finished.