Hollow-Core Fiber vs Multi-Core Fiber for AI Data Centers: Latency, Density, and the Future Role of G.652.D

Traditional G.652.D single-mode fiber is not disappearing. It is still inexpensive, standardized, globally available, and familiar to almost every fiber installation team. For conventional telecom networks, enterprise links, FTTH, and long-established backbone systems, that combination remains difficult to replace.

AI data centers are different. Large GPU clusters are forcing optical networks to handle two pressures that older network designs could often ignore: microsecond-level latency and extreme fiber-density growth. A fiber type that works well in traditional networks can become physically limiting when millions of optical channels must be routed through racks, rows, buildings, and campus interconnects.

For AI data center fiber planning, the problem is becoming a balance between three budgets: the time budget, the space budget, and the cost budget. Hollow-core fiber improves the time budget by lowering propagation delay. Multi-core fiber improves the space budget by increasing the number of optical paths per fiber. G.652.D remains the cost and maturity baseline. The future fiber plant is therefore unlikely to be a single-fiber story; it will be a layered architecture where each fiber type occupies the network level that matches its strongest constraint.

That is why two newer fiber architectures are gaining attention: hollow-core fiber, or HCF, and multi-core fiber, or MCF. They solve different problems. HCF is mainly a latency technology. MCF is mainly a density technology. Neither should be treated as a simple one-for-one replacement for G.652.D across all network layers.

The real question is not whether HCF or MCF will “kill” G.652.D. The more useful engineering question is: where does each fiber type fit inside future AI data center interconnects?

What Is the Difference Between Hollow-Core Fiber and Multi-Core Fiber?

Hollow-core fiber vs multi-core fiber is a comparison between two different ways of escaping the limits of conventional single-core silica fiber. Hollow-core fiber reduces latency by guiding most optical power through air, while multi-core fiber increases density by placing multiple independent cores inside one fiber. HCF mainly solves time delay; MCF mainly solves space and cable-count pressure.

                                     G.652.D vs HCF vs MCF Fiber Structure Comparison

Hollow-Core Fiber Reduces Latency by Changing the Propagation Medium

In standard G.652.D fiber, light travels mainly through solid silica glass. The silica core has a refractive index of about 1.468, so optical signals travel at roughly 68% of the speed of light in vacuum. That gives G.652.D a propagation delay of about 4.9 µs/km.

Hollow-core fiber changes the basic medium. Instead of guiding most of the optical field through glass, HCF uses a hollow air core surrounded by engineered glass microstructures. In practical hollow-core designs, more than 99.9% of the optical power can propagate through air rather than through solid glass. Because air has a refractive index close to 1.0003, HCF can reduce propagation delay to about 3.35 µs/km.

That is not a small tuning improvement. It is a change in physical path. In the context of AI data center interconnects, the difference between 4.9 µs/km and 3.35 µs/km can matter when multiple network hops and synchronization layers accumulate delay.

Multi-Core Fiber Increases Density by Changing the Fiber Geometry

Multi-core fiber takes a different route. It does not primarily try to make light travel faster. Instead, it places multiple independent optical cores inside the same outer fiber structure.

The current AI data center discussion often focuses on 4-core weakly coupled MCF. In this architecture, four separate cores are integrated inside a standard 125 µm cladding diameter. Each core can be designed to remain optically compatible with the existing G.652 / G.657 single-mode fiber ecosystem.

That compatibility is the key engineering point. MCF does not require every optical signal path to be reinvented. It mainly compresses several single-core paths into one physical fiber, reducing cable count, connector count, pathway congestion, and cable mass.

Why G.652.D Fiber Faces New Limits in AI Data Centers

G.652.D remains the baseline because it is inexpensive, standardized, and easy to deploy. Its cost is often described around $0.10/m, and its installation ecosystem is mature. It also belongs to the long-running ITU-T G.652 family of single-mode optical fiber specifications, which defines characteristics for single-mode optical fiber and cable.

However, AI clusters create a different type of stress. The issue is not that G.652.D suddenly stopped working. The issue is that its two strongest physical assumptions — solid-glass propagation and single-core geometry — become limiting when the network must support synchronized GPU computation and massive optical channel density.

