On July 9, 2026, an international conference hosted by the University of International Relations(国际关系学院) was held in Beijing under the theme "Artificial Intelligence and International Security: Challenges and Governance." The author presented on "The Impact of AI on Northeast Asian Security." The main point was that Northeast Asia is already burdened with a dense cluster of traditional security flashpoints, including the front line of U.S.-China rivalry, Sino-Japanese tensions, the Taiwan Strait, and the North Korean nuclear issue. Layering AI on top of this stands to make the region only more uncertain and volatile.

The conference was organized around four main themes. First, how AI is reshaping the traditional security agenda. Second, the dimensions of great-power competition over AI supremacy. Third, how to respond to AI-driven risk as it spreads across domains such as energy, power grids, the cognitive space, and terrorism. Fourth, the need for coordinated global governance and how to achieve it. Participants came from the United States, Japan, Russia, India, Pakistan, South Korea, and elsewhere.

Some may still think of artificial intelligence as a matter for the future, but given how fast AI has spread of late, it's fair to say "the future is already here." At nearly every conference I've attended in China recently, an AI panel is a fixture. AI experts are routinely called up to speak even at luncheons and dinners. China gives every impression of a country that has gone all-in on its AI rise, determined to go head-to-head with the United States on this front. This conference's entire panel lineup was built around AI. What follows is a brief look at the conference's key discussions, along with a review of how AI is transforming international security, and what Korea's strategy should be in the AI era.

| Changing Patterns of Warfare in the AI Era

AI is already widely used in manufacturing through Physical AI, in applications such as robotics and production process management. In the security domain, AI's impact is most visible in weapons and the changing character of warfare. Waging war first requires grasping battlefield conditions such as enemy movements, and intelligence gathered on the battlefield passes through human cognitive processes to become the basic material for policy decisions.

AI naturally makes intelligence collection and processing far faster. In traditional warfare, data collection, battlespace management, data analysis, and target identification were major challenges. With AI, situational awareness has improved, and the speed and accuracy of intelligence analysis, target identification, war planning, and weapons employment decisions have become far easier. AI has sharply lowered the cost curve of war, reshaped states' strategic psychology by providing new means of deterrence and punishment, and redefined the offense-defense balance in ways that favor great powers. In particular, agentic AI, through a process of recursive self-improvement, now operates and evolves faster than a human-in-the-loop system. For this reason, AI is a factor that adapts quickly to, and accelerates, changes in the nature and conduct of war.

Of course, AI is not yet flawless. A frequently cited example is the bombing of a girls' school in Iran. Early in the war, on February 28, 2026, a girls' elementary school in Minab, southern Iran, was destroyed in a U.S. airstrike, killing at least 175 people, including children and teachers. Most experts attribute the massive civilian casualties to the U.S. military's failure to update outdated targeting information, the site had once housed an Islamic Revolutionary Guard Corps base, before carrying out the strike. The building had been converted into a civilian school back in 2016, but the information update was disregarded. The episode is a lesson that while AI compresses human cognitive processes on the battlefield, the final decision must still rest with humans. AI further compresses the OODA loop (observe, orient, decide, act), but this comes with side effects such as automation bias, perception-action gaps, and inflexible mission fixation.

Strategic stability could also be shaken if AI is applied to nuclear weapons operations. The Netflix film "A House of Dynamite" illustrates just such a possibility. The film depicts a razor's edge crisis triggered by an unidentified nuclear missile launched toward Chicago. It closely portrays the political panic and confusion inside the White House and military leadership, locked in fierce dispute over whether to retaliate with nuclear force, all while the identity of the launcher, and even whether the threat is real or false information, remains unknown.

Drawing on the experience and lessons of its troops' involvement in the Ukraine war, North Korea is likely to begin applying AI to areas such as drones and cognitive warfare. This would pose a serious security threat to South Korea. North Korea still lags far behind the United States and China in AI innovation, but it is gradually incorporating AI into its asymmetric military strategy. There is no public evidence that Russia has directly transferred advanced military AI algorithms or foundation models to North Korea. Russia may instead be providing something no less valuable, including battlefield experience from the Ukraine war, drone tactics know-how, electronic warfare tactics, ISR integration, operational data, and military modernization experience.

