Hong Kong Telecommunications has announced the city’s first dedicated AI data center interconnect backbone, and the most technically significant part of the announcement isn’t the bandwidth figure — it’s the fiber it will run on.
HKT (SEHK: 6823) announced on July 17, 2026, that it will deploy a 3.2 Tbps AI Data Center Interconnect (DCI) Superhighway across Hong Kong’s major data center clusters, with the first phase connecting the Lok Ma Chau Loop in the Northern Metropolis to the Tseung Kwan O data center cluster. That route will be built on hollow-core fiber — a medium that transmits light through an air-filled channel rather than glass, delivering approximately 30% lower latency per kilometer than conventional single-mode fiber. The announcement matters because a distributed cluster of thousands of GPUs across two or more data centers is only practical when the network connecting them is fast enough that synchronization delays don’t eat the compute budget. Hollow-core fiber is the technology that makes that threshold achievable at city scale.
Why No Single Data Center Can Hold Hong Kong’s Largest AI Ambitions
The structural problem that HKT’s DCI Superhighway addresses is a geography problem disguised as a networking problem.
Frontier AI training workloads — the kind needed to develop or fine-tune large language models and scientific simulation engines — routinely require GPU clusters far larger than any single data center in Hong Kong can currently house. When a training job demands resources that exceed a single site’s capacity, operators must distribute workloads across multiple facilities. But distributed training has a hard networking requirement: the all-reduce communication that passes gradient tensors between GPUs at each training step must complete before the next step can begin. Any latency on the interconnect translates directly into GPU idle time, and GPU idle time at scale is an expensive problem.
Conventional enterprise WAN links between data centers were not designed to meet this requirement. They carry traffic through protocols that tolerate packet loss and queuing delay — behaviors that are catastrophic for synchronized GPU-cluster communication. The result has been a quiet constraint on Hong Kong’s AI ambitions: world-class individual facilities, but no city-scale cluster to speak of.
HKT’s DCI Superhighway is a direct attempt to close that gap. The 3.2 Tbps backbone supports four AI-native networking protocols — Remote Direct Memory Access (RDMA), RDMA over Converged Ethernet version 2 (RoCEv2), Multipath Reliable Connection (MRC), and the Ultra-Ethernet Transport standard — each specifically engineered to move data between GPU nodes with minimal CPU overhead and near-zero packet loss.
Steve Ng, Managing Director of Commercial Group at HKT, described connectivity as the “critical foundation of Hong Kong’s AI future,” saying the network would enable faster model training, real-time data processing, and cross-sector AI innovation while laying groundwork for the Northern Metropolis development.
How Hollow-Core Fiber Works — and Why It Changes the Calculus
Standard optical fiber guides light through a solid glass core. Light travels through silica glass at roughly 200,000 kilometers per second — approximately two-thirds the speed of light in a vacuum. Hollow-core fiber replaces the glass core with a precisely structured air channel, surrounded by a honeycomb-like arrangement of thin glass tubes that act as optical mirrors, confining the light inside the air gap. Because light moves through air at nearly the full speed of light in a vacuum, the latency per kilometer drops by roughly 30 to 35% compared with conventional glass-core fiber.
In practical terms, conventional single-mode fiber has a propagation latency of approximately 5 microseconds per kilometer. Hollow-core fiber brings that to approximately 3.3 to 3.5 microseconds per kilometer. On a metropolitan data center interconnect route, even a few hundred microseconds of round-trip latency improvement can mean the difference between a GPU cluster that communicates as if it is in one room and one that stalls constantly waiting for acknowledgements to return.
Hollow-core fiber carries a second advantage beyond latency: it reduces signal distortion because light interacts with air rather than glass. This lowers nonlinear effects, which in turn means signals can be transmitted at higher power over longer distances without degradation — an important property for extending the practical range of DCI links beyond the roughly 60 kilometers that conventional fiber DCI can reliably serve.
Microsoft has already deployed more than 1,280 kilometers of hollow-core fiber carrying live traffic in its Azure network as of early 2026, reporting zero field failures and a measured attenuation of 0.091 dB per kilometer — the lowest operational loss ever recorded at production scale. In September 2025, Microsoft announced a manufacturing scale-up through partnerships with Corning and Heraeus, with a stated goal of reaching 15,000 kilometers deployed by late 2026.
What RoCEv2, MRC, and Ultra-Ethernet Transport Actually Do
The protocols HKT has announced support for are not marketing terms — they describe a specific and recently standardized architecture for AI cluster networking that represents a decisive shift from the networking designs that powered the previous decade of data center infrastructure.
