Korean NPUs now handle 14 million daily AI requests for SK Telecom
Rebellions' ATOM-Max neural processing units have moved from pilot to production, powering four SK Telecom services including call summarization and fraud detection. The company reports 14 million daily requests and over 4 billion tokens processed. The deployment is positioned as complementing GPUs rather than replacing them.
The ATOM-Max rollout followed a staged path: a June 2025 testing agreement led to a December 2025 production deployment for one call-summary feature, with three additional services added sequentially from June 2026 onward. SK Telecom pools dozens of NPUs and hundreds of accelerator cards into shared infrastructure rather than dedicating hardware per workload. The streaming TTS generates audio sentence-by-sentence to reduce perceived delay, while Scam Vanguard's fraud models can be refreshed without interrupting service. The company frames this as complementary to GPU infrastructure, not a replacement.
This deployment could signal that specialized inference chips are becoming viable for real consumer workloads, potentially diversifying the AI hardware market beyond dominant GPU vendors. If Korean-designed NPUs sustain this scale reliably, telecom operators and other large enterprises may consider domestic or specialized silicon options, affecting supply chains and pricing dynamics. However, company-reported figures lack independent verification, so broader adoption depends on demonstrated reliability and efficiency gains over time.