South Korea has also begun applying AI in the military domain.¹⁾ Examples include the operation of an AI-based DMZ surveillance system and coastal surveillance AI, the AI-enabled Army Tactical Command Information System (ATCIS), the next-generation Allied ROK-U.S. Joint Command and Control System (AKJCCS), and AI-based decision support systems. South Korea's push to adopt AI stems from three main factors: preparations for the transfer of wartime operational control (OPCON), a shrinking pool of military manpower due to population decline, and the growing North Korean nuclear and missile threat. The United States, a leader in AI, is reported to have worked with frontier AI firms, deploying Claude (Anthropic) and Palantir (Maven) in the Iran war. As the United States moves to make full use of AI in military operations, South Korea's armed forces have little choice but to accelerate their own AI adoption for the sake of interoperability within the alliance.

South Korea is not among the world's top-tier AI powers, but it is ranked in the upper tier, roughly fifth to tenth globally. North Korea, by contrast, lags well behind South Korea in overall national AI capability, but its concentration of AI research on the military and cyber domains means the risk it poses should not be underestimated.

| U.S.-China Rivalry for AI Dominance

The two countries currently leading global AI development are the United States and China. Today's U.S.-China rivalry for AI dominance has moved well beyond a purely technological contest and taken on the character of an all-out struggle for national security and economic survival. The United States designs and leads, while China pursues and expands. OpenAI (o3), Google (Gemini), and Anthropic still set the pace at the frontier, but the gap between the top U.S. and Chinese models has narrowed to three to six months. Since the DeepSeek shock, Chinese firms such as Alibaba (Qwen) and ByteDance (Doubao) have flooded the market with ultra low cost, highly efficient models, coming to dominate the open source ecosystem and showing particular strength in reasoning specialized models. This can be summed up as "the gap narrows, the front lines multiply." While the performance gap at the very top has closed rapidly, the two countries are in effect aiming at different targets. The United States, drawing on overwhelming private investment, top tier frontier model performance, and its edge in semiconductors and data center infrastructure, is focused on building the "smartest model." China, by contrast, is focused on state directed resource allocation, open source diffusion, and integration into applications and manufacturing such as robotics and industrial automation, aiming to distribute "good enough" models widely and cheaply.

The policy environment is also highly fluid. Nvidia announced it had resumed production for the Chinese market after taking orders from Chinese customers for its H200 processor. Yet on May 31, 2026, the U.S. Commerce Department's Bureau of Industry and Security (BIS) issued guidance requiring licenses for sales to China headquartered firms, and on July 14, 2026, Nvidia removed more than half of its Asian customers from its whitelist to prevent transshipment. Easing and tightening have alternated repeatedly.

The most frequently cited metric for comparing U.S. and Chinese AI capability is Stanford HAI's AI Index.²⁾ What follows is a comparison of American and Chinese AI capabilities across the categories that define the race for AI dominance.

First, the most commonly compared dimension is model performance, or frontier capability. As of March 2026, the performance gap between the top U.S. and Chinese models on the Arena leaderboard stood at just 2.7 percent, with Anthropic's Claude Opus 4.6 leading the top ByteDance model by 39 Elo points. Across four industry standard benchmarks used to evaluate and compare the intelligence, reasoning, and multimodal capability of large language and vision models, MMLU, MMMU, MATH, and HumanEval, the gap that stood at 17.5 to 31.6 percentage points at the end of 2023 had narrowed to 0.3 to 8.1 percentage points by the end of 2024.³⁾ The Boston Consulting Group (BCG) suggests that a structural split is taking hold, with the United States betting on frontier performance and capital intensive compute and China betting on application diffusion, open source, and robotics, raising the possibility that the global AI ecosystem could split in two. This suggests that competition over standards and platforms is becoming just as important as competition over performance.

Second, in terms of capital investment, global corporate AI investment reached 581.7 billion dollars in 2025, up roughly 130 percent from the previous year. U.S. private AI investment stood at 285.9 billion dollars in 2025, 23 times China's 12.4 billion dollars. This figure reflects only the private capital gap, however. China's state guidance funds are estimated to have channeled roughly 184 billion dollars into AI firms between 2000 and 2023, so the gap in total investment is likely narrower than the private capital figures suggest. On private capital alone, the United States holds an overwhelming lead, but that gap narrows somewhat when measured by total national investment.

