SK Telecom and Rebellions Expand Korean AI Chip Infrastructure
On Aug. 7, SK Telecom and Rebellions demonstrated SK Telecom's A.X K1 sovereign language model running on Rebellions' RebelServer infrastructure, including a multi-user agent service. UPI reports that SK Telecom has deployed Rebellions NPUs for its A. service and operated the 519 billion-parameter mixture-of-experts model at an SK Telecom data center, while Nvidia remains dominant in sovereign-model training.
SK Telecom and Korean AI chipmaker Rebellions have expanded Korean AI inference infrastructure, demonstrating SK Telecom's A.X K1 language model on Rebellions' RebelServer system. Menlo Times reports that the demonstration included a map-based AI agent handling simultaneous user requests, such as locating pharmacies open during a weekend.
UPI reports that SK Telecom has deployed Rebellions neural processing units, or NPUs, in systems supporting its A. service and has operated A.X K1 using Rebellions technology at an SK Telecom data center. The outlet describes A.X K1 as a 519 billion-parameter mixture-of-experts, or MoE, language model. Menlo Times refers to the model as Korea's largest 500 billion-parameter LLM, an apparent rounded figure.
A domestic inference stack
According to Menlo Times, the companies ran the model on a single RebelServer configuration and presented the result as comparable to global high-end GPU server performance. The report attributes the claimed efficiency to A.X K1's MoE architecture, Rebellions' optimized software stack, and distributed-processing techniques designed for real-time requests.
Rebellions CEO Sunghyun Park told Menlo Times that the demonstration validated the use of a domestic LLM and Korean-built NPU for large-scale AI services. He added that Rebellions would continue working with SK Telecom on Korean AI infrastructure.
Training remains Nvidia-centered
The deployment occurs against a markedly different training-market backdrop. UPI, citing Counterpoint Research, reports that 92% of sovereign AI language models surveyed as of July were trained on Nvidia chips. Counterpoint's research covered roughly 170 sovereign models in more than 80 countries, according to UPI.
UPI attributes Nvidia's training advantage to both GPU compute and the CUDA software ecosystem. It also reports that SK Telecom is combining Nvidia GPUs with domestic processors as part of what the company calls an "AI factory," spanning data centers, semiconductors, networks, and software.
For ML infrastructure teams, the reported demonstration illustrates a narrower but consequential competitive arena: inference hardware can be evaluated around serving throughput, latency, energy use, software compatibility, and operational reliability rather than training performance alone. Comparable sovereign-stack efforts commonly face the practical challenge of proving those properties across production workloads, not only controlled demonstrations.
Key Points
- 1SK Telecom and Rebellions demonstrated A.X K1 on RebelServer in a multi-user inference workload using domestic Korean hardware.
- 2UPI reports Nvidia trained 92% of surveyed sovereign LLMs, underscoring CUDA's continued dominance even as alternative inference stacks emerge.
- 3For production teams, alternative inference platforms are typically judged on throughput, latency, efficiency, compatibility, and reliability across sustained workloads.
Scoring Rationale
The deployment is a notable example of a sovereign AI infrastructure effort using domestic NPUs for large-model inference. It is particularly relevant to teams assessing non-Nvidia serving hardware, although the evidence here centers on a reported demonstration.
Sources
Public references used for this report.
Practice with real Telecom & ISP data
90 SQL & Python problems · 15 industry datasets
250 free problems · No credit card
See all Telecom & ISP problems