RDMA — Remote Direct Memory Access — allows one GPU node to read from or write to the memory of another GPU node directly, bypassing the CPU on both ends. This is essential for the all-reduce communication patterns that dominate large model training, where gradient tensors must be exchanged among all participating nodes at each step.
RoCEv2 brings RDMA semantics to standard Ethernet networks, which matters because it opens GPU cluster networking to the open, multi-vendor Ethernet ecosystem rather than locking operators into proprietary InfiniBand infrastructure. The Ultra Ethernet Consortium released its UEC 1.0 specification in June 2025 — a 560-page open standard that adds multipath transmission, intelligent congestion management through Network-Signaled Congestion Control, and AI-specific transport optimizations designed to replace the Priority-Based Flow Control workarounds that earlier RoCEv2 deployments required.
MRC — Multipath Reliable Connection — is a protocol co-developed by AMD, OpenAI, Microsoft, and other industry partners that incorporates path-aware sender-based congestion control and round-trip-time-aware window adjustment, drawn directly from the UEC 1.0 NSCC specification. It allows large GPU clusters to distribute traffic intelligently across multiple paths at the path level rather than triggering network-wide backpressure — improving throughput and reducing head-of-line blocking at the scale of tens or hundreds of thousands of GPUs.
The combination of hollow-core fiber’s latency advantage with this protocol stack addresses a problem that conventional WAN links and glass fiber simply cannot solve: keeping thousands of distributed GPUs in tight enough synchronization that distributed training performs comparably to single-site training.
Phase One and the Northern Metropolis Backstory
HKT’s first deployment phase will connect Lok Ma Chau Loop to Tseung Kwan O, with completion targeted by year-end 2026. A second phase will extend coverage to the Hetao area, also in the Northern Metropolis.
The timing is not coincidental. Construction began on March 28, 2026, at the Sandy Ridge Data Facility Cluster — awarded by the Hong Kong government to Hong Kong Range Intelligent Computing Technology Company on a 50-year land grant for HK$581 million. The site covers more than 110,000 square meters in the Northern Metropolis, with a planned gross floor area of approximately 250,000 square meters, nearly all dedicated to high-tier data center use. By 2032, its target computing capacity of 180 ExaFLOPS would represent 36 times Hong Kong’s current aggregate capability — a figure that makes connectivity to the rest of the city’s GPU infrastructure not optional but structurally necessary for the facility to function as a distributed cluster rather than an isolated island.
Sandy Ridge’s first operations are expected to begin in approximately 2029. HKT’s Lok Ma Chau–TKO network, completing by end-2026, will be in place years before Sandy Ridge’s first cabinets are powered on.
Ricky Kwong, Head of Fixed Network Engineering at HKT and CEO of Fiber Link Global, described the first phase as the beginning of a program to upgrade the AI Superhighway across all of Hong Kong’s major data center clusters, with Sandy Ridge specifically named as a target facility.
HKT also operates an AI Exchange node at the Lok Ma Chau Loop — an existing interconnection point that serves as a peering hub for the Northern Metropolis — meaning the DCI Superhighway plugs into an operational fabric from day one rather than requiring clients to construct point-to-point dark-fiber arrangements independently.
What This Upgrade Replaces — and Why Four Times the Bandwidth Isn’t the Main Story
In February 2025, HKT announced Asia’s first 800 Gbps Wide Area Network for AI workloads — a CE2.0-based AI Superhighway targeting supercomputing centers, science parks, and academic institutions. The new 3.2 Tbps DCI Superhighway is a four-fold increase in raw capacity over that predecessor, but the capacity figure is not the primary advancement.
The architectural shift is the move from a Wide Area Network model — in which AI clients share managed enterprise bandwidth — to a true Data Center Interconnect model, in which the network is purpose-built for GPU cluster protocols. WAN networks are engineered for resilience and multi-tenant efficiency. DCI networks are engineered for single-digit-microsecond consistency, lossless transport, and the specific all-reduce communication patterns that define distributed AI training. These are genuinely different design targets.
The hollow-core fiber medium and the AI-native protocol stack — not the bandwidth ceiling — are what enable HKT’s DCI Superhighway to serve as the connective tissue of a distributed GPU cluster rather than merely a high-speed pipe.
Hong Kong’s Wider Data Center Momentum
HKT’s announcement lands in a market that has been accelerating on multiple fronts simultaneously.
Equinix launched HK6, its sixth Hong Kong International Business Exchange data center, on June 16, 2026, with an initial investment of USD 124 million. The purpose-built facility opened with 1,000 cabinets in its first phase, scaling to 3,550 cabinets at full build-out, and is equipped with liquid-cooling technology designed specifically for high-density AI workloads. It connects to the Hong Kong-Shenzhen Innovation and Technology Park, giving tenants cross-border access to the Greater Bay Area ecosystem.