Third, in terms of model output and the corporate ecosystem, the number of notable models produced in 2025 stood at 59 for the United States and 35 for China. The leading contributing firms globally were OpenAI (19), Google (12), and Alibaba (11), in that order. With a Chinese firm already ranking among the global top three, the United States retains a quantitative edge, but the gap is narrowing.

Fourth, in AI related research and patents, China leads in the volume of published papers, citation counts, and patent registrations, while the United States retains its edge in high impact patents. China leads in the share of global AI papers published (23.2 percent) and AI patent applications filed (60 percent), while the United States leads in patent citation impact, at 6.9 times the global average. China accounts for 74.2 percent of global AI patents and 17.8 percent of papers, yet U.S. patents generate more than half of all forward citations. In short, China leads on quantitative indicators, while the United States leads on qualitative, or influence based, ones.

Fifth, on talent, the United States still holds the world's largest pool of AI talent, but the pace of new talent inflow has fallen to its lowest level in roughly a decade. The number of AI researchers moving to the United States has dropped 89 percent over seven years and 80 percent in the past year alone, a decline some analysts link to H-1B visa restrictions under the Trump administration. The United States remains ahead on talent stock, meaning cumulative holdings, but is weakening sharply on talent flow, meaning new inflows. Talent acquisition is expected to become a key variable in medium to long term competitiveness going forward.

Sixth, on semiconductors and compute infrastructure, the United States operates 5,427 data centers, more than ten times the combined total of every other country. Huawei's AI chip yield rate stands at just 5 to 20 percent, far below Nvidia Blackwell's 60 to 80 percent, and SMIC remains stuck at the 7 nanometer node, putting it two to three generations behind TSMC's 3 nanometer process. Adding in the advanced AI chip manufacturing capacity of the United States and its allies, the combined lead over China is estimated at 35 to 38 times on a quality adjusted basis. The overwhelming dominance of the United States, together with Taiwan's TSMC, represents the most solid and unclosing gap of all.

Finally, in the real world application of AI, including robotics and manufacturing, and in open source and global diffusion, China holds a clear lead. China accounted for 54 percent of global industrial robot installations in 2024. Its robot density remains at roughly half that of the United States, but at the current pace of growth it is projected to overtake the United States by 2030. As of July 2025, China also accounted for 40 percent of the world's open source LLMs, and its corporate AI adoption rate of 86 percent exceeds the roughly 55 percent seen in the United States. Although China's share of global AI investment stood at only about 17 percent as of the second quarter of 2026, it is projected to process more than half of the world's token throughput. China is, in effect, leading the diffusion of low cost, open AI into markets such as the Global South.

Looking ahead, the U.S.-China rivalry for AI dominance is likely to narrow in the short term, over the next one to two years, but the United States structural advantage will probably hold. The private research firm Recorded Future, through its Insikt Group, assesses that when core indicators such as government and VC funding, industrial regulation, talent, technology diffusion, model performance, and compute capacity are taken together, China is unlikely to sustainably overtake the United States by its target date of 2030.⁴⁾ The Boston Consulting Group (BCG) suggests that a structural split is taking hold, with the United States betting on frontier performance and capital intensive compute and China betting on application diffusion, open source, and robotics, raising the possibility that the global AI ecosystem could split in two. This suggests that competition over standards and platforms is becoming just as important as competition over performance.⁵⁾

Even so, the U.S.-China rivalry is expected to grow tighter still, with compute and export controls as the key variable. In its own scenario analysis, Anthropic projected that if China's access to EUV (extreme ultraviolet) lithography is cut off and gaps in DUV related maintenance and servicing are also closed, Chinese semiconductor makers would be unable to produce chips in the volume and quality needed to challenge U.S. compute superiority. Under this scenario, if the United States releases a model marking a major performance leap in 2028, China might not reach comparable capability until 2029 or 2030. This projection comes from Anthropic, an interested party, so the possibility of optimistic bias should be kept in mind. There is also countervailing evidence that, despite tightened export controls, substantial volumes of advanced chips continue to reach China through smuggling and other channels.⁶⁾

In its 14th Five Year Plan (2021 to 2025), China identified technological innovation aimed at self reliance and self strengthening as a core task in pursuit of "building a fully modernized socialist country." Under this plan, China has set a goal of "comprehensive intelligent transformation," integrating AI into 70 percent of core industries by 2027, 90 percent by 2030, and 100 percent by 2035. This reflects a clear strategic choice to compete on the pace of adoption and industrial transformation rather than on top tier model performance alone.

