Insights

On July 17, 2026, Hong Kong telecommunications flag-bearer HKT officially announced its plan to roll out a 3.2Tbps AI Data Centre Inter-connect (DCI) Superhighway—the first ultra-high-capacity, hollow-core fiber DCI architecture of its scale in Hong Kong.

Representing a fourfold upgrade from HKT’s existing 800Gbps AI Superhighway, this next-generation network is designed to solve the critical physical bottleneck facing modern artificial intelligence: interconnecting thousands of GPUs across geographically distributed data centers to operate as a single, unified supercomputer.

1. Technical Innovations: Hollow-Core Fiber & Ultra-Ethernet Standards

As AI training models and real-time inference workloads surpass the power and space capacity of single data center facilities, infrastructure providers require seamless cross-facility clustering. HKT’s 3.2Tbps DCI Superhighway addresses this through significant optical and protocol advancements:

• Hollow-Core Fiber (HCF) Deployment: The core route will utilize advanced hollow-core fiber optic technology. Because light travels nearly 30% faster through an air-filled core than traditional silica glass fiber, the network achieves an approximate 30% reduction in transmission latency over standard fiber links.
• Next-Gen AI Protocol Support: The 3.2Tbps architecture natively supports enterprise-grade and high-performance computing (HPC) protocols, including Remote Direct Memory Access (RDMA), RDMA over Converged Ethernet version 2 (RoCEv2), Multipath Reliable Connection (MRC), and Ultra-Ethernet Transport.
• GPU Synchronization & Energy Efficiency: Minimizing transmission delay allows clusters containing thousands of GPUs across different sites to stay tightly synchronized during massive parallel processing tasks. This drastically cuts compute wait-times, optimizes total compute cost, and lowers power-per-bit consumption.

2. Strategic Topography: Linking Hetao, TKO, and Sandy Ridge

Phase one of the project focuses on constructing the high-speed optical pipeline along Hong Kong’s primary innovation and data infrastructure corridors:

• Lok Ma Chau Loop to Tseung Kwan O: The primary hollow-core route directly connects HKT’s newly established AI Exchange in the Lok Ma Chau Loop (Hetao / 河套)—a key zone in the Shenzhen-Hong Kong Innovation and Technology Co-operation Zone—to the high-density data center cluster in Tseung Kwan O (TKO / 將軍澳).
• Northern Metropolis Extension: The network is planned to extend to the upcoming megawatt-scale Sandy Ridge Data Facility Cluster (沙嶺數據中心集群) in the Northern Metropolis as new computing facilities come online.
• Completion Timeline: Phase one connecting Lok Ma Chau to Tseung Kwan O is targeted for full operational completion by the fourth quarter of 2026.

3. Executive Perspectives

The initiative underlines HKT’s commitment to building the digital backbone for Hong Kong’s emerging status as an international AI and Innovation & Technology (I&T) hub:

"At HKT, we see connectivity as the critical foundation of Hong Kong's AI future. This next-generation fiber network delivers ultra-low latency, high-speed, and secure data exchange to unlock the full potential of AI development across the city... It also lays vital groundwork for the Northern Metropolis, linking the Hetao area with existing data centers today while paving the way for seamless access to new computing resources as they come online."

— Steve Ng (吳家隆), Managing Director of Commercial Group, HKT

"HKT is firmly committed to supporting innovation in Hong Kong, with the ambition of positioning our city as one of the most advanced I&T hubs in the world. To realize this vision, we are upgrading the AI Superhighway across Hong Kong's major data center clusters. The first phase, connecting Lok Ma Chau to Tseung Kwan O, is targeted for completion by the end of 2026."

