An image visualizing the concept of semiconductors having a certificate of origin while models don't A visual representation of the article’s core concept.

A warning that wasn’t about performance

Lim Jong-in, chair of Soongsil University’s AI committee, raised the possibility of copyright infringement around Korea’s national AI foundation model project, known as DAKMO. What stands out in this warning, reported by News1, is that it never touched on performance. How smart the model is, what its benchmark scores are, none of that was the point. The question was whether the model could prove what it was actually built from.

The issue at the center is distillation, the practice of using a foreign commercial model’s outputs as training data for another model. The terms of service for most commercial AI services explicitly prohibit using their outputs for this purpose. Break those terms and you run into a contract dispute before you even get to copyright. Last February, Anthropic alleged that a Chinese AI company used roughly 24,000 accounts to generate over 16 million conversations, which it then used without authorization to train its own model. OpenAI has raised similar concerns about comparable attempts. Detection information is now being shared through the Frontier Model Forum, which means enforcement has moved past the stage of mere declaration and into active practice.

DAKMO is a national project with a plan to invest roughly 530 billion won by 2027 to reach 95% or more of the performance of top-tier global models. Teams are eliminated in six-month interim evaluations, so the pressure to perform is significant. When a fast, cheap path to higher scores is right there, but the legality of that path is unclear, the completed model risks getting tripped up later during export or international partnerships. In the worst case, API access could be cut off, or the project could become the target of legal action. This is the context behind Lim’s proposal to use Samsung Electronics’ and SK hynix’s standing in semiconductors as leverage for preemptive negotiations with the US. If there’s no good way to prove innocence after the fact, the argument goes, secure the rights beforehand.

On the same day, semiconductors moved the opposite way

An interesting contrast showed up in the news on the very same day. SK hynix is holding a groundbreaking ceremony on the 27th for an advanced packaging fab in West Lafayette, Indiana. The investment is $3.87 billion, roughly 5.5 trillion won. Under the CHIPS Act, the US government is providing up to $458 million in direct subsidies and up to $500 million in loans, with operations slated to begin in the second half of 2028 and more than 1,000 direct jobs written into the plan. As the first HBM packaging line in the US, it’s also being read as a case that could shift the balance of power in back-end packaging.

What all these numbers have in common is that they’re written down. Where something is made is a condition of the subsidy. When it starts operating is a condition of the contract. How many people it employs is a condition of the permit. Which line an HBM chip came from and which cluster it ended up in can be traced through certificates of origin and supply chain documents. Even the US Commerce Secretary’s public request that Samsung Electronics and SK hynix expand their front-end fabs is possible only because what exists where is already documented on paper.

The same holds for Nvidia’s guarantee of up to $105 billion in rent, power costs, and residual value for an Ohio data center. OpenAI has exclusive rights to lease the initial 4.25GW facility for 20 years, and Nvidia is separately investing $1.5 billion in SB Energy, which is building the facility. This structure drew accusations of circular financing, where money earned from selling chips flows back into buying more chips, and the originally discussed $250 billion guarantee was scaled back to phase one only, under warnings from the bond market. CEO Jensen Huang pushed back, framing it as supply chain management, arguing that residual risk shrinks as OpenAI pays rent and capacity comes online.

What’s worth noting here isn’t whether the controversy is right or wrong. It’s that the controversy was even possible. Because who guaranteed how much to whom, and where that money flowed, is documented on paper, the market could question it, and the size of the guarantee actually got adjusted as a result. The same is true of the tally showing Samsung Electronics and SK hynix executed 45 trillion won in capital expenditure in just six months. Physical things leave a trail. And in industries where a trail is left, arguments get settled with data.

What happens on the side that leaves no trail

Model training doesn’t work this way. No matter how thoroughly you dissect a finished model, you can’t tell which data went in through which path or which part of the weights it seeped into. Unlike pouring concrete, training erases its own history as it proceeds. That’s why the distillation dispute isn’t a matter of technical verification but of circumstantial evidence and negotiation, which is exactly why it leads to a proposal to talk things over with the US in advance. Questions that a customs office and a contract would answer in the semiconductor world have to be answered by lawyers and diplomats in the world of models.

Worse, this problem can’t be undone once it happens. Even if contaminated data is discovered after the fact, there’s no way to surgically remove just that part from the model, which effectively leaves retraining as the only option. A defective wafer can simply be scrapped, but contaminated training shakes the entire finished product. This asymmetry is likely to widen going forward. In DAKMO’s second-round evaluation, Artificial Analysis scores came in at 47 for Motif Technologies, 37 for Upstage’s Solar Open2, 35 for SK Telecom’s A.X K2, and 31 for LG AI Research’s K-EXAONE 2.0. The model with the most parameters, at 750B, scored the lowest, while a 314B architecture that activates only 8 of 384 experts per token took first place. As it becomes clear that scale alone doesn’t determine performance, the center of gravity in this competition is shifting toward architecture design and training recipes. The more sophisticated the design, the more stages the pipeline splits into, and the murkier it becomes which stage something entered at. The better a model is built, the harder it gets to prove.