The U.S.-China contest for AI dominance is evolving into something broader than a technology race. It is becoming a contest of institutional resilience across the whole of each nation. Looking at performance alone, the narrative that the U.S.-China gap is disappearing has gained traction. But on structural indicators such as infrastructure, capital, and patent influence, the American advantage remains solid. China's strength lies not in catching up but in setting a different competitive frame built around application, diffusion, and cost efficiency, and this could translate into substantial influence over global standards competition going forward. The three most important variables likely to shape the balance over the next three to five years are the actual effectiveness of export controls, the pace of energy infrastructure expansion, and whether the flow of talent reverses course.

On data scale and governance, China operates on a domestic scale while the United States operates on a global one. Their governance models also differ, with China emphasizing institutional norms and the United States emphasizing a security oriented approach. In compute capacity, covering chips, infrastructure, and services, and in talent acquisition, the United States currently holds the advantage. Their governance patterns and systems also differ, contrasting technology driven versus scenario driven approaches, and market first versus value oriented priorities. Ultimately, the outcome of the U.S.-China AI competition will be decided by who builds the greater institutional resilience, meaning a stronger innovation ecosystem, and the stronger technological foundation, meaning a stronger technology ecosystem.

| Emerging Security Threats in the AI Era

AI is not only changing the character of warfare. It is also spreading into a range of domains and becoming a source of new security threats. Two areas where concern has already become visible are AI enabled attacks on critical infrastructure and cognitive warfare.

First, attacks on and destruction of critical infrastructure include efforts to disrupt the foundations of economic and social life, such as power grids, transportation, telecommunications networks, and e-commerce. A Japanese expert who attended the international conference at the University of International Relations in Beijing cited Retsu Wakasugi's 2013 novel "Genpatsu White-Out" to illustrate the danger AI could pose if used to attack critical infrastructure. The novel is a work of faction, believed to have been written under the pen name Wakasugi Retsu by a serving senior Japanese bureaucrat. Its central storyline exposes the corrupt ties among politics, bureaucracy, and business, and the misconduct of the so called "nuclear mafia," as the Japanese government reversed its "zero nuclear power" policy and pushed to restart reactors in the aftermath of the Fukushima disaster. The book caused a considerable stir in Japanese society when it was published. On its own, the novel's narrative has nothing to do with AI technology. At the time of its publication in 2013, Japanese society was voicing strong calls for "zero nuclear power," a full exit from nuclear energy, in the wake of Fukushima. Today, however, the global tech industry is doing the opposite. As massive AI data centers run by big tech firms consume enormous amounts of electricity, industry is abandoning its earlier retreat from nuclear power and pivoting sharply back toward expanding it, in the name of stable electricity supply. The novel accused the "greed of the nuclear mafia" of driving the restart of reactors. In an odd inversion, today's "AI power crunch" hints at a similar connection, this time raising the possibility of AI enabled terrorism. Whiteout is originally a meteorological term describing a phenomenon in which snow or sand renders the entire field of view uniformly white, causing a total loss of orientation. In the novel, it refers to a catastrophe in which terrorists blow up transmission towers in the dead of winter, cutting off the power supply and freezing an entire city into paralysis amid brutal cold.

Attacks on critical infrastructure are, of course, nothing new. Stuxnet infiltrated an Iranian nuclear site and helped disable the Natanz facility in 2010. The remote shutdown of Ukraine's power grid is widely attributed to Russia's Sandworm group. The Barakah nuclear plant in the UAE came under physical attack from three drones on May 17, 2026. The UAE's Ministry of Defense said at the time that it shot down two of the drones, but the third struck a generator located outside the plant's inner perimeter, sparking a fire. What makes AI dangerous in this context is that it sharply multiplies the vectors available for such attacks. Volt Typhoon is an advanced persistent threat (APT) hacking group that U.S. and Western intelligence agencies assess to be linked to the Chinese government. Rather than simply stealing industrial secrets, it is known for pre-positioning itself for extended periods inside U.S. critical infrastructure, including power grids, telecommunications, water and sewage systems, and transportation networks spanning aviation, rail, and ports, before activating. There is a documented case in which the power grid on Guam was disabled during Typhoon Mawar.

