South Korea’s push for domestic foundation models is expanding into computing infrastructure, AI semiconductors and open-weight deployment, raising a broader question about what technological independence should mean in the AI era.
South Korea’s sovereign AI strategy is moving beyond the question of whether the country can build a competitive foundation model. In the first half of 2026, four teams participating in the government’s Sovereign AI Foundation Model project developed large-scale models ranging from 250 billion to 750 billion parameters, with all four released as open source. The government’s evaluation found that the gap between Korea’s leading models and frontier systems had narrowed compared with the project’s first phase. But a domestically developed model is only one component of technological sovereignty.
The harder question is whether Korean companies, public institutions and researchers can control enough of the surrounding stack to decide where AI runs, what data it uses, which hardware supports it, and how easily workloads can move between infrastructure providers. That is why Korea’s sovereign AI strategy is increasingly becoming a strategy for choice and control across the AI stack, rather than simply a race to produce a Korean alternative to foreign models.
The infrastructure challenge is becoming difficult to separate from the model challenge. The Korean government plans to secure more than 50,000 GPUs through its broader AI Highway initiative, including more than 15,000 GPUs through the National AI Computing Center by 2028. The center is intended to become infrastructure for models, AI services and advanced semiconductors, rather than simply another government computing facility.
Korea is simultaneously investing in technologies that could reduce dependence on a single computing architecture. Its K-Cloud R&D program, for example, is developing AI infrastructure based on domestic AI semiconductors, including hardware, software and cloud technologies, with a goal of creating vendor-independent and broadly compatible software. This is significant because AI dependence can emerge at several layers.
- Dependence on proprietary models can create API and pricing constraints.
- Dependence on a single accelerator ecosystem can constrain compute availability and costs.
- Dependence on external infrastructure can complicate data residency and operational control.
- Dependence on foreign software stacks can make domestic hardware harder to commercialize.
Sovereignty therefore becomes less about eliminating foreign technology and more about preventing any single external layer from becoming an unavoidable bottleneck.
Korea Is Betting on Open and Heterogeneous AI
That explains the significance of Korea’s growing emphasis on open models and heterogeneous computing. In July, Korea’s Ministry of Science and ICT signed an agreement with AMD to develop AI infrastructure combining AMD CPUs and GPUs with Korean AI semiconductors. The initiative also includes memory, networking, DPU, CXL, server and orchestration technologies. The underlying objective is to create an open environment in which different computing resources can be combined according to workload requirements.
This is a different interpretation of sovereignty from building everything domestically. It is closer to building interoperability. For Korean enterprises, the ability to deploy an AI workload across different combinations of GPUs, NPUs, cloud environments and on-premises infrastructure could ultimately matter more than whether every component is manufactured in Korea.
Korea’s sovereign AI models also illustrate why open weights are becoming strategically important. Upstage’s Solar Open 2, for example, has 250 billion total parameters but activates 15 billion per token. With quantization, it can run on two NVIDIA H200 GPUs and supports a context window of up to one million tokens. Its weights are available for commercial use, fine-tuning and deployment on an organization’s own infrastructure.
SK Telecom’s A.X K2 and LG AI Research’s K-EXAONE 2.0 similarly demonstrate that Korea’s sovereign AI program is producing models designed for broader deployment rather than remaining confined to government-funded research. K-EXAONE 2.0, for instance, was released under the Apache 2.0 license and supports 10 languages.
For enterprises, the distinction matters. An open-weight model can be evaluated against internal data, adapted for specific applications and deployed within controlled environments without automatically sending sensitive information to an external model provider.
Localization Is Becoming a Competitive Layer
Sovereignty also increasingly means developing AI around data and operating environments that global models may not fully capture.
A recent example is NAVER Cloud’s cybersecurity AI consortium, which brings together 33 organizations and is developing two approximately 700-billion-parameter models using 4,000 GPUs. Its training dataset is expected to include around 830 TB of real-world data from Korean critical infrastructure and industries. Field trials are planned across power, finance, telecommunications, semiconductors, defense, aerospace, and science and technology.
The significance goes beyond Korean-language performance.
The model is being designed around Korea’s actual cybersecurity environment, infrastructure and threat landscape. That points toward a broader direction for sovereign AI: local data and domain knowledge may become as strategically important as local model development.
Open Models Also Raise the Question of Hardware Independence
The question of openness does not stop at model weights. For enterprises adopting AI at scale, the ability to choose where and on what hardware a model runs can become an equally important part of technological independence.
This is particularly relevant as Korea develops a more heterogeneous AI computing ecosystem. The country’s partnership with AMD is aimed at combining CPUs, GPUs and Korean AI accelerators with memory, networking, servers and orchestration software. That approach reflects a broader shift away from treating AI infrastructure as a single-vendor stack and toward an environment where different computing resources can be matched to different workloads.
Open-weight models can strengthen that flexibility because they can, in principle, be deployed outside the infrastructure controlled by the model provider. For enterprises, that can create more room to optimize computing costs, keep sensitive workloads within controlled environments and adapt infrastructure as hardware options evolve.
While conversing with KoreaTechToday, Eugene Cheah, CEO and Co-Founder of Featherless.ai, highlighted this connection between open models and hardware choice, arguing that specialized models can reduce dependence on individual chip and cloud ecosystems.
“Specialised open models present another crucial advantage – the ability to have hardware and computing independence. Relying on a single vendor’s chip ecosystem or cloud monopolies creates unsustainable overhead for growing businesses; whereas open-weight architecture enables companies to deploy their models on alternative hardware platforms. Moving away from closed ecosystems also saves processing costs because open-weight solutions are significantly more cost-effective. These benefits also enable everyone to develop local AI solutions in their own native languages, built for their culture and ecosystem, rather than trying to fit a vendor’s culture and language into yours.”
The broader implication is that model openness and hardware diversity can reinforce each other. An open model may give an enterprise greater control over deployment, but that control is less meaningful if the organization remains structurally dependent on a single computing or cloud environment. Conversely, a diversified hardware ecosystem becomes more valuable when models can move across it without being locked to one provider.
For Korea, this makes interoperability an important part of the sovereignty equation. The objective is not necessarily to replace every foreign technology, but to ensure that domestic enterprises have enough alternatives across the AI stack to make infrastructure choices based on performance, cost, security and workload requirements.
Sovereignty Does Not Mean Building Everything Alone
Korea’s approach also shows why AI sovereignty should not be confused with technological isolation. The country’s collaboration with AMD demonstrates that domestic capability can be strengthened through global partnerships, provided those partnerships contribute to a more open and diversified infrastructure environment. Korea can simultaneously develop domestic models, AI chips and software while working with global semiconductor and technology companies.
The strategic objective is therefore not necessarily to replace every foreign component. It is to ensure that Korean organizations have meaningful alternatives. That could mean choosing between models, deploying them on different hardware, keeping sensitive workloads within controlled environments, adapting models using domestic data, and moving workloads across infrastructure without rebuilding the entire AI stack.
Sovereign AI Is Becoming a Stack
Korea has already demonstrated that it can produce large, competitive foundation models. The next stage is considerably broader.
Sovereign AI will ultimately be tested not by the nationality of a model, but by how much control Korean organizations have over the layers surrounding it: models, data, compute, chips, cloud infrastructure, software and deployment.
That makes Korea’s current strategy less a race to build one national model and more an attempt to create an ecosystem where no single external provider controls the path from model to machine. In the AI era, technological independence may not mean owning everything. It may mean having enough choice, interoperability and control that no single dependency can determine what Korea is able to build.
