Special Research No.02
2026.08.09
On the evening of June 12, 2026, the U.S. government placed export controls on Anthropic. The controls covered its latest models, Claude Fable 5 and Mythos 5, as well as foreign nationals. That included users outside the United States and foreign-national employees working at the company.1
The servers never stopped. The company did not fail. A paid service was closed by a single document, then reopened eighteen days later. The episode was brief, but it clarified one thing: neither the company selling the AI nor the customer buying it held the final key to access.
If Sovereign AI is read as total self-sufficiency, no country qualifies. No country can supply accelerators, data centers, power, software, open source and research talent entirely within its own borders. That is why the term is used to ask the four questions below.
The answers to these four questions do not sit in one place.
They are scattered across the layers.
Opening a model is not the same as building the model yourself. Open weights give you permission to inspect and modify a model built by someone else. An independently built model asks a different question: who created those weights in the first place? A fully open model can still be someone else's model, while a closed model built from scratch can still be your own. The distinction comes to a head on page 6.
| Layer | What control of this layer means | What happens without it |
|---|---|---|
| MODEL | Create the weights yourself · able to retrain | The improvement path can be cut off by changes in the original developer's policy or license |
| COMPUTE | Own a domestic path to secure compute | Supply schedules and pricing depend on someone else's roadmap |
| DATA | Both storage location and legal jurisdiction are domestic | Sensitive data remains within reach of another country's laws |
| INFRA | Operate data centers and power infrastructure domestically | Little room to respond to outages or sanctions |
| OPERATE | Domestic teams handle deployment, security and operations | Technology comes in, but operating know-how does not accumulate |
| EXIT | Able to migrate to other chips, models and clouds | No bargaining card when terms deteriorate |
If all six layers are bundled into one test, the only conclusion is that no country has sovereignty. In practice, each country chooses different layers to protect. Those choices form the map on the next page.
All six cases fall within the debate over AI sovereignty. What they seek to protect is different: language, industrial data, auditability, public safety, Arabic, jurisdiction. What each country grabs first also reveals what it fears most.
| Case | Layer held most tightly | Layer not directly controlled |
|---|---|---|
| India | Model · data | Accelerator ecosystem |
| Japan | Data · application domain | Top-tier general-purpose models |
| Switzerland | Model · auditability | Frontier performance race |
| Singapore | Data · operations | Foundation model · commercial cloud for less-sensitive workloads |
| UAE | Arabic model · infrastructure ownership | Frontier model |
| France · EU | Infrastructure · jurisdiction | Accelerator hardware |
The right-hand column is not always the result of deliberate strategy. Some layers were consciously outsourced; others reflect limits in technology or capital. Even so, none of the six cases leaves the right-hand column empty.
Sakana AI's name itself refers to fish. Instead of growing a single giant model, it chose to combine multiple existing open models using evolutionary algorithms. Its Series B announcement put it this way: rather than “reinventing the wheel by training yet another large foundation model from scratch,” it opened a path to combining existing open models through evolution.11
Instead of paying to grow one more model, it pays to combine models that are already open.
| Item | Details | Timing |
|---|---|---|
| Evolutionary model merging | 「Evolutionary optimization of model merging recipes」 · Nature Machine Intelligence | 2025.01 |
| Automation of AI research | “Towards end-to-end automation of AI research” · Nature, Vol. 651 | 2026.03.25 |
| Mitsubishi UFJ Lending workflow | About six months of validation · roughly 100 participants (30 core members) · moved into live-case use at selected branches | 2026.03.06 |
| Daiwa Securities | Joint development of a total-asset consulting platform · separate from the Mitsubishi UFJ project | 2025.10.03 |
The widely repeated claim that “an AI-written paper was published in Nature” is incorrect.
What appeared in Nature itself was a paper by eight human authors describing the system.
The case in which an AI-generated manuscript passed review involved a conference workshop, and the authors later withdrew it themselves.
The Japanese government is not backing only this route. GENIAC continues to support the development of domestic foundation models. But in May 2026, it separately selected projects on manufacturing data and robotics foundation models. Japan is not betting on a single axis.
Finance is the proving ground. Loan review is a workflow buried under regulation and historical documents. The contest is less about how large the model is than how well it handles that institution's own documents. That is the position Sakana chose.
Sakana disclosed a validation period of about six months, roughly 100 participants including about 30 core staff, and a scope ranging from initial analysis and information organization to financial simulation and drafting approval documents. It is now being used on live cases at selected branches and divisions, while rollout to all branches remains only a stated goal. Start and end dates for the validation and metrics such as time saved have not been disclosed. This report does not estimate undisclosed figures.
The government speaks of an “AI G3.” Some indicators offer a basis for that claim. But model counts, patents and performance do not measure the same thing.
Korea's 2025 Notable AI Models. This is a count of models selected for technical significance, not a performance ranking.12
Stanford AI Index 2026, based on the count as of April 22, 2026.
It is difficult to say Korea stands on the same line as the United States and China. But within the group behind them, Korea is near the front.
Two different sets of numbers from the same report are circulating. The first edition released in April 2026 showed U.S. 50 · China 30 · Korea 5. The figures currently posted by Stanford are U.S. 59 · China 35, because the database continues to update. Korea's figure of 5 was separately revised after the Ministry of Science and ICT flagged missing domestic models and requested a review, bringing it to 8. The U.S. and China figures changed for one reason; Korea's changed for another. Mixing the two vintages creates inconsistencies within the same document. Always cite the snapshot date alongside the figures.
Multiple teams have directly built HyperCLOVA X · EXAONE · A.X · Solar · Motif
A core producer in the HBM and DRAM supply chain—the memory on which AI compute depends
14.31 AI patents per 100,000 people, No. 1 globally. Luxembourg 12.25 · China 6.95 · U.S. 4.6812
Frontier model · Accelerator ecosystem · hyperscale cloud. All three remain external
Korea's answer does not have to be a “Korean ChatGPT.” Factories, shipyards and logistics networks where AI is embedded most deeply can represent sovereignty on the same scale.
NAVER Cloud was eliminated from the government's independent-AI review. Five months later, NAVER announced with NVIDIA that it would build ‘sovereign AI infrastructure’. At first glance the two events seem contradictory, but they refer to different layers of sovereignty.
Among five teams, NAVER Cloud and NC AI were eliminated. The government named both. NAVER Cloud failed the independence standard; NC AI fell short on its overall score. The government's minimum threshold was explicit: Even when using external open source, reset the weights and retrain independently.
Ministry of Science and ICT announcement, 2026.01.1524NVIDIA released a model with 550B total parameters and 55B active parameters under the Linux Foundation's OpenMDW 1.1 license. It opened not only the weights, but also the training data and training recipe. Commercial use is also permitted.
Hugging Face model card · NVIDIA Developer Blog15NAVER's GAK Sejong data center will host an NVIDIA DSX-based AI factory, starting at 55MW and scaling toward gigawatt capacity. The next-generation HyperCLOVA X will be built by fine-tuning Nemotron 3 Ultra on NAVER's own data.
NVIDIA Newsroom, 2026.06.0713Brookfield joined the project, setting a 200MW target for 2028 and a long-term ambition of 1GW. NVIDIA announced a planned $1bn strategic investment, while Brookfield announced a non-binding term sheet for up to $9bn.
NVIDIA Newsroom, 2026.07.2414Four teams remain: LG AI Research, SK Telecom, Upstage, and Motif Technologies, which joined through an additional February call. From Aug. 8 to 11, 200 people randomly selected with gender and age balance in mind will use the systems and score them directly. The result will feed into the second-stage evaluation.
Ministry of Science and ICT · NIPA application notice24The government's dividing line was whether a team simply continued from weights created by someone else. Five months later, the model NAVER chose was one that disclosed the weights, training data and recipe together.