Second, another domain where AI is having a major impact is cognitive warfare, carried out through AI enabled virtual reality (VR) and augmented reality (AR). Cognitive warfare treats the human cognitive space itself as a new operational domain. It refers to operations that deliberately spread disinformation or manipulated information to influence the perceptions and decisions of individuals and societies. In an age of information overload driven by the spread of the internet and social media, truth has gone missing, and distrust of governments and institutions has grown more dangerous still as it converges with AI. One expert who studied the lessons of the Russia-Ukraine war assessed that AI has shifted the center of gravity of hybrid warfare toward cognitive warfare. AI has moved the battlefield from territory into the domain of human cognition, and shifted the object of war from territory, to society, to the human cognitive space itself.

AI is also being used by extremist groups for propaganda, recruitment, and virtual attack training. AI powered metaverse platforms are said to be creating new opportunities, and thereby new tools, for terrorist organizations. Cases of AI enabled propaganda are already widespread online. There has been a rise in the generation and distribution of neo-Nazi, antisemitic, and racist imagery on social media, along with cases in which text, images, and video supporting the Islamic State (IS) have been generated and distributed by breaching servers.

On Christmas Day in 2021, a 19 year old named Jaswant Singh Chail broke into Windsor Castle armed with a crossbow in an attempted assassination of Queen Elizabeth II. Chail had exchanged more than 5,000 messages with a girlfriend named "Sarai." Sarai was not a person. She was a chatbot.

Taken together, the uncontrollable new patterns of warfare AI is bringing about, the widespread harm that could result from attacks on critical infrastructure, and the reckless spread of disinformation and false information across a chaotic cyberspace, naturally push the discussion toward the norms and regulations needed to ensure the "responsible use" of AI.

| Considerations for International Norms and Regulation

International concern over the negative side of AI is now converging on a single question: how should it be regulated? Artificial intelligence remains an area without any centralized global regulation or control. As AI use has grown explosively, demand for governance has surged, but governance frameworks are only just beginning to take shape. Many countries are eager to use AI, but preparation for its risks remains inadequate, and it is only the major powers that are racing to stake out regulatory ground first. The output generated by technological innovation is moving far faster than the rules and institutions meant to govern it. Given how far human governance awareness and experience lag behind the pace of AI development, user caution remains, for now, the only real line of defense.

In the AI era, frontier AI companies have come to play a role as security actors in their own right, taking on part of what was once purely a national security function. These firms now perform security roles that governments themselves cannot. In the past, governments alone determined security outcomes. Today, governments have grown heavily dependent on the role these firms play. One expert described this situation as "distributed security governance." Global tech firms such as Anthropic (Claude), Palantir (Maven), Google, OpenAI, and Nvidia already occupy an important position in security discussions.

AI clearly offers major benefits for security, the economy, and technological productivity, but no effective means yet exists to control its side effects or enforce responsible use. One Chinese expert argued that at least four paradoxes arise between the explosive growth in AI use and calls to regulate it: security control versus innovative vitality, national security versus the public interest, the divergence between transnational coordination and national level implementation, and ethical empathy versus corporate interest. These are the core tensions running through the debate. He suggested that the world should lower unrealistic expectations around AI use and regulation, work to maximize common ground, and begin with small but substantive forms of cooperation. To this end, he advised identifying AI related security red lines and defining risk thresholds, while building a more open and inclusive global governance system.

