Ha Jung-woo: AI data centers should become 'token factories,' not just rental businesses

Ha Jung-woo, standing vice chair of the National AI Strategy Committee, attended a forum at the National Assembly on the 2nd hosted by Democratic Party lawmaker Kim Eui-kyeom titled "Forum for South Korea's Leap as a Global Robot Foundry and Realizing K-physical AI as a Leading Nation" and gave a presentation on "A national infrastructure strategy for an indispensable South Korea in the AI era." At the event, he called for fostering AI data centers.

He said AI data centers should not be seen as mere computing infrastructure, but as a national strategic industry for producing high-value-added "intelligence tokens." He also outlined an industrial structure linking AI semiconductors, data centers, foundation models, and physical AI, and said power, water, and cloud operations should be designed in an integrated way.

He also stressed an approach that goes beyond simply attracting data centers. He said this should be aimed at expanding domestic industrial AX, demonstrating local AI semiconductors, and fostering regional industries, with the benefits of investment spreading more broadly.

Ha diagnosed generative AI as rapidly evolving beyond question-and-answer interactions into agentic AI that can plan, use tools, and carry out tasks. He explained that when AI is given a problem, it can plan, execute according to the plan, and then reassess the results before reporting back.

He said real industrial productivity innovation is already taking place through agentic AI. In manufacturing, finance, bio, and other industrial settings, productivity gains are occurring through the use of AI agents, he added.

He said the importance of computing infrastructure is rising in step with these changes. He identified memory, data, and power as the key factors in AI competitiveness. He also said securing high-performance memory has emerged as a critical supply chain that determines the AI competitiveness of countries and companies. He added that the amount of high-performance memory secured will determine which companies and countries are strong in the AI era, stressing that memory is an important part of the core supply chain in the AI era.

He said there are capabilities that matter beyond open data. He stressed the importance of converting decades of accumulated experience and know-how on factory floors into AI training data.

He likened AI data centers to "token factories" that produce intelligence by inputting power, water, GPU, NPU, and data. He explained that if you put GPU and data into an AI data center, AI generates pieces of knowledge and information, emphasizing that these are production facilities that create intelligence by putting in resources.

He said the collection of generated pieces of knowledge and information can be embodied as reports, images, and complex program code. He also said the scope of data center output should include not only text, images, and code, but also action intelligence for manufacturing equipment and robot operations, and expressed the view that such outputs should be regarded as high-value-added assets.

He also pointed out that domestic users' use of ChatGPT and Gemini is a consumption structure for tokens produced by overseas AI data centers. He said domestic users are currently limited to consuming tokens produced by overseas AI data centers.

He said the direction needed in the future is to produce intelligence in domestic data centers for manufacturing, shipbuilding, automotive, and robotics. He also mentioned domestic industrial applications and overseas supply as uses for the intelligence produced.

He said paying subscription fees for ChatGPT and Gemini means using products from token factories based in the United States. He added that what is produced matters more than the scale of the data center being built.

He explained that it is necessary to link domestic manufacturing and AI semiconductor competitiveness with token production. He also said that when data centers produce intelligence that can be used directly on industrial sites, investment in data centers can translate into industrial competitiveness.

He also said the approach of only providing infrastructure and leaving utilization to others should be avoided. He said AI and token factories should be operated in a way that incorporates our technological assets to create high-value-added tokens.

He said sovereign AI cannot be completed with only domestic foundation models. To that end, he said a certain level of self-reliant capability is needed across the entire spectrum of power, semiconductors, computing infrastructure, data centers, cloud, AI models, and utilization.

He said sovereign AI cannot be achieved simply by training a lot of Korean-language data to develop a model with high Korean-language understanding. He added that areas lacking sufficient capability can be handled in cooperation with global companies, but stressed that even in cooperation, technological capabilities that can lead the way must be secured.

He then proposed easing regulations, shortening permit and approval times, expanding global investment, strengthening power infrastructure, and fostering neo-clouds as tasks for developing the AI data center industry. He also mentioned the need to shorten data center construction periods.