Total data center investment in Hong Kong is projected to exceed USD 5.81 billion by 2031, according to research firm Arizton.
Against this investment backdrop, HKT’s DCI play represents a particular strategic posture: the city’s incumbent telecommunications operator positioning itself not as a passive connectivity provider but as an active infrastructure layer. By offering guaranteed latency, AI-native protocol support, and the operational certainty of a managed service, HKT is betting that AI operators will pay a premium over dark-fiber self-build economics — particularly during the period when hollow-core fiber manufacturing remains specialist and capacity constrained.
That bet has a variable attached to it. Nokia Bell Labs analysts note that while manufacturing costs for hollow-core fiber are falling as volumes increase, the installation ecosystem — connectors, splicers, test equipment — is still optimized for standard conventional fiber. If hollow-core fiber commoditizes rapidly, through programs such as Microsoft’s Corning and Heraeus partnerships, the technical barrier that underpins HKT’s premium service narrows. The analogy Nokia draws is instructive: single-mode fiber was once specialist and expensive before becoming the universal standard. The trajectory for hollow-core fiber is likely similar — the question is the timeframe.
Carriers across Asia-Pacific facing fragmented, land-constrained data center geographies — Singapore, Seoul, Tokyo, Mumbai — will be watching Hong Kong’s experience closely. HKT is not the first telco to reposition as AI infrastructure: TelecomLead analysis from July 2026 notes that SK Telecom, SoftBank, Singtel, and others are pursuing comparable pivots. Whether the managed DCI model can sustain a premium over self-provision is the question this deployment will test in practice.
Frequently Asked Questions
What is hollow-core fiber and why does it matter for AI data centers?
Hollow-core fiber replaces the solid glass core of conventional optical fiber with an air-filled channel. Because light travels faster through air than through glass — roughly 50% faster — hollow-core fiber delivers approximately 30 to 35% lower latency per kilometer compared with standard single-mode fiber. For AI data center applications, where thousands of GPUs across multiple buildings must exchange data in sub-millisecond synchronization windows, that latency reduction can determine whether distributed training is economically viable. Without it, GPU idle time from network stalls can consume a significant portion of compute capacity. Hollow-core fiber is entering commercial deployment now, with Microsoft having deployed more than 1,280 kilometers in its Azure network by early 2026.
Why do distributed GPU clusters need special network protocols, not just fast fiber?
Fast fiber addresses the physical latency of signal propagation — the time it takes a signal to travel from one data center to another. But GPU-to-GPU communication also requires the network to be lossless, to bypass the CPU entirely (via RDMA), and to handle the all-reduce communication pattern of training jobs without head-of-line blocking or packet-drop retransmission. Conventional internet protocols tolerate packet loss and queuing delay; distributed GPU training cannot. The protocols HKT’s DCI Superhighway supports — RoCEv2, MRC, and the Ultra-Ethernet Transport standard (UEC 1.0, June 2025) — are specifically architected for GPU cluster communication, enabling thousands of nodes to stay in sync across multiple buildings rather than being co-located in one.
What is Sandy Ridge and why does the DCI Superhighway matter for it?
Sandy Ridge is Hong Kong’s largest planned AI data center development: a 110,000-square-meter site in the Northern Metropolis awarded in March 2026 to Range Intelligent Computing Technology Company on a 50-year land grant. It is expected to reach 180 ExaFLOPS of computing capacity by 2032 — 36 times Hong Kong’s current total. Sandy Ridge will not begin operations until approximately 2029. HKT’s DCI Superhighway, targeting completion by end-2026, is effectively pre-wiring the city’s connectivity layer for Sandy Ridge years before its first GPU racks are installed, meaning the network needed to operate Sandy Ridge as part of a distributed cluster will already exist when it opens.
Will hollow-core fiber become a commodity technology, and what does that mean for managed DCI services?
Hollow-core fiber manufacturing is maturing rapidly but remains constrained. The installation ecosystem — specialized splicers, connectors, and test equipment — is still optimized for conventional single-mode fiber. Manufacturing partnerships such as Microsoft’s agreements with Corning and Heraeus are accelerating production capacity, but hollow-core fiber production at the volumes needed for widespread deployment remains a bottleneck. Nokia Bell Labs analysts draw a parallel with single-mode fiber, which was once an expensive specialist technology before becoming the universal network standard. If hollow-core fiber follows the same trajectory, the technical barrier supporting premium managed DCI services will eventually narrow — making the next few years a window in which carriers who move early can lock in customer relationships before the technology commoditizes.