NVIDIA and Korea expand sovereign AI work from chips to agent security

NVIDIA has outlined a wider partnership with South Korea's sovereign AI effort, spanning chip connections, software for serving AI models, Korean-language models and security for AI agents. Most details come from NVIDIA's own blog posts, and its one performance figure is NVIDIA's claim and has not been independently verified; the account of Korea's model competition comes from The Korea Times.

What did NVIDIA agree with Korea's science ministry?

NVIDIA says it is broadening its work with the Ministry of Science and ICT (MSIT) beyond AI chips and data centers to include talent, startups, models, agents and security. According to NVIDIA, the work lays the groundwork for Korean NPU and XPU chip companies to explore joining its NVLink Fusion ecosystem, and includes training and certification programs, hackathons and technical seminars. NVIDIA also says it announced a tie-up with SK Telecom and Rebellions, backed by MSIT, on sovereign AI infrastructure for Korean models and workloads built on NVLink Fusion; the details are in a separate press release we did not review.

This builds on NVIDIA's earlier announcement at the APEC Summit in Gyeongju, where it said MSIT planned to deploy up to 50,000 of its latest GPUs for sovereign AI through NAVER Cloud, NHN Cloud and Kakao, beginning with 13,000 Blackwell GPUs, and that Korea's sovereign AI would be supported by "more than a quarter-million NVIDIA GPUs," a total that includes AI factories planned by Samsung, SK Group, Hyundai Motor Group and NAVER.

What are the Dynamo recipes?

Under Korea's AI for All initiative, NVIDIA says it will collaborate with three chosen consortia and four of Korea's leading foundation model developers using Dynamo, its open source software for serving AI models across many GPUs and servers. It is also releasing ready-made deployment recipes for four Korean open models: SK Telecom's A.X K2, LG AI Research's K-EXAONE 2.0, Upstage's Solar Open 2 and Motif Technologies' Motif-3, tuned for NVIDIA B200 GPUs, and says they deliver up to 2x performance out of the box. It gives no baseline for that figure, and it does not say whether these are the same four developers it is working with under AI for All.

Why does Motif stand out?

One of the recipe models comes from Motif Technologies, which was eliminated from the government's national AI foundation model project. The Korea Times reported in late August 2026 that Upstage, SK Telecom and LG AI Research advanced to the project's next stage, while Motif Technologies was dropped after the second-stage evaluation even though its Motif 3 model scored highest on the Artificial Analysis Intelligence Index (47, versus 37 for Upstage, 35 for SK Telecom and LG AI Research's 31).

The evaluation weighted benchmarks at 40 points, expert assessments at 35 and user evaluations at 25, the paper said. Motif, noting that its model received the lowest overall score despite leading the benchmark, filed a formal objection asking for detailed scores and a reassessment, while saying it was not trying to overturn the result. The ministry released only the highest and average scores in each category, citing possible harm to companies. It defended its criteria, saying the project values both performance and practical use, with more weight on model independence after the first stage; that usability's weight in expert reviews was raised to 15 points from 10; and that the criteria were agreed in advance with the participating companies and made public. ICT Minister Bae Kyung-hoon said the ministry needs to "fundamentally consider" whether a system set up a year and a half ago still suits AI trends that change every few months. The ministry said it would decide on the objection within 15 days; none of the sources we reviewed reports the outcome.

The dispute adds to earlier fairness concerns, according to the paper: in January the government changed its original rules to let an extra team join through a separate selection round, after Naver Cloud and NC AI were eliminated, and Motif entered the project through that round.

NVIDIA's Dynamo post does not mention the elimination. NVIDIA's earlier APEC announcement listed NVIDIA itself as a collaborator in MSIT's Sovereign AI Foundation Models project, alongside Upstage, SK Telecom, NC AI, NAVER Cloud and LG AI Research; the sources do not say whether that is the same competition The Korea Times describes, or how it relates to AI for All.

What is new for AI agents?

NVIDIA says Hyundai Motor Group is exploring its NemoClaw blueprints and OpenShell runtime, which NVIDIA describes as a secure runtime that sandboxes AI agents and applies policy-based controls to them. AIM Intelligence, an AI security firm in Korea, is integrating its Starfort agent control platform with OpenShell, with the goal of combining OpenShell's access controls with Starfort's ability to spot malicious instructions and data leaks. NVIDIA adds that AIM Intelligence is weighing an early proof of concept for the public energy sector.

On models, NVIDIA says LG AI Research is opening API access for K-EXAONE-2.5, which was built with NVIDIA Nemotron datasets such as Nemotron-Personas-Korea, an open dataset of 7 million entirely synthetic Korean personas built on official Korean statistics.

What should readers take from it?

Most of what NVIDIA describes is plans and exploration rather than systems in use, and its single performance figure comes without a baseline. The announcements show NVIDIA hardware and software running through Korea's push to build domestic AI models and infrastructure, at a time when the government's model competition faces questions over how ministry scores and benchmark results line up.