Nvidia remains overwhelmingly dominant in the global AI infrastructure market as countries race to develop their own sovereign large language models, but South Korea's SK Telecom and AI chipmaker Rebellions are expanding an inference infrastructure ecosystem built around domestically developed semiconductors.
As of July, 92% of sovereign AI large language models worldwide were trained using Nvidia AI chips, according to market research firm Counterpoint Research. The research covered about 170 sovereign large language models developed across more than 80 countries and found that most rely heavily on Nvidia graphics processing units and its CUDA software ecosystem during model training.
Nvidia's dominance in AI training stems not only from the computing power of its GPUs but also from the extensive software and development ecosystem built around CUDA. As a result, South Korean and other global AI companies continue to rely heavily on Nvidia GPUs when developing hyperscale AI models.
The shift to inference
The competitive landscape, however, is beginning to change as the industry's focus shifts from developing models to operating AI services. Inference is the process through which a trained AI model generates answers or other outputs for users. For large-scale commercial services, computing cost and energy efficiency can be as important as raw processing performance.
Counterpoint cited cooperation between SK Telecom and Rebellions as an example of the emerging competition in AI inference infrastructure. The partnership is seeking to expand the use of South Korean AI chips in an infrastructure market that remains centered on Nvidia technology.
SK Telecom has been broadening its use of AI semiconductors through its partnership with Rebellions. The telecommunications company has deployed Rebellions' neural processing units in systems supporting its A. service and has successfully operated its hyperscale A.X K1 AI model using Rebellions technology at an SK Telecom data center. A.X K1 is a 519 billion-parameter mixture-of-experts large language model developed by SK Telecom.
Building an AI factory
The effort is part of SK Telecom's broader strategy to build what it calls an "AI factory," integrating data centers, AI semiconductors, networks and software needed to develop and operate AI models. SK Telecom is seeking to improve the competitiveness of those facilities by combining different types of AI processors, including Nvidia GPUs and domestically developed neural processing units, rather than relying on a single semiconductor architecture.
The company's AI business is also beginning to generate revenue. SK Telecom said its AI data center business generated 136.2 billion won ($92.1 million). The company has emphasized that its AI operations are increasingly translating into commercial revenue and earnings.
SK Telecom plans to develop 5 gigawatts of domestic AI data center capacity in stages beginning in 2029 as part of a longer-term plan to expand its total capacity to as much as 15GW.
Mixed processor strategies ahead
Industry observers expect competition in AI infrastructure to evolve beyond a simple contest between Nvidia and South Korean AI chips. Instead, data center operators are expected to increasingly select different processors for different workloads, using GPUs where their capabilities are best suited for large-scale model training and alternative accelerators where cost and energy efficiency offer advantages in inference.
South Korean AI companies are therefore likely to remain heavily dependent on Nvidia GPUs for training large models while expanding the use of domestically developed chips in inference, following the approach being pursued by SK Telecom and Rebellions.
Why this matters for IT and development
The shift toward inference workloads changes which hardware and software skills matter. For AI for IT & Development professionals, the SK Telecom-Rebellions partnership signals that production systems will increasingly mix GPU and NPU architectures, meaning infrastructure teams need to plan for heterogeneous deployments rather than single-vendor stacks. The A.X K1 model running on Rebellions hardware also shows that Generative AI and LLM workloads can be optimized for cost and energy efficiency at scale, not just raw performance - a factor that will shape procurement decisions as AI data centers expand.
Your membership also unlocks:
