This development has created interest in hollow-core fiber, or HCF, as an alternative to conventional solid-core single-mode fiber. Its most widely discussed advantage is lower propagation latency: because most of the optical power travels through an air-filled core rather than solid glass, light can move through HCF more than 30% faster than through conventional silica fiber.

That does not mean HCF is automatically the better choice for every AI data center link. G.652.D remains a mature, standardized and widely supported transmission medium. Whether lower-latency fiber materially improves an AI system depends on link distance, workload behavior, network architecture and the proportion of end-to-end latency caused by propagation through the fiber itself.

What Is G.652.D Fiber?

G.652.D is a widely deployed category of conventional single-mode optical fiber. It uses a solid silica core and is designed with a zero-dispersion wavelength near 1310 nm. It can operate in both the 1310 nm and 1550 nm regions and supports a broad range of Ethernet, coherent, CWDM and DWDM applications.

The current ITU-T G.652 recommendation, approved in August 2024, defines the geometrical, mechanical and transmission attributes of this fiber family. G.652.D is the low-water-peak category, allowing transmission across a wider wavelength range than older G.652 variants.

Its main strength is not that it represents the newest fiber physics. Its strength is the mature ecosystem built around it:

• internationally established specifications and test methods;
• broad availability from multiple fiber and cable manufacturers;
• proven cable constructions for ducts, direct burial, tunnels and buildings;
• standardized fusion-splicing and connector practices;
• familiar OTDR, loss-testing and fault-location procedures;
• compatibility with a large installed base of optical equipment.

For most campus and metro networks, G.652.D is therefore the reference transmission medium against which emerging fibers must be evaluated.

What Is Hollow-Core Fiber?

Hollow-core fiber guides most of its optical power through a central air-filled region surrounded by a precisely engineered glass structure. Modern telecommunications research is focused largely on anti-resonant designs, including nested and double-nested anti-resonant nodeless fiber.

The surrounding glass membranes confine light inside the hollow core without requiring the light to travel primarily through solid silica. This different propagation mechanism creates several potential advantages:

• lower propagation latency;
• much lower optical nonlinearity;
• potentially lower attenuation;
• low chromatic dispersion over useful wavelength ranges;
• reduced Rayleigh backscattering;
• the ability to tolerate higher optical launch powers.

These are properties of the transmission medium, not automatic guarantees of faster applications. HCF does not by itself change an 800G interface into a faster Ethernet interface, eliminate switch latency or remove the need for forward error correction.

Why Light Travels Faster in Hollow-Core Fiber

The speed of light in a medium is determined by its refractive index. In conventional silica single-mode fiber, the group index is approximately 1.46 to 1.47. The resulting propagation delay is about 4.9 microseconds per kilometer. Because the optical field in HCF travels mainly through air, its propagation delay can approach approximately 3.3 microseconds per kilometer.

The approximate latency difference is therefore:

These values are useful for planning, but they represent only fiber propagation. They exclude transceiver processing, FEC, switching, routing, buffering, congestion, protocol overhead and application processing.

HCF vs. G.652.D: Technical Comparison

Is Hollow-Core Fiber Now Lower-Loss Than Conventional Fiber?

The answer requires an important distinction between a fiber record and an installed cable link.

A 2025 paper in Nature Photonics reported broadband HCF attenuation below 0.1 dB/km. On a 15 km fiber, the researchers measured approximately 0.123 dB/km at 1310 nm and 0.095 dB/km at 1550 nm using OTDR. The work demonstrates that hollow-core designs can cross a loss threshold once considered exclusive to solid-core silica fiber. See Broadband optical fibre with an attenuation lower than 0.1 decibel per kilometre.

However, a record attenuation value should not be treated as the guaranteed loss of every commercial HCF cable. Real links also include:

• variations between production lengths;
• losses introduced during cabling and handling;
• bends and installation stress;
• fusion splices;
• HCF-to-solid-core transitions;
• connectors and patching hardware;
• contamination or gases inside the hollow structure.

For example, a 2025 Microsoft research demonstration used HCF sections with measured fiber loss between approximately 0.11 and 0.24 dB/km. The reported splice loss was about 0.1 to 0.15 dB per splice, and additional loss appeared after fiber handling and transfer to practical storage bobbins. Despite these constraints, the experiment transmitted a full C-band system rate of 25.6 Tb/s over 200.5 km without inline optical repeaters. See Unrepeated HCF Transmission over Spans up to 301.7 km.

The correct conclusion is not that all HCF is already lower-loss than all G.652.D. It is that HCF has demonstrated the physical potential to equal or outperform conventional silica fiber while also offering lower latency and lower nonlinearity. Manufacturing consistency and installed-link performance remain central engineering questions.

