Policy outcomes should be measured through faster AI deployment, productivity gains, cost reductions, private investment and exports—not data volume alone
In This Series
Part 1 — Data as National Infrastructure
Part 2 — How Data Becomes an Industrial Asset: Yang Jae-soo’s Framework for Verification
Part 3 — DataStreams Takes Trusted, AI-Ready Data Ecosystems Global
Part 4 — FLock.io CEO Jiahao Sun on Keeping Data Local While AI Learns
Part 5 — Hasibul Haque Links Operational Context to AI Accountability
Part 6 — Analysis: Korea’s Data Competitiveness Rests on Verifiable Industrial Results
The role of data in AI competition is changing. The industry is advancing from accumulating the largest possible datasets for model training to using data with the quality, domain expertise and operational context required for real-world training and evaluation. Across Korea IT Times’ Data as National Infrastructure series, Korean and international experts identified data quality, metadata, lineage, ontology, privacy protection, contribution measurement and operational context as essential components of an effective AI data infrastructure.
The decisive question is whether that infrastructure produces measurable economic value. Snorkel AI, a San Francisco-based AI data company founded at the Stanford AI Lab in 2019, provides one example. The company recently raised $350 million at a valuation of $3.5 billion. Reuters reported that its annualized revenue had risen from about $20 million a year earlier to more than $350 million, while Snorkel AI said on Sept. 22 that the figure had surpassed $375 million.
The source of that revenue is significant. Snorkel AI combines human expertise with AI automation to build training data and reinforcement-learning environments for specialist fields such as coding, law and healthcare. According to the company, its data-services business expanded rapidly within about a year of launch.
From usable data to operational context
Snorkel AI’s growth connects the issues raised throughout this series with a market willing to pay for specialist data and evaluation infrastructure.
Yang Jae-soo, president of the Korea Data Agency, identified quality-assured and trustworthy “usable data” as the starting point for AI competitiveness. He argued that quality certification, data valuation and secure distribution systems are necessary before companies and public institutions can put data to work in AI development.
Lee Young-sang, CEO of DataStreams, described data infrastructure as the foundation supporting AI performance and reliability. He emphasized the need to connect data fabric, metadata, quality, lineage and ontology so that companies and institutions can manage the meaning and relationships embedded in information dispersed across multiple systems.
Jiahao Sun, the London-based founder and CEO of FLock.io, outlined the conditions for using data across institutional boundaries. He said federated learning, which allows institutions to collaborate without moving raw data into a central repository, still requires privacy protection, secure aggregation and mechanisms to measure and reward each participant’s contribution.
Hasibul Haque, CEO of San Francisco-based Causal Dynamics Lab, argued that access to data alone does not enable AI agents to act safely inside real enterprise systems. Agents also need operational context connecting code, infrastructure, security policies, ownership and past incidents. Companies must also be able to trace the evidence and authority behind an agent’s actions, as well as the resulting operational outcomes.
Taken together, these arguments define the scope of modern data infrastructure. It begins with ensuring data quality, extends to structuring meaning and relationships, enables secure collaboration across institutions and ultimately supports AI systems that can make decisions and act within real operational environments.
Snorkel AI has commercialized one part of this infrastructure by supplying expert data and evaluation environments. Instead of providing data-labeling labor alone, it combines domain specialists with AI to produce data and reinforcement-learning environments for AI laboratories, hyperscalers, enterprises and U.S. government agencies.
Alex Ratner, Snorkel AI’s co-founder and chief executive, has said the company was founded on the premise that progress in AI would become increasingly data-centric. He argues that data development should be treated as a research and engineering challenge rather than a labor-intensive outsourcing function.
As models become more capable, their data requirements are also becoming more complex. A coding agent, for example, needs more than source code paired with correct answers. It requires an environment in which it can perform multi-step tasks, verify outcomes and demonstrate that it has not manipulated or bypassed the evaluation criteria.
The AI data market is consequently expanding from static datasets into training and evaluation environments. How an AI system’s behavior is tested and measured is becoming as important as the data on which it is trained.
Korea’s data policy needs a performance scorecard
Korea’s next challenge is to measure what its data investments produce in the real economy. Counting datasets, storage capacity and participating organizations is no longer sufficient. Policy assessment should examine how long a company takes to launch an AI service after gaining access to public data, how much data cleaning and integration reduce development costs, and how AI adoption affects productivity.
The scorecard should also track revenue generated by data-based services, follow-on private investment and export performance. These indicators would make it possible to test the connection between data policy and industrial competitiveness. If substantial public spending on data does not shorten AI development cycles or lower costs, the policy design should be reassessed.
The reverse is equally important. When an industry dataset accelerates product development and generates overseas revenue, that result should inform subsequent investment and policy decisions.
Korea has accumulated decades of data in manufacturing, semiconductors, healthcare, finance, construction and shipbuilding. The priority in these sectors is to structure the expert judgments and operational context embedded in production, diagnosis, design and quality-control processes so that AI systems can use them.
Data containing personal information or trade secrets must often remain within the institution that holds it. Korea therefore needs systems that allow organizations to retain raw data internally while using it for collaborative training and analysis. Federated-learning projects require privacy safeguards and clear methods for measuring each institution’s contribution.
When AI agents enter production systems, organizations must be able to trace the data they accessed, the permissions they received, the evidence behind their decisions and the results of their actions. Expanded data use must be designed together with governance, control and accountability.
Snorkel AI’s $3.5 billion valuation should not be treated as a target for Korea to replicate. Valuations reflect investors’ expectations of future growth, while annualized revenue represents the current pace of sales projected over a year and should not be confused with confirmed full-year revenue.
The more important signal is what companies and investors are willing to pay for. Businesses are investing in data containing specialist industry knowledge and in environments where AI can train, perform realistic tasks and have its capabilities rigorously evaluated. These data, training and evaluation services are developing into an industry in their own right.
Korea should apply the same test to its data policy. Its outcomes should be demonstrated through shorter AI development cycles, lower costs, higher productivity, new revenue and export growth.
Data infrastructure in the AI era extends beyond accumulation. It must produce data that AI can use and convert that capability into productivity, revenue and exports. Korea’s data competitiveness will ultimately be judged by those results.
