KT bets the AI buildout needs its own lanes: 800Gbps dedicated lines, 8x the old ceiling
KT announced on Wednesday that it is launching dedicated artificial-intelligence network lines with bandwidth of up to 800 gigabits per second — eight times the 100Gbps maximum its enterprise dedicated lines previously topped out at. The new product runs on an all-photonic network architecture designed to keep latency near zero between the GPU clusters and AI data centers now spreading across South Korea.
Eight times the ceiling
The offering comes in two tiers, 400Gbps and 800Gbps, aimed squarely at AI data center (AIDC) operators and companies building large-scale AI infrastructure. KT says its roughly 840,000 kilometers of nationwide optical cable let it design shortest and dual paths between data centers or sites, "minimizing latency and ensuring stability." The lines use an APN (All Photonic Network)-based, optics-centered structure — photonics end to end, minimizing the optical-to-electrical conversions where latency typically accumulates on long-haul links. KT has not disclosed pricing for either tier.
"The more AI data centers and graphics processing unit infrastructure expand, the more important ultra-fast, ultra-low-latency networks will become," a KT official said, adding that the new lines would help customers "build and scale large-scale AI transformation infrastructure quickly and flexibly," as reported by the Seoul Economic Daily. KT also says it will keep tailoring high-speed, low-latency services to business customers' network environments, per Aju Press — a signal the AI-dedicated line is a product line, not a one-off speed bump.
Why AI traffic gets its own product
Per Aju Press's report of the announcement, KT's rationale is that generative AI adoption and AI-infrastructure investment have made the high-speed, low-latency transfer of massive datasets a real-time requirement — and with AI data centers distributed nationwide, a reliable inter-site network has become key infrastructure in its own right.
The technical case is easy to see. A single large training run has thousands of GPUs exchanging parameter updates after every batch; when those GPUs sit in different buildings, the link between the buildings is effectively part of the computer. Inference is a different strain: millions of users, traffic fanned across regional data centers, where milliseconds decide responsiveness. KT is formalizing what those workloads already demanded — and charging for it as a first-class product rather than generic enterprise bandwidth.
The race for the interconnect
KT is not starting from scratch. In September it said it would turn about 3,500 of its telecom facilities into AI computing hubs, linking a central AI data center with edge nodes near industrial sites for latency-hungry physical-AI workloads. The 800Gbps lines look like the connective tissue for that plan: distributed compute is only as good as the network between its pieces.
Rivals have not answered yet. No directly comparable AI-specific dedicated-line tier has been publicly confirmed by SK Telecom or LG Uplus, according to Telecompaper's coverage of the launch. What to watch next: pricing, actual operator uptake, and whether the all-photonic bet holds up under real GPU-cluster traffic. The pattern is bigger than Korea — the industry is racing to secure not just compute but the networking that connects it, and KT just tried to own that layer first.