Does Lower Fiber Latency Improve AI Performance?

Sometimes, but not universally. AI infrastructure contains several different networks, and they do not have identical latency requirements.

Scale-up networks

Scale-up networks connect accelerators inside a server or tightly integrated compute system. Their performance targets are extremely demanding, and they generally use specialized short-reach electrical or optical interconnects. Inter-campus fiber is not a substitute for this local fabric.

Scale-out training networks

Scale-out networks connect large numbers of GPU or accelerator nodes. Distributed training can be sensitive to communication delay because collective operations may require repeated exchanges among participating nodes. A slower communication phase can leave expensive accelerators waiting for data.

However, tightly coupled training clusters are normally designed to keep the highest-intensity communication within a data hall or closely controlled campus fabric. Extending a synchronous training domain across separate campuses introduces more than propagation delay: it also adds switches, optical interfaces, route length, failure domains and congestion risks.

Storage, checkpoint and data replication networks

Inter-campus links may carry training datasets, checkpoints, storage replication and backup traffic. These workloads often value bandwidth, availability and predictable completion time more than the last few microseconds of latency. HCF can still be useful, but its economic case may be weaker if throughput rather than response time is the binding constraint.

Inference and distributed cloud services

Real-time inference pipelines, distributed databases, financial services and interactive cloud applications can be sensitive to round-trip time. For these services, removing tens or hundreds of microseconds from a repeated request path may create measurable benefits, especially over metro-scale routes.

The workload must therefore be evaluated before the fiber is selected. "AI data center" alone is not a sufficient technical requirement.

The Distance Question: When Do the Microseconds Become Material?

At 1 km, replacing solid-core fiber with HCF saves only about 1.6 µs in one direction. At 20 km, the saving increases to roughly 31 µs one way. At 50 km, it approaches 80 µs one way.

Whether this matters depends on the rest of the network. If a link includes multiple switch stages, heavy FEC, buffering and software processing, the propagation improvement may represent only a small fraction of total latency. If the route is a carefully engineered optical express path with minimal equipment delay, the fiber contribution becomes much more visible.

A useful evaluation is to multiply the propagation saving by the frequency of latency-sensitive transactions. This is not a monetary formula by itself. It is a reminder that the same 30 µs saving can be insignificant for a bulk transfer but important when it occurs repeatedly inside a synchronized or interactive workflow.

HCF Does Not Automatically Increase Link Capacity

Latency and capacity are different properties. Replacing G.652.D with HCF while keeping the same transceivers, channel plan and interface rate does not automatically change an 800 Gb/s link into a 1.6 Tb/s link. The interface and optical system still determine the carried data rate.

HCF may nevertheless support future capacity growth because it offers:

• low nonlinearity at high launch powers;
• potentially broad low-loss transmission windows;
• low dispersion in suitable designs;
• reduced nonlinear interaction between WDM channels;
• the possibility of extending unamplified reach.

These advantages can change the design envelope of a transmission system. They should not be presented as an immediate bandwidth multiplier independent of the active equipment.

Deployment Challenges That Matter Outside the Laboratory

A production inter-campus route must be judged as a cable system, not only as a fiber sample. The main challenges are splicing, cabling, field testing, contamination control, repair planning and interoperability.

Splicing and mode conversion

HCF has a different internal geometry and mode-field behavior from standard solid-core fiber. Joining two HCF sections, or transitioning between HCF and G.652.D, requires control of alignment, collapse, contamination and mode coupling. A low-loss laboratory splice is not automatically equivalent to a repeatable field process performed in a closure or equipment room.

Cabling and mechanical behavior

The microstructured glass surrounding the hollow core must survive proof testing, cabling, transport, pulling, bending and long-term environmental exposure. Cable design must protect the fiber without introducing unacceptable attenuation or modal effects.

Testing and fault location

Conventional OTDR relies on backscattered light. HCF's low Rayleigh backscatter is beneficial for transmission but can make standard monitoring and fault-location methods more difficult. The Microsoft long-span experiment used a newly developed long-range OTDR approach with high peak pulse power to address this characteristic.

Gas and contamination control

Because the guiding region is hollow, gas absorption can affect particular wavelengths. The long-span Microsoft experiment observed carbon-dioxide absorption features in part of the C-band. Fiber design, manufacturing, termination and sealing therefore become part of optical performance management.

Repair strategy

A production network needs more than an initial installation method. Operators must define emergency restoration, technician training, spare lengths, approved splice processes and fallback routing. Mean time to repair can be more important than a small propagation advantage if the specialist repair ecosystem is limited.

