K-water bets on AI Division and 2030 data center to fix water ops

Flood damage downstream of K-water's dams in 2020 started the utility's serious move into AI. Five years later, in late 2025, the Korean public enterprise formalized the effort by creating an AI Division. Its brief: merge digital twins, AI-run treatment plants and smart distribution management into one framework.

Korea Water Resources Corporation, known as K-water, operates as a national water wholesaler. It manages dams, treats water at purification plants and supplies municipalities. Historically, dam operations relied on accumulated climate patterns and operator experience. More frequent rainfall and drought extremes broke those forecasting methods, and management concluded AI was the replacement tool.

Ha-Dong Jung, VP of the AI Division, joined K-water in 1997 and became the unit's first head. He frames the division as a continuation, not a rupture, of the company's technology-driven history. That framing deserves scrutiny, but the groundwork claim is concrete: ERP work refined metadata and data standards before AI was on the agenda, letting K-water start AI projects earlier than peers.

What does the data actually need?

Collected data does not equal usable data. Jung is blunt on this point. "In the past, the purpose of our data work was to collect and store data and display it so people could understand it at a glance," he says. "Today, that accumulated data is just a collection of numbers. To use it for AI, we need to define what each piece of data means and how the different data points relate to one another."

To connect rain upstream of a dam to water quality, treatment output and final supply, K-water is investing in an ontology that defines data concepts and their relationships. It pairs that with retrieval-augmented generation (RAG) so systems can find and apply relevant information across facilities. No timeline or budget figure was disclosed for either effort.

Where did the infrastructure fail?

Execution exposed hard limits. Attempts to extract large data volumes for analysis sometimes stopped existing equipment outright. K-water responded by upgrading collection infrastructure and adding multiple data-acquisition routes.

"The existing equipment was designed with specifications suited to its original tasks," Jung says. "It wasn't built to accommodate the extraction of the large volumes of data that AI projects require. Sometimes we'd be told things were working, only to check and find that the process had stopped partway through."

The fix has a date and a place: a secured site in Sejong for a dedicated AI data center, targeted for completion by 2030.

How does a public enterprise handle workplace AI?

On the employee side, K-water first built an internal service called K-water GPT on Meta's Llama models, swapping in newer models over time to improve performance. In-house development could not keep pace with commercial rivals. Jung cites the benchmark employees already use.

"Employees already judge AI against global services such as ChatGPT, Gemini, and Claude," he says. "It's difficult to keep up with the pace of development in global AI models through in-house development alone, so we introduced external services that allow employees to choose among major AI models as needed in an environment with enhanced security."

Two further adjustments followed. Because the internal work network and internet network are separated, employees could not use internal documents with external AI; K-water built connections to make generative AI services reachable from the work network. And after initially allowing broad usage to drive adoption, it imposed individual usage limits to control costs.

What problem remains unsolved?

The structural gap Jung identifies is timing. A typical IT project takes six months to a year to prepare, about a year to implement and roughly six months to close out — while AI models keep changing underneath it. "Things have changed by the time you finish preparing, and they change again while you're working your way through implementation," he says. His proposed answer: build internal capability so employees can respond directly instead of relying entirely on outside contractors. He acknowledges that hiring staff and passing project and budget reviews takes considerable time.

External validation exists. OpenAI has highlighted K-water as a prime example of AI use in disaster response. Jung credits employee persistence, learned from earlier projects that initially fell short: "Earlier efforts build on one another."

His management method is similarly direct. "If it's a problem I can resolve at my level, I do everything I can to resolve it," he says. "If it isn't, I spend more time listening to what employees have to say." Jung calls current AI use early-stage. The 2030 data center deadline is the next measurable test of that ambition.