The Latency Ceiling: Solid Silica Propagation and GPU Synchronization

In ordinary web traffic, an extra microsecond per kilometer rarely changes the user experience. A page request that takes 1.5 ms longer is not usually noticeable. GPU clusters are more sensitive because distributed training depends on repeated synchronization.

During All-Reduce, thousands of GPUs may compute a mini-batch and then wait for results to be aggregated across the cluster. If one layer of the network adds only a few microseconds, that may look insignificant. But when several layers and many communication rounds accumulate delay, microseconds can begin to affect effective GPU utilization.

G.652.D has about 4.9 µs/km of propagation delay. HCF can reduce this to about 3.35 µs/km, a difference of roughly 1.54 µs/km. Over 10 km, that is about 15.4 µs of propagation-delay difference before considering switching, serialization, DSP, or protocol overhead.

For traditional networking, that number may look small. For tightly synchronized AI training clusters, it becomes part of the physical-layer budget.

The Space Ceiling: Fiber Count, Cable Weight, and Routing Density

The second limit is physical space. At hyperscale AI data center levels, fiber scale can reach extraordinary levels: up to 20 million fiber channels inside a single data center, more than 1 million fibers between buildings, and cable weights that can reach 100 pounds per foot in extreme cable-bundle cases. A single NVIDIA GB200 NVL72 node has also been described as requiring around 10,000 fibers.

These numbers are not normal enterprise cabling problems. They are pathway, tray, duct, rack, installation, and building-load problems. When physical space becomes the bottleneck, adding more single-core fibers is no longer the cleanest answer.

That is where MCF becomes attractive. A 4-core MCF can combine four optical cores into one fiber. For the same channel count, a representative 144-fiber to 36×4-core MCF comparison shows a 75% reduction in fiber count and about a 45.7% reduction in cable cross-sectional area.

Hollow-Core Fiber in AI Data Centers: Using Air to Reduce Latency

Hollow-core fiber is the more radical technology. Its primary advantage is not just lower attenuation or wider bandwidth. Its most distinctive feature is that it changes where the light travels.

Instead of moving mainly through solid silica, HCF guides optical power through air. This directly attacks the propagation-delay limit of conventional glass-core fiber.

How HCF Reduces Propagation Delay

The physics is straightforward:

The result is about 31% lower latency and a signal-speed improvement commonly described around 47% compared with conventional solid-core single-mode fiber.

                                                HCF Low-Latency Propagation Principle

In a short patch-cord environment, this advantage may not justify the cost. In cross-building DCI, campus interconnect, or latency-sensitive financial networks, it can become meaningful.

Why Lower Nonlinearity Matters More Than It First Appears

Latency is the headline feature of HCF, but the larger engineering change may be its much lower nonlinearity.

In G.652.D, increasing launch power eventually increases nonlinear impairments. Kerr effect, four-wave mixing, and stimulated Brillouin scattering can distort the signal. This is one reason engineers cannot simply raise optical power indefinitely to extend reach.

HCF changes this balance. The nonlinear coefficient is described at about 0.001 W⁻¹km⁻¹, compared with around 1.3 W⁻¹km⁻¹ for G.652.D. That is roughly a 1,000x reduction. With far less optical power interacting with glass, HCF can tolerate higher optical power before nonlinear distortion becomes a limiting factor.

In the DCI comparison used here, HCF supports about 1.5x longer unamplified spans than G.652.D, which can reduce intermediate equipment, power consumption, and potential failure points in multi-building AI campuses.

HCF Performance Metrics: Latency, Attenuation, Dispersion, Spectrum, and Power Handling

HCF should not be evaluated only by latency. Its broader value comes from a combination of propagation speed, low nonlinearity, dispersion behavior, and potentially wider usable spectrum.

Recent hollow-core fiber research reported in Nature Photonics has shown attenuation below 0.1 dB/km across wide bandwidths, reinforcing why HCF is now being taken seriously as more than a low-latency laboratory concept. That does not mean every deployed HCF link will match a record laboratory result. It does mean that HCF has crossed an important credibility threshold.

