Korea’s challenge is to turn industrial experience into trusted data assets that work across institutions and borders
Launching the Series
Governments and companies worldwide are investing heavily in graphics processing units, data centers and large language models. Yet advanced computing infrastructure cannot deliver industrial results when corporate and public-sector data remains fragmented, its quality cannot be verified and usage rights are unclear. In this series, Korea IT Times examines what it takes to turn data into national and corporate AI capability. It connects policy perspectives from the Korea Data Agency (K-data) and enterprise experience from DataStreams with international expertise in data sovereignty, federated learning, operational context and provenance.
Series Overview
September 16 | Part 1 — Data as National Infrastructure
How shifts in the global data industry and Korea’s structural challenges are increasing demand for usable data.
September 17 | Part 2 — How Data Becomes an Industrial Asset: Yang Jae-soo’s Framework for Verification
The interview examines how quality certification, valuation, secure distribution and international cooperation can support national data infrastructure.
September 18 | Part 3 — DataStreams Takes Trusted, AI-Ready Data Ecosystems Global
How data fabric, metadata, quality, lineage and ontologies can prepare fragmented enterprise data for AI.
September 21 | FLock.io CEO Jiahao Sun on Keeping Data Local While AI Learns
How federated learning can support joint AI development while protecting sensitive information and data sovereignty.
September 22 | Part 5 — Hasibul Haque on Giving AI Agents Operational Context and Accountability
September 23 | Part 6 — KIT Analysis: Turning Korea’s Data Capabilities into Verifiable Evidence for AI
Industrial Data Becomes the New Front in the AI Power Race
The AI industry is moving from securing computing capacity and general-purpose models to connecting proprietary data with real operations. Models trained largely on public internet data cannot fully understand the causes of equipment failures, a patient’s treatment history, financial risks or accountability within an administrative process. Such judgments require internal data and knowledge of how it was generated.
AI competitiveness depends on whether data is ready for use. Organizations must manage accuracy, timeliness, provenance, usage rights, processing history and operational meaning together. Even data from the same industrial equipment may be unsuitable for training unless its sensor environment, error-handling procedures, maintenance history and relationship to productivity are understood.
“The key to developing the AI industry is data above all,” said Yang Jae-soo, President of K-data. “What matters most is not the absolute volume of data, but data that can actually be used.” Yang said AI infrastructure investment should include systems that prepare and govern data. These systems must manage quality, rights, metadata, update cycles and conditions of use.
Expanding data lakes and cloud storage does not resolve unclear meaning and accountability. If departments use different definitions for customers, products or equipment, or cannot trace how information was generated, they may be unable to verify AI output. Personal information, trade secrets and copyrighted material also require controls over users, purposes and permitted scope.
A data fabric connects the location, meaning, quality and usage conditions of information across multiple systems. Metadata explains what data represents, while lineage records how it was created and altered. Ontologies define relationships among customers, products, equipment, transactions and organizations in a form that AI can interpret.
“Ultimately, the performance of AI is determined by the reliability of the data it learns from and references,” said DataStreams CEO Lee Young-sang. “If an AI model is a building, data infrastructure is the ground and foundation supporting it.” Enterprise AI therefore requires an integrated framework covering data collection, standardization, quality testing, metadata, lineage and access control. It must also capture the reasons behind system changes, operational failures, employee decisions and customer outcomes.
Hasibul Haque, CEO of Causal Dynamics Lab, previously told Korea IT Times that today’s AI coding agents “still cannot see production reality.” An agent may analyze code and propose changes without understanding why the code was written or what happened when it ran in a live service. Agents connected to corporate databases and business applications therefore need both operational context and controlled access. Each agent requires a verifiable identity, limited permissions and records of the data it accessed, the actions it performed and the human approvals involved.
The European Union is establishing industrial data access rights, conditions of use, cloud switching and interoperability through the Data Act and common European data spaces. Its rules increasingly determine who may use data generated by connected products and industrial equipment and under what conditions it can be shared. The United States is developing AI risk-management and security frameworks alongside its semiconductor, cloud, model and software industries. As AI agents gain direct access to enterprise systems, identity, delegated authority, activity records and supply-chain security are becoming areas of standards competition.
Wanja Oberhof, CEO of Subsquid Labs, previously told Korea IT Times that AI would “run on trusted data.” Data volume alone cannot establish whether its source, accuracy and modification history are verifiable. Organizations must be able to trace where data came from, when it was collected, how it changed and who processed it. Provenance and lineage provide the basis for validating AI output and determining responsibility when failures occur.
