NVIDIA said on November 8 that it is expanding its collaboration with Korea’s Ministry of Science and ICT (MSIT) across the country’s sovereign AI ecosystem. According to NVIDIA’s posts, the work now covers AI semiconductors, data center infrastructure, talent, startups, models, agents and security. Two technologies anchor the announcement: NVLink Fusion, which could let Korean chip designers plug into NVIDIA’s platform, and Dynamo, NVIDIA’s inference software, which is being applied to Korean open models.

What NVIDIA announced

The company published three related posts on November 8 and 9. The main one describes cooperation with MSIT and with the consortia chosen under Korea’s AI for All initiative: three selected consortia and four leading Korean foundation model builders. NVIDIA says it will work with them on training and certification programs, hackathons and technical seminars aimed at improving GPU utilization.

None of NVIDIA’s posts we reviewed give GPU counts, investment amounts or delivery dates. That limits what can be concluded: this is a framework for cooperation, not a disclosed purchase order.

NVLink Fusion is NVIDIA’s program that lets other companies’ processors connect to its interconnect. NVIDIA says the collaboration lays groundwork for Korean NPU and XPU companies to explore joining that ecosystem. The post also mentions a separate collaboration, announced the same day, between SK Telecom and Rebellions, a Korean AI chip company, with MSIT support.

An NVIDIA DGX GB200 rack-scale AI system
NVIDIA rack-scale systems such as the DGX GB200 are the kind of infrastructure the Korean consortia build on. Photo: Pokiiri / Wikimedia Commons, CC BY-SA 4.0

For Korean chip designers this is a trade-off. Access to NVIDIA’s interconnect can make custom accelerators easier to deploy next to NVIDIA GPUs in the same racks, but it also ties them to NVIDIA’s platform. NVIDIA describes the effort only as exploratory.

Dynamo for Korean open models

The second post focuses on software. NVIDIA Dynamo is an inference framework for scaling models across many GPUs. NVIDIA says it is publishing “recipes” that bring Korea’s leading open models to production, and that it works with the consortia and model builders to optimize and scale inference. For model builders, serving cost is often the barrier between a published checkpoint and a usable service, so tuned deployment recipes are a practical form of help.

The third post covers securing agentic AI in Korea’s ecosystem, in line with the broader scope NVIDIA lists.

How this fits the sovereign AI trend

Sovereign AI means a country building models, compute and skills it controls, rather than relying only on foreign services; our explainer on sovereign AI covers the idea in depth. Korea is a notable case because it has strong memory and chip manufacturing alongside domestic model builders, so cooperation spans hardware to applications.

It also sits against the background of US export rules. As we reported in our explainer on AI chip export controls, access to advanced accelerators is shaped by policy, which gives countries a reason to diversify their chip supply. Partnering with NVIDIA while helping local NPU makers is consistent with that hedge. We note this as context; NVIDIA’s posts do not mention export controls.

What it means for developers and companies

Korean startups and model builders gain training, certification and optimization help rather than money. Companies deploying Korean open models may see faster, cheaper inference if the Dynamo recipes perform as described. Chip startups such as Rebellions gain a possible path into NVIDIA-based systems, although NVIDIA calls it an exploration.

What to watch next

  • Whether NVIDIA or MSIT disclose GPU volumes, funding or timelines for the consortia.
  • Which Korean NPU or XPU makers actually adopt NVLink Fusion.
  • Published performance numbers for the Dynamo recipes on Korean models.
  • Details of the SK Telecom and Rebellions collaboration.

Key takeaways

  • NVIDIA is widening its cooperation with Korea’s science ministry across chips, data centers, talent, models, agents and security.
  • NVLink Fusion and Dynamo are the two named technologies.
  • The announcement includes no investment amounts or GPU counts.
  • It reflects a wider push for national AI capacity and diversified chip supply.

Frequently asked questions

It is an NVIDIA program that lets non-NVIDIA processors, such as custom NPUs or XPUs, connect to NVIDIA’s interconnect technology. NVIDIA says Korean chip companies can explore joining it.

What is NVIDIA Dynamo?

Dynamo is NVIDIA’s software for optimizing and scaling AI inference across GPUs. In Korea it is being used with consortia and model builders to put open models into production.

How much is NVIDIA investing in Korea?

NVIDIA’s posts do not state an investment amount or GPU count, so none is known from these announcements.