He cited site selection, construction, and power as regulations that need to be improved, and proposed power, water, and telecommunications as permits and approvals that should be processed quickly. He also named transmission and substation facilities, energy storage systems (ESS), uninterruptible power supplies (UPS), and renewable energy as equipment and resources that need to be secured.

He also proposed using AIops (AIOps) to optimize operations as a capability that should be fostered. This was presented as a task requiring not only broader investment and a stronger power base for the AI data center industry, but also operational competitiveness.

Concerns were raised about dependence on demand from global big tech data centers. Since local governments support infrastructure such as land, power, and water, there were also calls to create a structure in which the benefits of investment flow back to domestic companies and regional industries.

He said that even if local governments solve land, power, and water issues, there will be insufficient national spillover effects if the benefits are monopolized by global operators. He said overseas demand should be utilized in parallel, but that it is necessary to build a token factory ecosystem based on our own technology and capabilities.

As goals for the token factory ecosystem, he proposed the shared growth of startups, deep tech firms, and large corporations. He also said that in addition to large-scale training data centers, the importance of inference data centers will grow.

He explained that inference data centers handle AI services in locations close to industrial sites. He said inference data centers will grow as trained AI is used more broadly in services, factories, and other fields, and proposed designing mid-sized data centers inside industrial complexes as physical AI inference data centers specialized for those industries as an important task.

He proposed a K-XPU demonstration to verify domestic AI semiconductors in actual data centers. The target resources are heterogeneous computing resources such as domestic GPU, NPU, ASIC, and CPU, and the operating venue is an actual data center. The proposal aims to reduce dependence on overseas GPU, verify domestic AI semiconductors in data centers, create an initial market, and expand into industrial sites.

He said the view that semiconductors and data centers should be considered separately should be avoided. He explained that a full-stack perspective encompassing power, data centers, AI models, and applied physical AI is important.

He also said that data center locations need to be distributed. The standard is functional dispersion.

For the greater Seoul area, he proposed small and medium-sized inference facilities for real-time services and ultra-low-latency processing. He said data centers in the greater Seoul area can be small facilities for actual ultra-low-latency services, and mentioned the need to separate roles between the greater Seoul area and the regions.

For the regions, he proposed large-scale training and industry-specific data centers. He explained that the regions should handle large-scale training and operate industrially specialized data centers linked to industrial complexes. As a concept in the outline, he proposed a Saemangeum-Gunsan AX hub.

He proposed that support from the central government and local governments should not be simple provision, but should be designed to align with regional industries, with power, water, permits and approvals, regulation, and investment support based on regional industrial linkages. This also includes a flow in which data center operators participate in regional industrial AX, energy efficiency improvement, talent development, and domestic AI semiconductor demonstration.

As an example of a regional AX model, he cited Saemangeum and Gunsan, outlining a concept that connects hyperscale AI data centers based on renewable energy such as solar and offshore wind, the power grid, and ESS, as well as neo-cloud and GPU cloud links, AI computing optimization platforms, and robot foundries.

The AI data center computing resources secured in this way would be used for AX at manufacturing companies in nearby areas such as Gunsan and Wanju, while robots would be deployed in manufacturing, logistics, agriculture, and service sites. The expansion path would also connect robot design, prototyping, components, assembly, and mass production into a regional physical AI ecosystem.

The three major mega projects are semiconductors, AI data centers, and physical AI. The idea is to combine them into an industrial structure that creates demand for one another rather than推进ing each separately, and this concept aligns with that direction.

He stressed that the employment effect of data centers themselves is limited. He also said that because direct employment at data centers is small, linking them with regional industries is important.

He said priority should be given to contributing to the AI transition of industries near the region. He also said it is necessary to include providing cheaper computing power to local schools, startups, and research institutions.

The competitive phase for AI data centers is shifting from expanding facility scale to building an industrial ecosystem. The key variable for future competitiveness will be the degree of organic connection among domestic semiconductors, cloud, manufacturing, and physical AI.

Source: TECHWORLD · Kim Seung-gi
Original: https://www.epnc.co.kr/news/articleView.html?idxno=406445

References

• https://www.epnc.co.kr/news/articleView.html?idxno=406445