HCF Commercial Deployment: Production Use, Cost Barriers, and Adoption Path

HCF is already beyond pure research. Microsoft Azure has publicly discussed scaling hollow-core fiber production through manufacturing collaboration with Corning and Heraeus, and HCF has been reported in production use across more than 1,280 km of European Azure data center links. The reported operating data includes zero field failures, 47% speed improvement, and 32% latency reduction.

Another hyperscale cloud operator has also moved into HCF deployment, with links reported across roughly 10 data centers. Financial trading networks have used HCF in production for more than four years, which is consistent with the technology’s strongest early value proposition: in some financial environments, microsecond-level latency differences can affect trading outcomes.

Still, HCF faces severe cost and ecosystem barriers. In the current cost comparison, HCF remains roughly 50–100x more expensive than G.652.D, while its share of global fiber installations is still below 0.1%. In China, reported HCF capacity gaps reach 70%, and the price gap can be far higher than in overseas markets because production remains constrained.

That cost structure makes broad near-term replacement unlikely. HCF’s likely adoption path is staged:

1. Financial trading networks

2. Hyperscaler DCI

3. High-end enterprise interconnect

4. Select telecom backbone use cases

Each step requires lower cost, more standardized testing, easier installation, and broader transceiver support.

Multi-Core Fiber in AI Data Centers: Using Fiber Geometry to Increase Density

MCF is less dramatic than HCF from a physics perspective, but it may be more urgent from a deployment perspective.

MCF does not try to make light travel through air. Instead, it treats physical space as the bottleneck. If a data center cannot keep adding single-core fibers at the required rate, the logical next step is to put multiple cores inside each fiber.

4-Core MCF Structure and 125 µm Cladding Compatibility

A 4-core MCF places four independent cores inside a standard 125 µm cladding. This detail matters because the outer fiber size remains familiar to the existing fiber ecosystem. The goal is not to rebuild every duct, panel, and pathway around a larger fiber diameter. The goal is to multiply optical paths inside the same physical envelope.

The ITU-T G Supplement 87 standardization framework prioritizes weakly coupled multicore fiber with standard 125 µm cladding and backward compatibility with the existing G.65x single-mode fiber ecosystem. That is important because it supports the idea that MCF is not merely a custom specialty fiber. It is being shaped around compatibility with existing single-mode infrastructure.

G.657 is also relevant because G.657 Category A fibers are compliant with G.652 and used across transport, data center, and access environments. For MCF, the broader compatibility logic is that each core can behave like a standard single-mode channel while the overall fiber provides much higher spatial density.

MCF Performance Metrics: Fiber Count, Cable Area, Weight, Crosstalk, and Reach

The most important MCF metrics are not only optical. They are physical deployment metrics: fewer fibers, fewer cables, fewer connectors, less mass, and shorter installation time.

Commercial MCF solution literature also describes four cores inside a standard 125 µm footprint, with up to 4x optical pathway density, up to 75% fewer cables or connectors, and major reductions in cable mass and installation time. These values should be treated as solution-level claims, not universal guarantees for every installation, but they show why MCF is attractive for AI data center cabling.

                                          MCF Density Improvement in AI Data Center Cabling

MCF Ecosystem: Connectors, FIFO, Splicing, Modules, and Standardization

MCF is moving faster than HCF in ecosystem readiness because it does not require a complete change in optical propagation physics. The key components are already emerging across the chain:

Commercial MCF offerings are also beginning to appear as integrated fiber, cable, and connectivity systems rather than only specialty bare fiber. This matters because data center operators usually do not adopt a fiber architecture in isolation. They need connectors, fan-in / fan-out devices, test procedures, installation training, and supply-chain availability.

HCF vs MCF vs G.652.D: Key Engineering Trade-Offs

The easiest mistake is to ask which technology is “best.” That is not how the engineering problem works.

G.652.D, HCF, and MCF optimize different constraints.

Latency, Density, and Cost Are Three Different Problems

G.652.D wins when cost, standardization, and field familiarity matter most. HCF wins when the network is genuinely latency constrained. MCF wins when space, pathway capacity, connector count, cable mass, and installation time become the limiting factors.

