LG AI Research's Lee Moontae: "Sovereignty is the key to superintelligence"

In a conversation about the most distant future, the most immediate, present-day word became the anchor. Lee had also appeared before Bloomberg's cameras at the same event last November, when the video was titled "AI Development Strategy." Same person, same event, same outlet — but in nine months, the central theme shifted from "strategy" to "sovereignty."

The event itself has grown too. This year's AI Summit Seoul & Expo, running from August 19 at COEX, drew 118 companies from seven countries operating 307 booths, according to the Korea International Trade Association. One of the theme tracks on the second day of the two-day conference is "Sovereign AI."

A day earlier, the government trimmed the field to three

The timing looks almost too precise to be coincidence. On August 18, the day before the interview, the Ministry of Science and ICT announced the results of the second-stage evaluation for its "homegrown AI foundation model" program. The three teams that passed were LG AI Research, SK Telecom, and Upstage. Motif Technologies, which had drawn attention for its proprietary architecture, was eliminated.

The program has been designed from the start to narrow the field over time. It began with five teams in August 2025; three survived the first-round evaluation in January; one more was added back in February through a second-chance round, bringing the total to four; and now it's back down to three. The original plan called for narrowing the field to just two teams by the third-round evaluation. However, in announcing this round's results, the Ministry noted there was consensus that the scale of support and the competitive structure itself needed to be redesigned — leaving room for the framework to be revised.

The evaluation is weighted 40 points for benchmarks, 35 for expert review, and 25 for user assessment — meaning human judgment carries more weight than machine-scored metrics. Motif's elimination illustrates this structure well: its model scored highest among the four teams on international intelligence benchmarks, yet lost out on usability and applicability.

The three surviving teams will each be allocated roughly 1,000 GPUs in the second half of the year. Choi Dongwon, Director General for AI Infrastructure Policy at the Ministry of Science and ICT, explained that assuming a six-month lease, this amounts to roughly 40 billion won worth of resources per team. In other words, Lee sat down for his interview the day after his team had just cleared one of these gates.

The government side had already been moving well ahead of this. On August 3, construction began on the National AI Computing Center in Solasido, Haenam County, South Jeolla Province. The project carries a total budget of 2.4065 trillion won, with a target of 15,000 GPUs by 2028 and 50,000 by 2030. Power capacity will start at 40MW and scale up to 80MW.

This year's government AI budget approaches 10 trillion won — roughly triple last year's level. At the groundbreaking ceremony, Deputy Prime Minister and Minister of Science and ICT Bae Kyunghoon said the goal was to make South Korea a "token factory" producing tokens for the world. At a briefing marking his first year in office in July, he also said, "Frontier-class models will become national strategic assets on the level of nuclear weapons."

Bae was the founding president of LG AI Research. The person who once built the models is now, in a role overseeing budgets and GPU allocation, making the same argument.

Sovereign AI ultimately comes down to who holds the keys

Sovereign AI refers to the ability of a specific country or company to process its own data using its own technology, without depending on external infrastructure or models. As concerns over data security and regulatory compliance have converged, the term has become one that governments and companies worldwide are increasingly quick to invoke.

It's not unlike the difference between renting and owning a home. You can cook and sleep in either. But you can't tear down a wall to reroute the plumbing, and you can't hold your ground when the landlord decides to change the terms.

That's the difference between borrowing a model and building one. The gap is invisible in ordinary times — it only shows up the moment contract terms change. Lee himself made a similar point in a lecture last May, noting that "in 2023, the industry's concern was AI's side effects; in 2024, it was AI sovereignty" — suggesting the word had already swept through the industry once before.

K-EXAONE 2.0, which LG AI Research released on July 31, is its answer to that question. Of its 750 billion total parameters, only 37 billion are active at any given time, thanks to a mixture-of-experts (MoE) architecture that calls on only the experts it needs — scaling up size while keeping electricity costs in check.

The first generation, unveiled in January, had 236 billion parameters — meaning the model has more than tripled in size within half a year. The number of layers grew from 48 to 78, and the number of experts per layer from 128 to 256. Rather than training from scratch, the team reused first-generation weights and scaled up from there. Supported languages also expanded from six to ten.

According to the research institute, performance rose to an average of 70.1 points across 24 benchmarks, up from 63.3 for the first generation. That said, these figures are based solely on the institute's own reporting so far, with no third-party verification yet available.