The trouble is that rather than moving toward an integrated global norm or regulatory regime anytime soon, the world appears headed instead toward a deepening "AI divide," an extension of the broader U.S.-China rivalry for dominance. Amid the ongoing contest between the United States and China centered on AI technology and semiconductors, an international AI cooperation body led by China, with a total of 29 countries including Russia and Brazil, has now been formally launched. At the G7 summit held in France in June 2026, the CEOs of firms including Anthropic and Google DeepMind proposed building a U.S. led AI cooperation framework. In the end, it was China, not the United States, that launched such a body first. Discussion of AI governance within an international cooperative framework has continued on the Western side as well. At that same June 2026 G7 summit, leaders of major American tech firms proposed forming a U.S. led international framework to discuss AI norms and technical standards. OpenAI CEO Sam Altman even called for building a global standards system modeled on the International Atomic Energy Agency (IAEA). In response to these moves, China's state news agency Xinhua reported on July 17 that representatives from 29 countries, including China, signed an agreement in Shanghai the previous evening to establish the World AI Cooperation Organization (WAICO). This can be read as China moving to actively expand its own influence in the U.S.-China contest over AI technology and international norms, by launching its own cooperative body. In his keynote address at the opening of the "2026 WAIC and AI Global Governance High Level Conference" held in Shanghai, President Xi Jinping said that AI development "should not be a solo performance by a single country, but a symphony achieved through international cooperation," adding that the world should jointly oppose the excessive expansion of the concept of national security in the AI field and move quickly to establish a global AI governance system built on consensus. The remarks are widely read as directed at the United States, which has been tightening its restrictions on China in the AI and semiconductor domains.⁷⁾

Reuters assesses this speech not as a routine technology policy announcement, but as a sign that the U.S.-China rivalry for AI dominance is expanding beyond a technology race into a contest over the international order itself. China is pursuing three goals at once: technological leadership, by pushing open source AI and its own AI firms to strengthen competitiveness; normative leadership, by taking the lead in setting international AI rules and standards; and diplomatic leadership, by building a China centered AI cooperation bloc anchored in the Global South. This stands in clear contrast to the U.S. strategy of deepening cooperation with allies around advanced AI and semiconductor supply chains.⁸⁾

This trend suggests that competition over AI governance is likely to intensify going forward, meaning the U.S.-China rivalry is likely to move in earnest beyond technology and into a full contest over international norms and standards. At the same time, support for developing countries in AI is likely to become a key instrument of diplomatic competition, so the race between the United States and China to court the Global South can be expected to grow even more intense.

| What Should Korea Do?

According to the Stanford AI Index 2026, South Korea ranked fourth in the world in overall score, behind only the United States, China, and India, and placed fifth globally on Oxford Insights' Government AI Readiness Index, where it was recognized for particular strength in policy and governance. By category, South Korea ranks first in the world in AI patents per 100,000 people, at 14.31, and is assessed to rank among the top three globally in the number of notable AI models produced. In Stanford's 2024 national AI vitality assessment, South Korea also ranked seventh, behind the United States, China, the UK, India, the UAE, and France.

At the same time, South Korea's greatest weakness lies less in any individual technology gap and more in the small scale of its frontier AI ecosystem and the limited amount of capital it can mobilize. Korea ranks 35th among OECD countries in AI talent and is experiencing a net outflow, and there remains a clear gap between research and development output, measured in patents and models produced, and the country's actual human capital base.

The core slogan of the Lee Jae Myung government's AI development strategy is the goal of becoming one of the "Big Three" AI powers, or "AI G3." In September 2025, the government launched the National AI Strategy Committee, chaired by the president, as the control tower for driving AI development, and in January 2026 elevated it to a statutory body under the Basic AI Act. The committee oversees the establishment of the nation's AI vision and mid to long term strategy and coordinates policy across ministries.

Given this landscape, what capabilities does South Korea need, and where should it focus, if it hopes to keep pace with the leading countries rather than fall behind in the global AI race?

First, Korea needs to expand its national compute capacity. Core infrastructure in the AI era includes GPUs, data centers, electricity, and telecommunications networks. Compared with the United States and China, South Korea has so far lagged in AI data centers and large scale computing investment. This is precisely why the government has been working since 2025 to secure GPUs and expand national AI computing infrastructure. Developing frontier AI models requires tens of thousands to hundreds of thousands of GPUs, large scale data centers, and a stable power supply. Semiconductor production, in turn, requires vast amounts of water and a reliable, round the clock electricity supply.