That distinction is central. HCF is not a better MCF. MCF is not a cheaper HCF. They solve different layers of the physical network.

Compatibility, Splicing, Testing, and Standardization

HCF has a more disruptive adoption path. It may require new transceivers, different DSP assumptions, new OTDR and test approaches, and new training for field teams. Its physical advantages are strong, but its ecosystem must catch up.

MCF has a more incremental adoption path. Each core can remain compatible with familiar single-mode optical behavior, while the infrastructure around it changes through connectors, FIFO devices, splicing procedures, and standardization.

That is why MCF may become urgent sooner. Its deployment model does not require the entire ecosystem to be replaced at once.

Commercial Timing: Why MCF May Arrive Faster Than HCF

HCF is more exciting from a pure physics standpoint. A 31% latency reduction is easy to understand, and the nonlinearity reduction is even more important for certain long-span designs. But HCF’s cost, manufacturing scale, testing requirements, and standardization gap keep it concentrated in high-end use cases.

MCF is less radical, but more deployable. Because it can preserve more of the existing single-mode ecosystem, its adoption barrier is lower. With 4-core commercial solutions, connector development, FIFO miniaturization, MCF modules, and standardization activity all moving together, MCF could reach broader AI data center use earlier than HCF.

Based on its compatibility path, connector ecosystem, FIFO development, module activity, and standardization progress, MCF could move toward broader commercial adoption around 2027–2028, potentially 3–5 years earlier than broad HCF deployment. That should be treated as a conditional market judgment rather than a guaranteed timeline. The timing depends on standardization, connector supply, module availability, test procedures, and installation training.

Where Each Fiber Fits in an AI Data Center Network

AI data center networks are layered. Each layer has a different bottleneck, so the right fiber choice changes with distance and function.

In this article, the following practical labels are useful:

• Scale-Up: tightly coupled compute expansion over very short distances

• Scale-Out: horizontal expansion inside a building or data center fabric

• Scale-Across: cross-building or campus-level AI infrastructure interconnect

Why MCF Fits Scale-Out and HCF Fits Scale-Across

                                         Layered AI Data Center Fiber Network Architecture

MCF is strongest where the problem is physical density. If thousands or millions of fibers must be routed through trays, ducts, panels, and buildings, reducing the number of fibers by 75% can be more valuable than shaving propagation delay.

HCF is strongest where the problem is time. Cross-building and campus-level links can accumulate enough distance that propagation delay becomes visible in the network budget. HCF is especially relevant when low latency and fewer intermediate powered sites justify the cost.

This is why HCF and MCF should be viewed as complementary. MCF compresses the fiber plant. HCF compresses time.

Could Multi-Core Hollow-Core Fiber Combine Both Advantages?

A future fiber could theoretically combine both ideas: multiple cores, each using hollow-core guidance. Such a multi-core hollow-core fiber would aim to combine HCF’s latency advantage with MCF’s density advantage.

The concept is physically plausible because both approaches involve microstructured fiber design. The barrier is manufacturing complexity. Combining multiple independent cores with hollow-core guidance would make geometry control, loss control, crosstalk control, splicing, connectorization, and yield much harder.

For now, this should be treated as a future research and manufacturing direction, not a near-term data center deployment option.

Supply Chain and Industrialization: Why HCF and MCF Are Scaling Differently

Technical records do not automatically create industrial adoption. A fiber technology must be manufacturable, installable, testable, connectable, and available at a cost that matches its use case.

HCF and MCF are scaling differently because their industrial challenges are different.

HCF: Strong Technical Records but Limited Production Scale

China has reported strong HCF technical indicators, including a 0.05 dB/km low-loss result in 2025, a 7.5 km Hangzhou Unicom pilot in Binjiang, and multiple operator tests for cross-border financial lines.

The gap is production scale. Overseas HCF deployment is more advanced in hyperscaler networks, with Microsoft’s 1,280+ km deployment and another hyperscale deployment involving roughly 10 data centers. China’s HCF capacity gap is reported around 70%, and the price gap can be far higher than in overseas markets because production remains constrained.

