SK Telecom, Rebellions expand Korean AI chip infrastructure
Asia Today reports SK Telecom and AI chipmaker Rebellions are expanding an AI inference infrastructure using South Korean semiconductors. Counterpoint Research says 92% of sovereign AI LLMs were trained with Nvidia chips as of July. SK Telecom deployed Rebellions NPUs for its A. service and operated A.X K1, and said its AI data center business earned 136.2 billion won ($92.1m).
How this was made
The 30-second read
Why it matters
It suggests a gradual transition in AI infrastructure from a single-vendor training-centric approach toward workload-optimized heterogeneous compute, with SK Telecom as an example.
Market read
Traders may use this as a signal that inference economics are driving incremental adoption of non-Nvidia accelerators, while Nvidia’s training dominance remains intact.
What to watch
The article lacks details on performance, unit economics, and customer scale for A.X K1 inference, which are key to determining whether this becomes a meaningful procurement shift.
Background
The article cites Counterpoint Research on sovereign AI model training reliance on Nvidia GPUs and CUDA, then describes SK Telecom’s partnership with Rebellions to expand inference infrastructure using domestically developed semiconductors.
Ticker impact
SK Telecom is expanding its AI inference infrastructure by deploying Rebellions neural processing units and operating its A.X K1 model at an SK data center.
Low near-term impact; any repricing would likely depend on follow-on commercial wins and scale of the AI factory deployments.
The piece provides concrete deployment and revenue figures for SK Telecom’s AI data center business, but it does not quantify margins, customer contracts, or near-term capacity ramp beyond planning.
The article frames Nvidia as still dominant in sovereign AI training, with 92% of sovereign models using Nvidia chips and CUDA during training.
Neutral for NVDA; any downside would be gradual and workload-specific rather than a sudden share loss.
The data point reinforces Nvidia’s entrenched position in training, while the competitive change is described as emerging and focused on inference efficiency rather than replacing training demand.
Market effects
Supports a broader thesis that AI infrastructure procurement will split by workload, with GPUs favored for training and alternative accelerators gaining share in inference.
Highlights South Korea’s push for sovereign AI compute, potentially increasing local accelerator adoption in domestic data centers.
Reinforces that Nvidia’s software ecosystem remains a moat for training while inference becomes a new battleground for cost and energy efficiency.
Counterpoint
Mixed-accelerator inference adoption may not materially reduce Nvidia demand because many deployments still require Nvidia for training and for parts of the stack.
Key entities
- companySK Telecom
South Korean telecom operator building an AI factory and deploying Rebellions NPUs for its A.X K1 hyperscale model operations.
- companyRebellions
AI chipmaker whose neural processing units are used in SK Telecom’s AI inference infrastructure.
- companyNvidia
Dominant provider of GPUs and CUDA ecosystem for sovereign AI model training, per Counterpoint Research.
- research_firmCounterpoint Research
Market research firm cited for the 92% sovereign AI training reliance on Nvidia chips.