Energy in particular is emerging as a new bottleneck. In the United States, the International Energy Agency (IEA) projects that data center electricity demand will more than double between 2024 and 2030, reaching 426 terawatt hours and accounting for roughly 9 percent of total U.S. electricity demand. The Brookings Institution has termed this the "electron gap," one capable of reshaping the U.S.-China compute balance, and points out that securing energy, electricity above all, is emerging as a variable no less important than semiconductors themselves.⁹⁾

The outcome of the AI race is increasingly likely to hinge not just on securing semiconductors but on the ability to secure energy, and in the long run, sustainable energy is likely to become a core element of AI competitiveness. Until now, the AI race has been understood mainly as a contest over securing semiconductors, or GPUs. Going forward, however, a shortage of electricity supply could become the single greatest constraint on AI development. The pace at which AI demands more electricity is outrunning what the energy industry can deliver. Building power plants, transmission networks, and transformers takes the energy industry years, while the AI industry needs large scale data centers within months, or at most a year or two. In other words, the most immediate problem is that the pace of growing AI demand and the pace of growing power supply are simply out of sync. Korea's AI competitiveness depends not only on securing GPUs but on its ability to guarantee a stable power supply. Data center siting policy and power grid expansion need to be tied directly to national AI strategy. The realistic solution lies in finding the right energy mix, combining renewables, energy storage systems (ESS), gas power, fuel cells, and eventually small modular reactors (SMRs).¹⁰⁾

Samsung Electronics and SK Hynix plan to invest 800 trillion won in the Honam region to build four semiconductor fabs. Operating a single semiconductor fab requires at least 1 gigawatt of electricity, roughly the generating capacity of one nuclear reactor. The Honam region is currently home to the Hanbit 1 through 6 reactors, located in Yeonggwang County, South Jeolla Province. Running all four planned fabs in Honam would require roughly 6.3 gigawatts of power. The government's position is that because renewable energy, mainly solar, accounts for a large share of power generation in the Honam region, making supply highly variable, securing a large scale baseload source such as nuclear power, capable of running uninterrupted around the clock, is essential. Given low local acceptance in the Honam region, the government is reportedly leaning toward adding two large reactors on unused land within the existing Hanbit plant site in Yeonggwang, rather than seeking an entirely new location. Samsung is reportedly considering Gwangju Air Base as a candidate site for its semiconductor plant, raising concerns that the safety controversy once seen over Seongnam Air Base during the construction of Lotte Tower could resurface.

Second, building out Korea's AI industrial ecosystem, and securing top talent in particular, is an urgent priority. The fact that Korea ranks near the top on patent and model production indicators, yet 35th among OECD countries on talent and losing people to outflow, is a signal that the very researchers producing these results are leaving the country. Unless Korea addresses compensation, research environment, and visa policy for attracting talent from abroad together, and builds a research ecosystem capable of retaining top tier talent, improvements in patent and paper indicators are unlikely to translate into real industrial competitiveness. Mobilizing private capital also matters. The fact that Korea ranks just 12th globally in private AI investment is a structural problem that cannot be solved by expanding the government budget alone. The key lies in using policy finance, such as deep tech and AI startup funds and the National Growth Fund, as seed capital to draw in follow on investment from private venture capital.

Third, rather than trying to win the kind of frontier model performance race the United States and China are running, Korea needs to focus on AI applications and operations tailored to its own strengths. Securing "technological sovereignty" through sovereign AI carries real significance for data sovereignty and security, but rather than competing head on for the top tier of frontier models, it is more realistic for Korea to concentrate on specialized, lightweight applications tied to industry. Building on its strength in HBM and memory, Korea should focus on AI semiconductors and inference infrastructure. AI inference refers to the stage at which a trained AI model receives new data, makes judgments, and produces results in actual service operation. While raw computational performance matters most during the training stage, power efficiency, cost savings, and processing speed matter far more in the inference domain, where AI services run in everyday settings such as smartphones, PCs, autonomous vehicles, and the cloud. This is opening up new markets, driving surging demand for NPUs (neural processing units) and custom AI chips from AMD, Intel, and major tech firms that can serve as alternatives to Nvidia's costly GPUs.

Another area where Korea can specialize, and arguably its strongest bet, is manufacturing based physical AI. Semiconductor fabs, automobiles and autonomous driving, shipbuilding and autonomous navigation, robots and humanoids, batteries and energy systems, and defense, space, and disaster response should all be fostered as strategic industries at the national level. Korea should also actively consider a strategy of fusing AI with K-defense, which has drawn considerable international interest of late. Rather than pursuing general purpose models that try to cover every domain, Korea would do better with a strategy of building world class AI within specific industries. Promising areas include medical imaging and drug discovery, shipbuilding and maritime, education, defense and intelligence analysis, and the production and translation of K-content. Even where Korea relies on foreign general purpose models, it should aim to build a structure in which Korean firms control the industrial data and service layers.