The real bottleneck isn’t training. It’s execution.

But look closer at DAKMO’s evaluation criteria and a different story emerges. In this competition, which narrows four teams to three and then picks a final two in December, the core evaluation axes are agentic AI capability and operational efficiency in real enterprise and public-sector settings. Scoring is split into benchmarks (40 points), expert evaluation (35 points), and user evaluation (25 points), with a 200-person citizen evaluation panel actually using the models hands-on and scoring them. The bar for a national model has already shifted from conversational quality to the ability to actually do work.

The problem is the state of the place where that work is supposed to happen. Hyundai Motor has deployed its own H Chat Pro internally, and Samsung Electronics has formally adopted three external LLMs, but in some organizations usage is actually slowing down because of concerns about confidential information leaking. This isn’t just a domestic concern either. In a Deloitte survey, 86% of companies in the Asia-Pacific region cited security vulnerabilities as their top worry about AI usage. The 2023 incident where internal source code and meeting notes were sent to an external server is still cited today for one reason: reconstructing what actually happened afterward proved difficult. What lingered longer than the leak itself was the inability to pin down its scope.

It’s also worth noting that hesitation isn’t due to lack of demand. Korea’s three major cloud service providers all reported double-digit growth on the same day. KT grew 20%, Naver 21%, and NHN 53%. This means AI workloads are already converting into revenue for domestic providers. To the question the circular financing controversy raises, whether the demand is real, the domestic figures answer fairly plainly. So the current hesitation isn’t because organizations don’t understand the value, it’s because there’s no way to confirm what leaks out in the process of capturing that value. What’s holding back adoption isn’t skepticism, it’s opacity.

To sum up: organizations are trying to run a model whose provenance is hard to prove, in an environment that leaves no trace of what it executes. Two layers of unprovability, stacked on top of each other. This doesn’t show up clearly when you’re just getting a summary from a chat window, but the moment an agent starts accessing internal systems, editing files, and moving to the next step without approval, it becomes a real problem right away. And one of those two layers can be fixed today.

Execution can be proven by design

The provenance of training data is extremely hard to trace after the fact, because by the time you’d want to check it, it’s already dissolved into the weights. Execution is different. Which tools an agent called, what data it read, whose approval it went through, and what it changed can all be documented from the start, as long as the system is designed to do so. Semiconductors didn’t get a certificate of origin because a trail happened to be left behind on its own. They got one because the system was built to leave one.

This is exactly why ThakiCloud’s Paxis treats skills, tools, policies, and audit logs as first-class resources. Policy determines what an agent can do; the audit log records what it actually did. The L0 through L3 autonomy tiers let you dial, on a per-task basis, how much requires human approval and how much can pass through automatically. Because execution happens inside an isolated sandbox, a question like what an agent handling internal documents sent, and to where, becomes something you can look up, not something you have to guess at. The same applies to external system connections. When the tools and skills attached through connectors are managed as resources subject to policy, you can increase both the pace of adoption and the scope of control at the same time.

The purpose of keeping an audit log isn’t limited to incident investigation. As a record accumulates of what passed and what was blocked, a security team can decide whether to raise or lower autonomy for the next task based on history rather than gut feeling. Contrary to the assumption that control slows down adoption, control backed by a record actually speeds up expansion.

Model selection sits on the same axis. Just as the 314B architecture outperforming the 750B one showed, the right model varies by task, and sensitive work and general work need to be processed in different places. This applies especially to organizations like Samsung Electronics, which has adopted multiple external models. Paxis is designed to let you choose a model per task, and workloads with sovereignty requirements can be kept inside a closed network on on-premises Kubernetes. That data never crossed a border is also something that needs to be answered with a record, not just a claim.

These requirements tend to show up first in procurement documents, especially in regulated industries. In finance, the public sector, and defense, access control and audit items often outrank performance items. If a national model becomes the standard for these sectors, the execution environment it runs on has to pass through the same documents.

The two threads in today’s news ultimately meet in a single sentence. We can already answer, on paper, where a chip came from. We still can’t answer where intelligence came from and what it did. The first half is something negotiation and institutions will need time to resolve. The second half depends on what we decide to use today. Now that agentic capability is part of the criteria for selecting a national model, where that model runs has to be decided with the same weight.

References

This article was compiled from the following news sources.

Tags: agent-native-cloud, ai-governance, audit-log, enterprise-ai, model-provenance, paxis, sovereign-ai

Categories:

Updated:

You may also enjoy