Standardization and interoperability

G.652.D has a mature international framework. HCF does not yet have an equivalent universal deployment specification covering every fiber, cable, splice, connector and field-test attribute. In October 2025, ITU-T Study Group 15 agreed to investigate HCF technologies and applications to identify potential standardization paths. It also initiated work on optical technologies for AI data center interconnects. See the ITU-T Study Group 15 Executive Summary.

Commercial Maturity: Beyond a Laboratory Fiber

HCF has moved beyond being only a university research subject. Microsoft announced deployment of cabled HCF in its Azure network and described an ecosystem that includes cable, field splicing, installation, testing and compatible data center equipment. See The Deployment of Hollow Core Fiber in Azure's Network.

In September 2025, Microsoft also announced manufacturing collaborations with Corning and Heraeus to expand HCF production. Corning said its work would include fiber, cable and connectivity products needed for an end-to-end network solution. See Corning Collaborates with Microsoft to Advance AI Networks.

These developments are significant because commercialization requires repeatable long fiber draws, stable yields, cabling processes, connection methods and field support, not only an excellent loss measurement on one sample. They do not yet make HCF interchangeable with G.652.D. They indicate that HCF is entering an early deployment and manufacturing-scale phase.

When G.652.D Remains the Better Choice

G.652.D remains the rational default when:

• the link is short and propagation represents little of the total latency;
• conventional optics already meet the required reach and loss budget;
• availability and repairability are more important than microsecond savings;
• the project requires competitive multi-vendor sourcing;
• existing technicians, splicers and test equipment must be used;
• the business workload is throughput-sensitive rather than latency-sensitive;
• no measurable operational value has been assigned to the lower delay.

Calling G.652.D "older technology" can be misleading. It is a mature platform that continues to support modern high-capacity transmission systems. Maturity is an infrastructure advantage, especially for networks expected to operate for decades.

When HCF Deserves Serious Evaluation

HCF becomes more compelling when:

• the physical route is long enough for propagation delay to be material;
• the workload repeatedly crosses the inter-campus link;
• microsecond-level round-trip improvement has measurable value;
• the optical route can minimize switch and processing delays;
• high launch power or low nonlinearity changes the system design;
• the operator can support specialized installation, monitoring and repair;
• a controlled deployment can be qualified before wider adoption.

The strongest early case is not necessarily replacing every campus fiber. It is applying HCF to a limited number of latency-critical optical paths where its physical advantages are visible and measurable.

Questions to Ask Before Selecting HCF

A technically credible comparison should answer the following questions:

1. What is the actual route distance, not just the straight-line distance?
2. How much of current end-to-end latency comes from fiber propagation?
3. Is the target workload sensitive to one-way delay, round-trip time or throughput?
4. How many times does a critical transaction cross the link?
5. What delay is introduced by optics, FEC, switches and protocol processing?
6. What are the guaranteed cabled-fiber and installed-link loss values?
7. How are HCF sections spliced, terminated, tested and repaired?
8. What loss is introduced at each HCF-to-solid-core transition?
9. Which optical bands and transceiver types have been qualified?
10. What fallback path is available if specialist HCF repair is delayed?
11. Which parameters are standardized and which remain supplier-specific?
12. Can the performance benefit be demonstrated at the application level?

These questions prevent the selection process from being reduced to a single "33% lower latency" claim.

Final Assessment

Hollow-core fiber represents a genuine change in optical transmission physics. By moving most of the optical field from glass into air, it can reduce propagation latency by more than 30%, suppress nonlinear effects and potentially achieve attenuation below that of conventional silica fiber.

For AI data center inter-campus links, those advantages are real, but conditional. HCF creates the greatest value when the route is long enough, the workload is sufficiently latency-sensitive and the surrounding network is engineered so that fiber delay remains a meaningful part of the total latency budget.

G.652.D remains the lower-risk and more interoperable choice for general-purpose campus connectivity. Its installed ecosystem, field-service maturity and standardized performance continue to outweigh HCF's propagation advantage in many networks.

The practical decision is therefore not whether hollow-core fiber is universally better than G.652.D. The decision is whether a particular inter-campus path is valuable enough and latency-sensitive enough to justify a new fiber ecosystem.

FAQ

References

1. ITU-T G.652 (08/2024): Characteristics of a Single-Mode Optical Fibre and Cable
2. Broadband Optical Fibre with an Attenuation Lower Than 0.1 Decibel per Kilometre - Nature Photonics
3. Unrepeated HCF Transmission over Spans up to 301.7 km - Microsoft Research
4. The Deployment of Hollow Core Fiber in Azure's Network - Microsoft Azure
5. Microsoft Azure Scales Hollow Core Fiber Production
6. Corning Collaborates with Microsoft to Advance the Performance and Reliability of AI Networks
7. ITU-T Study Group 15 Executive Summary, October 2025