The important interpretation is that China’s HCF challenge is not simply technical. It is demand-side and industrialization-side. Without very large procurement orders from Chinese hyperscalers, production scale is harder to build, and cost is harder to reduce.

MCF: A More Complete Industrial Chain Around Fiber, Cable, Connectors, FIFO, and Splicing

MCF looks different. In China, YOFC has been described as participating in ITU-T MCF standardization since 2020, with product coverage across 4 / 7 / 8 / 19-core MCF, continuous drawing lengths of ≥1,000 km, MCF LC and MPO connectors, miniaturized FIFO, splicing solutions, and multiple field deployments.

This is why MCF may be strategically important. It is not only a fiber. It is becoming a system-level supply chain: fiber, cable, connectors, fan-in / fan-out, splicing, test, and field deployment.

Practical Engineering Takeaways for AI Data Center Fiber Planning

The future AI data center fiber plant is unlikely to be built around one universal fiber type. It will be layered.

When G.652.D Still Makes Sense

                                      Engineering Selection Matrix: Time, Space, Cost

G.652.D remains the practical choice where cost, standardization, and deployment maturity matter more than ultra-low latency or extreme density. It will continue to be used in FTTH, many enterprise networks, traditional transport systems, and parts of backbone infrastructure.

It is not obsolete. It is simply no longer the best answer for every AI data center layer.

When HCF Is Worth Evaluating

HCF is worth evaluating when latency is valuable enough to justify cost and ecosystem complexity. That includes financial trading networks, hyperscaler DCI, cross-building AI cluster interconnect, and campus links where lower delay and longer unamplified spans can reduce system complexity.

The caution is clear: HCF requires new thinking around transceivers, DSP, testing, splicing transitions, standards, supply chain, and cost.

When MCF Becomes the Practical Upgrade Path

MCF becomes attractive when the bottleneck is physical density. If cable trays, ducts, patch panels, connector counts, and installation time are limiting growth, MCF offers a direct path to higher fiber density without requiring each optical channel to abandon the existing single-mode ecosystem.

For AI data centers, that makes MCF a strong candidate for scale-out and short-to-medium internal interconnect layers.

FAQ

Is hollow-core fiber faster than G.652.D fiber?

Yes. Hollow-core fiber can reduce propagation delay from about 4.9 µs/km in G.652.D to about 3.35 µs/km, because most optical power travels through air rather than solid silica glass. That is roughly a 31% latency reduction, which can matter in cross-building DCI, campus interconnect, and latency-sensitive AI cluster networks.

Does multi-core fiber reduce latency in AI data centers?

Not in the same way HCF does. MCF mainly improves density, not propagation speed. A 4-core MCF places four cores inside one fiber, so it can reduce fiber count, cable mass, and pathway congestion. Its per-core latency is generally closer to conventional single-mode fiber than to hollow-core fiber.

Why is G.652.D still used if HCF and MCF offer advantages?

G.652.D remains widely used because it is inexpensive, standardized, easy to splice, globally available, and supported by a mature ecosystem. HCF and MCF offer important advantages in specific AI data center layers, but they also bring cost, standardization, testing, connector, and supply-chain challenges.

Which is better for AI data center interconnects: HCF or MCF?

It depends on the bottleneck. HCF is better when the main problem is latency, especially across buildings or campuses. MCF is better when the main problem is physical fiber density, especially inside data center buildings or scale-out fabrics. In large AI campuses, both may be used in different layers.

What are the main barriers to large-scale HCF deployment?

The main barriers are cost, manufacturing scale, standardization, specialized transceiver requirements, testing equipment, splicing transitions, and field training. HCF has strong latency and nonlinearity advantages, but it is still expensive and concentrated in high-value use cases such as hyperscaler DCI and financial networks.

Why could MCF commercialize faster than HCF?

MCF may commercialize faster because it is less disruptive to the existing single-mode fiber ecosystem. Each core can remain optically compatible with familiar G.65x-type systems, while the main changes occur in connectors, FIFO devices, splicing, and test procedures. That makes MCF easier to scale in density-constrained AI data center routes.