HKT deploys 3.2 Tbps hollow-core fiber backbone for distributed AI training
Hong Kong Telecommunications has deployed a 3.2 Tbps AI-focused data center interconnect backbone using hollow-core fiber technology to reduce latency by approximately 30%. This infrastructure is designed to facilitate distributed GPU training across multiple data centers by supporting advanced networking protocols like RoCEv2 and Ultra-Ethernet Transport.
Key Takeaways
- Hollow-core fiber guides light through air channels to reach propagation speeds roughly 47% faster than standard silica glass.
- The 3.2 Tbps link connects major data center clusters in Lok Ma Chau Loop and Tseung Kwan O to support the upcoming Sandy Ridge facility.
- Native support for RoCEv2 and Ultra-Ethernet Transport allows for distributed GPU clusters without the constraints of proprietary InfiniBand hardware.
- Latency is reduced from approximately 5 microseconds per kilometer in standard fiber to roughly 3.3 microseconds in HKT's new deployment.
Why It Matters
This deployment shifts the bottleneck for large-scale AI from facility capacity to network speed, allowing Hong Kong to treat multiple data centers as a single logical GPU cluster. By using hollow-core fiber and open standards like UEC 1.0, HKT is building a vendor-neutral blueprint for metro-scale AI training that bypasses the traditional trade-off between physical distance and GPU synchronization. This establishes a high-performance interconnect model that other global streaming and cloud hubs must adopt to remain competitive in the training of frontier models. Watch for HKT to expand this 'Superhighway' to the Sha Ling Data Park by late 2026.
Additional Context
The HKT deployment coincides with a broader industrial pivot toward hollow-core fiber (HCF) as existing silica-core networks hit physical scaling limits. Per Microsoft, late 2025 marked a major inflection point for the technology, with the company announcing strategic manufacturing partnerships with Corning and Heraeus to scale HCF production for its global Azure infrastructure. Microsoft has already reported zero field failures across its initial 1,280 kilometers of HCF, recording an operational loss of just 0.091 dB per kilometer—the lowest attenuation ever achieved at production scale. Simultaneously, the networking stack fueling these high-speed links is moving toward open interoperability. The Ultra Ethernet Consortium (UEC) released its 1.0 specification in June 2025, providing a standardized Ethernet-based communication stack specifically designed to handle AI and high-performance computing (HPC) workloads. This open standard, backed by major members including AMD, Broadcom, and Nokia, directly challenges the dominance of proprietary interconnects by introducing advanced congestion management and multipath transmission optimized for GPU traffic. Institutional adoption is also accelerating in other major markets. Per Nokia and BT Group, March 2025 saw a successful trial in London achieving 1.2 Tbps over a 100-km HCF link. Further trials in the UAE by Nokia and e& in early 2026 demonstrated massive bidirectional traffic of 153 Tbps on a single HCF strand. These developments indicate that HCF is maturing out of the laboratory phase and into the standard toolkit for telecom operators building 'AI-ready' digital fabrics to support distributed compute requirements.
Read full article at techtimes.com
Get this in your inbox → Subscribe
Enjoy our coverage?
Add StreamingMeme as a preferred source on Google to see more of our streaming news at the top of your Search results.
Add as preferred source