"In the 'brain' of artificial intelligence (AI) models, the United States and China are ahead, but the actual deployment of physical AI is still in its early stages. I think this is exactly where Korea has a big opportunity to seize." (Abhinav Gupta, co-founder of Skilled AI)
At a panel discussion during the country's largest tech conference, "SMARTCLOUD SHOW 2026," held at the Westin Josun Hotel in Sogong-dong, Seoul, on the 26th, global AI industry leaders said Korea must avoid remaining a follower in the AI race by linking its existing strengths—including manufacturing, semiconductors, infrastructure, and talent—to the application of AI on the industrial front lines. The real-world deployment of physical AI, On-device AI, and a sovereign AI ecosystem independent of the United States and China were cited as areas Korea could lead.
Moderating the discussion was Yoon Seong-ro, a professor in the Department of Electrical and Computer Engineering at Seoul National University (former Chairperson of the Presidential Committee on the Fourth Industrial Revolution). Panelists included Abhinav Gupta, co-founder of Skilled AI; Kim Tae-su, corporate vice president for security at Microsoft (professor in the School of Computer Science at Georgia Tech); Lim Hyo-jun, head of the Next-Generation Computing Research Lab at LG Electronics; and Jeong Seok-geun, CEO of SK Hyper (head of SK Telecom AI DC Integration Task Force). They offered views on a range of topics, including strategies for Korea to leap into an AI power and the challenges to solve when applying technology to real industrial settings.
◇ A range of views from leading in physical AI to a No. 1 strategy and building a third force
Professor Yoon opened the discussion with a shared question: "Where should Korea focus in the AI race?" Founder Gupta answered with "rapidly deploying physical AI in industrial sites." He said, "There aren't many physically deployed AI robots in the United States, and in China most robots appear to be for entertainment," adding, "If you combine need and opportunity, Korea can join physical AI experiments first and secure deployment experience that is hard for other countries to accumulate."
Founder Gupta also saw "regulatory innovation" as important for Korea to seize this opportunity. He said, "You shouldn't require 100 certifications to deploy a single robot," advising, "Lower regulations and speed up the process to actual deployment."
Vice President Kim said Korea should set its sights higher. "The strategy to finish third and the strategy to finish first are different," he said. "It would be better to be a bit more ambitious and aim for No. 1." He added, "Among startups developing open-weight models, many have 20 to 30 truly brilliant people," and said that, in the long term, it is important to build a sustainable industrial ecosystem that makes Korea an attractive market for Koreans active overseas.
CEO Jeong advised that a "Korean-style strategy" different from those of the United States and China is needed. "Even if it's third place, let's set the right strategy and get to third quickly," Jeong said. "Demand for a third player outside the United States and China is extremely strong." He added, "If Korea can package everything well from models to semiconductors and data centers, it can fully become a third force." As the United States leads global standards and China builds its own AI ecosystem, demand for sovereign AI could grow in countries that find it difficult to rely entirely on the two nations' technologies. Jeong said that if Korea offers its AI models, semiconductors, and data center infrastructure as one package to these countries, it can establish itself as an alternative outside the United States and China.
Director Lim proposed that Korea should focus strategically on "On-device AI" and "physical AI." The more AI uses sensitive personal data such as video and audio, the more limited it becomes to send everything to the cloud for processing. On-device AI, however, has the drawback of being more limited in computing resources than the cloud, but Lim said advances in technology and model slimming will gradually narrow this gap.
He said, "During the Go match between Lee Se-dol and AlphaGo, dozens of graphics processing units (GPUs) were needed, but now a single inexpensive GPU at home can run a Go AI that performs even better," adding, "In a few years, running AI on actual on-device hardware will come into view." He said, "Korea is very strong in manufacturing and has vast data generated in factories, so there are opportunities in physical AI as well."
◇ Beyond 99.9% accuracy, the key is 'how fast you deploy'
The discussion moved to the conditions needed to deploy physical AI in real industrial sites and the scope of decisions that can be entrusted to AI. Asked, "What is the criterion for judging that physical AI robots can move beyond research and demos into actual field deployment?" founder Gupta presented three main metrics: accuracy, throughput/cycle time, and the effort required to deploy.
Founder Gupta said, "Accuracy is a kind of gate because unless you reach 99.9%, no one wants to deploy." He added, "The most important thing is how much effort is required to deploy robots," saying it is crucial to improve generalization to reduce the data and time needed for deployment.
Asked, "Among vulnerability discovery and reproduction, patch generation, validation, and actual rollout, how far can AI go?" Vice President Kim answered, "Most of the vulnerability discovery and analysis/validation process is already automated." However, "In actual production environments, a lot of information not in the code is needed," he said, adding that you need to consider cloud configuration, network architecture, and firewall settings together to assess real risk. He said that while patch generation, rollout, and validation can be automated technically today, areas with accountability are structured so that people verify and approve.
There was also the view that the higher the autonomy, the more layers of safeguards are needed. Director Lim said, "As you move from general AI to agent AI to physical AI, the impact of failing to maintain safety increases," adding, "Physical AI operates in the physical world and can harm humans, so the risks are much higher." He said a step-by-step safety framework is needed, from model training to checking motion/trajectories and verification just before actual operation.
◇ For robots, 'data' is the bottleneck… For AI data centers, the 'capital problem' must be solved
Specific audience questions closely tied to field deployment followed. Topics covered included physical AI, On-device AI security, and the capabilities talent needs in the AI era—major issues in the tech industry recently.
Asked about the biggest bottleneck in Robotics, founder Gupta answered, "Robotics is a data problem." He said, "If robots aren't useful, they aren't deployed; if they aren't deployed, no data is generated; and without data, they can't become useful again—it's a chicken-and-egg problem." He continued, "I don't think we're ready for household/consumer robots, but enterprise robots are ready," adding, "The biggest bottlenecks are data and large-scale deployment."
Asked, "Beyond the number of GPUs or the scale of infrastructure, what is the key metric for evaluating the success of an AI data center?" CEO Jeong cited fundraising capability as the most important factor. "As an engineer by training, I used to think technology was key, but these days it seems the most important thing is money," he said. "We're not talking trillions of won but hundreds of trillions of won as the amounts needed for infrastructure. What matters is whether you can build a case to persuade financial investors."
He also cited how much efficiency you can achieve with the same capital as a key metric. "Korea has many platform companies and many semiconductor companies," Jeong said. "We can collaborate on how to produce more tokens with the same capital expenditure (CAPEX)."