— Ricky Kwong (鄺偉圖), Head of Fixed Network Engineering at HKT & CEO of Fiber Link Global

4. Strategic Shift: Sandy Ridge, Hetao, and Tseung Kwan O in Hong Kong’s AIDC Blueprint

Hong Kong’s digital infrastructure is undergoing a fundamental spatial transformation. While Tseung Kwan O (TKO) has historically served as the city's primary data center hub and subsea cable landing gateway, the Hong Kong SAR Government’s Northern Metropolis Development Strategy has catalyzed two massive new AI Data Center (AIDC) clusters directly along the Shenzhen border: the Sandy Ridge Data Facility Cluster (沙岭数据园区) and the Hetao Shenzhen-Hong Kong Innovation and Technology Co-operation Zone (河套深港科技创新合作区).

I. Sandy Ridge Data Facility Cluster: The Mega AIDC Flagship

Located in the North District directly adjacent to the Lo Wu and Man Kam To border control points, Sandy Ridge represents one of the largest dedicated AIDC land allocations in Hong Kong's history.

• Massive Investment & Scale: In March 2026, the government awarded the 11-hectare Sandy Ridge site under a "Two-Envelope" technical tender to Hong Kong Runjiang AI Compute Technology Co., Ltd. , a subsidiary of Mainland compute operator Runze Intelligent Computing Technology Group(润泽智算科技集团股份有限公司). The developer has committed HK$23.8 billion (~$3 billion USD) in capital expenditure over its first three years of development.
• 250,000 m² GFA & 180 EFLOPS Capacity Target: The facility will deliver up to 250,000 square meters of gross floor area, with 90% reserved exclusively for high-density, liquid-cooled AIDC facilities. By 2032, Sandy Ridge is projected to house a staggering 180,000 PFLOPS (180 EFLOPS) of AI compute capacity—equivalent to 36 times Hong Kong's total compute capacity in early 2026.
• CLP Power Grid Infrastructure: CLP Power Hong Kong is constructing dedicated, high-voltage substation connections tailored for multi-megawatt GPU power density, ensuring continuous power reliability for intensive large language model (LLM) training.

II. Hetao Innovation & Technology Zone: The Cross-Border Data & AI Exchange Hub

Straddling the Shenzhen River under a "One River, Two Banks" and "One Zone, Two Parks" framework, Hetao acts as a regulatory and technological sandbox bridging Hong Kong's Lok Ma Chau Loop with Shenzhen's Futian district.

• Cross-Border Data Flows: Hetao serves as the primary testing ground for "Data Express" pilot initiatives, allowing Mainland scientific, industrial, and medical datasets to cross into Hong Kong's open regulatory regime under specialized security protocols.
• Lok Ma Chau AI Exchange Node: HKT's newly launched AI Exchange in the Lok Ma Chau Loop provides the critical optical switching fabric where cross-border data meets high-performance computing assets from both Hong Kong and Mainland enterprise tenants.

III. Comparing Hong Kong's AIDC Triad: TKO vs. Sandy Ridge vs. Hetao

The emergence of Sandy Ridge and Hetao creates a complementary, dual-engine AIDC topology spanning northern and eastern Hong Kong:

IV. Strategic Importance: Next Door to Shenzhen’s Tech Ecosystem

The positioning of Sandy Ridge and Hetao directly along the Hong Kong-Shenzhen border offers profound geopolitical, technical, and economic benefits:

• Ultra-Low Latency to Shenzhen Tech Giants: Positioned mere kilometers from Shenzhen's technology districts (home to Tencent, Huawei, DJI, and Ping An), compute workloads deployed in Sandy Ridge and Hetao achieve sub-millisecond round-trip times (RTT) to Shenzhen's software and hardware development teams.
• The "International Data Port" Concept: By utilizing Hong Kong's free-flow-of-information environment, international common law, and robust intellectual property protections, Sandy Ridge and Hetao act as a dual-facing "Data Port." They allow global firms to aggregate and process Asian data locally while granting Mainland enterprises an international-grade compute launching pad to expand AI services globally.
• Seamless North-South Convergence: As HKT's 3.2Tbps DCI Superhighway links the high-density GPU compute capacity of Sandy Ridge and Hetao in the north directly to the subsea cable landing stations in Tseung Kwan O in the south, Hong Kong effectively unifies its internal compute assets into a single territory-wide, ultra-low latency AI supercomputer.