More notable than the numbers is the license. K-EXAONE 2.0 was released on Hugging Face under an Apache 2.0 license, opening it up for commercial use — a departure from the EXAONE line's long history of being restricted to research and educational purposes. Rather than locking the door while talking about sovereignty, the team chose to open it.

More countries are saying the same thing

This trend isn't unique to South Korea. NVIDIA reported that sovereign AI revenue topped $30 billion in fiscal year 2026 (ended January 2026) — more than triple the prior year. Every time a country declares it will build its own model, the bill tends to land at the same company.

The European Union opened applications for its "AI Gigafactories" program on July 30, planning to build up to seven facilities backed by €10 billion in public funding matched with €20 billion from the private sector.

In Japan, SoftBank, NEC, Sony Group, and Honda launched a physical AI foundation model consortium in April. Industry observers expect that up to ¥1 trillion in AI support funding the Japanese government has budgeted over five years starting in fiscal 2026 will flow to this entity. India unveiled three major homegrown models at once at a summit in New Delhi in February.

Back in South Korea, one particularly interesting scene has emerged. Canada's Cohere announced on August 5 that it would establish a Seoul subsidiary and designate it as its Asia-Pacific headquarters — with LG CNS as its Korean partner. The entity is set to open in the fourth quarter of this year.

Within the same corporate group, one arm builds its own model while another partners with a foreign sovereign AI company. It's a scene that shows sovereignty isn't a concept neatly divided along national borders.

The counterargument: "Middle powers can't have frontier models"

Not everyone buys into this narrative. Michael Bhaskar, who leads AI strategy at Microsoft, was asked at a forum in Seoul on June 17 whether middle powers like South Korea would ever have their own frontier-level foundation models. "My answer is no," he said.

His reasoning centers on the cost gap: what four major U.S. tech giants spend on AI infrastructure in a single quarter can exceed the annual AI budget of an average country. It's a point that keeps resurfacing — that national-level investment overlaps with what a hyperscaler spends in a matter of months. Some analysts also argue that the share of AI applications that genuinely require sovereign control isn't actually that large.

Skepticism exists within the industry itself. Cohere CEO Aidan Gomez told a European outlet in June that there's "way too much sovereignty-washing" going on — solutions being repackaged as sovereign AI without actually contributing anything to diversification or resilience.

On August 10, while explaining his company's own open-source strategy, Gomez listed three things people actually want from open-source models: customizability, low cost, and security. The implication: claiming sovereignty requires real control, not just the label.

Editor's take

The first time I downloaded and ran EXAONE, my impression was of "a mid-sized model that's good at Korean." Even when 3.5 introduced a 32B version, and 4.0 added reasoning, the story felt like one of efficiency rather than scale. That matched what Lee himself said in a lecture last May: the focus was on "whether emergent capabilities can be preserved without maximizing parameter count."

Seeing the number 750 billion felt like a shift in direction. Co-head of research Lim Woohyung said much the same right after the team passed the second-round evaluation: "To compete on equal footing with global frontier-class models, scaling up model size significantly was an inevitable strategic choice."

A team that had survived on efficiency has now declared it will compete on sheer scale. Keeping active parameters capped at 37 billion reads almost like a footnote to that declaration — stepping into the ring, but conserving stamina along the way.

Reading Lee's focus purely through the lens of "sovereignty" captures only half the picture. His topic on the summit's opening-day keynote panel was "The Era of AI Building AI." At a conference in June, he argued for a structure in which AI performs tasks, incorporates expert feedback, and improves from there — citing a roughly 25% performance gain from a pilot with the National Pension Service.

EXAONE Foundry, which he leads, is a framework that ties together everything from domain-specific training data generation to evaluation metrics into a single loop. Sovereignty is a question of where that loop sits; superintelligence is a question of how far that loop can spin. That's why it's not so strange, in a conversation about superintelligence, for the topic of sovereignty to come up.

What domestic teams should actually take away from this is the licensing angle. A 750-billion-parameter model released under Apache 2.0 can now be downloaded by anyone and adapted to their own domain data. Regardless of how the government program ultimately plays out, the fact remains: there's now one more resource available for teams to use.

Whether "sovereignty" stays confined to policy documents or becomes a model actually running on someone's own servers will be determined by how many teams pick up that resource. Fortunately, that barrier to entry has dropped noticeably just this month.