Page 2 made the distinction that matters here. Under the government's standard, using open source itself is not the problem. But a model that simply inherits someone else's weights and fine-tunes them is not counted as an independently built model. To qualify, the weights must be reset and the model retrained from the beginning. That line remains even if the chosen model is fully open. By contrast, the “sovereign infrastructure” described by NAVER and NVIDIA does not target the model layer in the first place. It means keeping the data center, data and operations inside Korea. The two sides use the same word to point to different layers.
| Party | Amount | Original wording in the press release |
|---|---|---|
| NVIDIA → NAVER | $1bn | plans to invest $1 billion into NAVER Corp. · subject to customary closing conditions and NAVER securing at least $9bn of separate financing as a condition |
| Brookfield → project | Up to $9bn | nonbinding term sheet · legally non-binding agreement |
| NAVER | Remaining amount | NAVER will fund the remaining amounts · amount undisclosed |
None of the three lines represents money already fully deployed. For 200MW, the releases use proposed · planned; for 1GW, they use intends to. Until the transaction structure is finalized, this report goes no further than “planned strategic investment.” Neither press release uses the term joint venture.
Calling it dependence alone misses what NAVER actually controls. Calling it full sovereignty misses how deeply the stack embeds new dependencies. Both sides need to be held in view.
No need to redesign from chips through operating tools · adopt a validated DSX design as a package
55MW · 200MW · Secure a single expansion path that can extend toward 1GW over the long term
Nemotron 3 Ultra · weights· training data · recipe also disclosed · relatively low licensing-restriction risk15
Different from relying on an external API · operate the data center domestically · adapt the model with proprietary data · deploy it directly
Not limited to accelerators · system design · operating software · agent tools · world models all sit inside one ecosystem
The deeper the full-stack integration, the higher the cost of moving to another chip or cloud · vendor lock-in
Even a domestic data center still uses U.S.-made accelerators and software stacks · it is not outside the reach of changes in U.S. export policy
The planned $1bn investment assumes NAVER separately secures at least $9bn in committed financing · if financing slips, the schedule can move
A full stack does not eliminate dependence. It changes the layer where dependence sits.
Stanford HAI makes a similar point about commercial sovereign-AI offerings: rather than eliminating supply-chain dependence, they reconfigure it and, in some cases, deepen it.3 The place to examine is what Page 2 sets out as the sixth layer: EXIT. Can the system move to another chip, model or cloud, and what would that move cost? Public materials do not disclose that value.
From here, these cases are farther from Sovereign AI in the strict sense. That is precisely why they are worth examining. A country's AI strategy is not determined by one question—whether it owns a domestic model.
Denmark built the national supercomputer Gefion with 1,528 NVIDIA H100 GPUs and has operated it since October 2024. It is used for drug discovery, weather forecasting and quantum-circuit simulation. The operating company is majority-owned by the Novo Nordisk Foundation, with the state export and investment fund holding 15%. The chips are American, but the country decides what they are used for and in what order.22
New Zealand stated in its July 2025 strategy that it would “put more weight on adoption and use than on developing foundation models.”, making the position explicit. The document also says New Zealand will not position itself to compete with Google or OpenAI. It addresses Māori cultural knowledge and data-sovereignty issues as well, but only in a single case-study box within a 19-page document. That is why it would be misleading to call the strategy “centered on Māori data sovereignty.”23
“Domestic or foreign” does not answer the question. Total self-sufficiency is expensive and slow; global collaboration is fast but creates new dependencies. Countries draw their line somewhere between the two. That is why the four questions from Page 2 remain unchanged at the end.
Across the fifteen countries and regions examined in this report, not one held all six layers entirely in its own hands.
So measuring sovereignty by whether something is “domestic” is not enough.
The questions worth counting are different:
Which layers were entrusted to others, what was received in return,
and when conditions deteriorate, is there still a door to leave through?
The door that stayed shut for eighteen days was an event that forced that calculation once.
Tap the underlined terms in the report to see their definitions. The two sections below provide the full definitions and supporting sources. Official announcements, press releases, papers and government documents were used as primary sources. Where a point could not be cross-checked against a primary source, the report states that limitation explicitly.