Fourth, Korea needs to actively participate in shaping international AI norms. Korea remains a latecomer in the AI technology race compared with the United States and China, but it has played a fairly active role as a norm entrepreneur in global AI governance. International AI norms are not forming along a single track but simultaneously across multiple layers. The EU AI Act, entering full implementation in 2025 and 2026, classifies AI systems by risk level and imposes corresponding obligations, and has positioned itself as a de facto candidate for a global standard. The 2023 G7 Hiroshima Process, comprising international guiding principles and a code of conduct, along with the OECD AI Principles, also serve as reference points for policy coordination across countries. In 2024, Korea co-hosted the AI Seoul Summit together with the United Kingdom, a follow-up to the UK's 2023 Bletchley Park AI Safety Summit. Korea is also an active member of the Global Partnership on AI (GPAI).

Korea needs to position itself not as a consumer of norms but as a designer of them. There are many areas where Korea can contribute to AI norm setting. Building on the momentum of the 2024 AI Seoul Summit, Korea should work toward the permanent institutionalization of the Seoul Process. Good examples include AI safety assessment and joint international testing, interoperability among AI regulatory regimes, safety standards for manufacturing, robotics, and physical AI, and human control and accountability over military AI. Military AI in particular is an area where Korea has already built up real experience in norm leadership. Korea hosted the 2024 Seoul REAIM Summit, the high level meeting on the Responsible AI in the Military Domain, which produced a blueprint for action covering compliance with international law in military AI, human responsibility and control, reliability, explainability, and risk assessment. Given the deepening U.S.-China rivalry extends into the normative space as well, Korea should also prepare for the possibility that it will be asked to selectively participate in China led international AI discussions and standard setting processes, even while maintaining AI cooperation with the United States. Korea has room to serve as a bridge, in its capacity as a middle power, in the fields of AI safety, ethics, and international norms. This connects naturally with the AI global governance diplomacy Korea has been pursuing.

1. Suon Choi, "South Korean Military-AI Integration: Opportunities and Risks," APLN(Asia-Pacific Leadership Network), March 2026.
2. Stanford University Human-Centered Artificial Intelligence (https://hai.stanford.edu/), 2026 Artificial Intelligence Index Report (2026) 참조.
3. The four benchmarks referred to are MMLU (Massive Multitask Language Understanding), MMMU (Massive Multi-discipline Multimodal Understanding), MATH (a dataset of 12,500 challenging competition level math problems drawn from contests such as the AMC and AIME), and HumanEval (a coding benchmark created by OpenAI, consisting of 164 hand written Python programming problems).
4. Recorded Future, by Insikt Group, "Measuring the U.S.-China AI Gap," May 8, 2025 (https://assets.recordedfuture.com/insikt-report-pdfs/2025/ta-2025-0508.pdf).
5. Nikolaus Lang, Sylvain Duranton 외, "The Great Divide: How the US and China Are Splitting the AI World," BCG Institute, June 30, 2026 (https://www.bcg.com/publications/2026/us-and-china-ai-strategy-causing-global-ai-divide).
6. "2028: Two scenarios for global AI leadership," ANTHROPIC, May 14, 2026 (https://www.anthropic.com/research/2028-ai-leadership).
7. “中 주도 29개국 AI동맹체 출범... 習, 1국 독주 아닌 교향곡돼야," 『매일경제』, 2026.07.18 (https://www.mk.co.kr/news/it/12101102).
8. "Xi pitches China as leader of new global AI order, challenging US dominance," Reuters, July 17, 2026.
9. Ryan Hass and Patricia M. Kim, "How will the United States and China power the AI race?" Brookings Institution, January 8, 2026 (https://www.brookings.edu/articles/how-will-the-united-states-and-china-power-the-ai-race/).
10. Kim Sung-woo, "Energy for AI," The Korea Times, July 12, 2026 (https://www.koreatimes.co.kr/opinion/20260712/energy-for-ai).

※ The opinions expressed in 'Sejong Focus' are those of the author and do not represent the official views of Sejong Institute.