Federated learning offers a way to combine data sovereignty with shared AI development. A model is sent to a hospital or corporate server and trained where the data resides, after which model updates are combined without transferring raw information. This approach has potential in healthcare, finance and public services.
“As compute costs stabilize, the true constraint in scaling AI has shifted to the scarcity, centralization and legal risks surrounding high-quality data,” said Jiahao Sun, founder and CEO of FLock.io. He added that federated learning enables collaborative training while keeping raw data within the organization. Keeping data local, however, does not eliminate security risks. Attackers may attempt to infer training information from model updates, requiring safeguards such as differential privacy, secure aggregation and homomorphic encryption.
Participants in federated learning also need rules for evaluating data contributions, assigning liability and distributing the benefits of jointly trained models. A technically decentralized system still requires governance over quality, participation and accountability. Without these safeguards, collaboration may preserve data location while leaving other forms of risk unresolved.
Turning Korea’s Industrial Experience into Verifiable Data Assets
Korea’s data industry has grown to more than KRW 33 trillion a year and employs about 240,000 people. Manufacturing, healthcare, finance, telecommunications and mobility have accumulated high-density operational data. Korea can also connect AI with established industries including semiconductors, automobiles, batteries, shipbuilding and healthcare.
Much of this data remains divided among institutions, companies and departments. The same information may be stored under different formats and definitions, while legacy systems often lack records explaining why data was created or how it changed. Privacy, trade-secret and intellectual-property concerns further limit sharing. Data providers also face uncertain returns and potentially significant legal liability.
Trusted environments and clearer contracts are needed to define permitted purposes, duration, reuse, rights to results, responsibility for incidents and revenue sharing. Public data programs also require stronger outcome measures. Participant numbers, data volumes and budget execution do not show whether data produced a service or commercial result. Assessments should track preparation time, AI service launches, productivity gains, cost savings, private investment and overseas contracts.
According to K-data, more than 196 data valuations had identified over KRW 155.5 billion in value by the end of 2025. Supported companies used the assessments to secure approximately KRW 19 billion in guarantees and investment. A follow-up assessment of companies supported by the 2025 Data Voucher Program recorded KRW 140.6 billion in sales contributions and KRW 73.7 billion in investment.
These figures indicate that data can function as an economic asset, but valuations require follow-up. Quality, usage rights, update status and contribution to earnings should be reviewed regularly. Such monitoring is needed to determine whether assessed value produces lasting revenue and competitive advantage.
Korea’s first task is to include data preparation in AI infrastructure investment. Common metadata, standardized business terminology, quality rules, lineage and rights information should be designed at the start of a project. Organizations should measure both the share of data immediately available for AI and the share converted into operational services.
The second task is to expand environments in which companies can collaborate while retaining control over their data. Secure data zones, federated learning and synthetic data can limit raw-data transfers. Standard contracts can define permitted use, accountability, rights to results and revenue sharing.
The third task is to establish access controls and audit systems for AI agents. Records of data access, actions, human approvals and permission changes are needed to trace incidents and assess automated decisions. These records should form part of the operating infrastructure rather than a compliance exercise conducted after deployment.
The fourth task is to convert domestic deployments into evidence that international customers can verify. Installing a system at a Korean public institution or conglomerate provides limited proof on its own. Vendors should document the scale of connected systems, reductions in errors and preparation time, security validation and measurable user outcomes.
International expansion also requires compliance with local privacy laws, data-storage rules, industry standards and procurement requirements. Certification, local trials, multilingual support, disaster recovery and long-term maintenance must form part of the offering. Government support should connect overseas testing, certification, legal review, procurement and sustained operations.
The value of Korea’s industrial data will be established through verifiable results. Korea must document operational meaning, usage rights, processing history and accountability under international standards. Converting experience from manufacturing, healthcare, finance and telecommunications into measurable evidence can strengthen Korea’s position in industrial AI and global data cooperation.
National AI infrastructure includes the systems that continuously prepare, govern and secure data. The ability to manage quality, provenance, rights and operational context determines whether AI creates trustworthy results and commercial value. Korea’s competitiveness will depend on how reliably it can turn accumulated information into data ready for action.
Next in the Series
Part 2, to be published on September 17, features an interview with Yang Jae-soo, President of K-data. Yang discusses quality certification, data valuation, secure data zones and the measurement of government support programs. He also addresses international cooperation and the conditions Korean data companies must meet to expand overseas.
