AI clusters are outgrowing single buildings. As GPU counts climb into the hundreds of thousands, power and space limits are pushing operators from single-campus "scale-out" designs toward "scale-across" architectures that spread one training job over several metro or regional data centers. The moment a job crosses that distance, one physical property starts to matter more than almost anything else in the optical layer: how fast light travels through the fiber. Hollow-core fiber (HCF) is drawing attention in AI networking precisely because it changes that number.
This article explains what hollow-core fiber is, how it lowers latency, where it fits in an AI data center, how it compares with single-mode fiber, and what still limits its deployment. The goal is a realistic view for network planners and buyers, not a claim that HCF replaces conventional fiber. For the wider context of why AI infrastructure depends on high-performance optical fiber, that discussion is a useful companion to this one.
What Is Hollow-Core Fiber (HCF)?
Conventional optical fiber guides light through a solid glass core using total internal reflection at the core-cladding boundary. Hollow-core fiber does something different: the center of the fiber is air (or near-vacuum), and a precisely engineered glass microstructure around it keeps the light confined to that air channel instead of the glass.
Two cladding designs dominate the discussion. Photonic-bandgap fiber uses a periodic lattice of tiny air holes that blocks certain wavelengths from leaking into the cladding, effectively trapping light in the hollow core. Anti-resonant (nested antiresonant, or "nested tube") fiber uses thin curved glass membranes whose anti-resonance reflects light back into the core. The anti-resonant family, including the double-nested antiresonant design Microsoft uses, is the one now moving fastest toward commercial production. For a deeper primer on the structure, see this overview of how hollow-core fiber breaks free from solid glass.
How Hollow-Core Fiber Reduces Latency Compared with Conventional Fiber
Light travels through glass at roughly two-thirds of its speed in a vacuum, because glass has a refractive index near 1.5. Air is close to 1.0, so light in an air core travels faster. That single fact is the whole basis for HCF's latency advantage.
The headline numbers come from the largest deployer of the technology. Microsoft reports that its hollow-core fiber is about 47% faster than standard silica fiber, with correspondingly lower latency, and Nokia describes HCF as delivering roughly 30% less propagation delay than today's single-mode fiber. Both framings point at the same physics.
Two cautions matter here. First, this is propagation latency, the time light spends in the fiber. End-to-end latency also includes transceivers, switches, routing, and protocol overhead, so a 30% reduction in fiber delay does not translate into a 30% reduction in application latency. Second, there is no such thing as "zero latency" on any physical link. HCF lowers propagation delay; it does not remove it.
Why AI Data Centers Need Lower-Latency Optical Interconnects
In distributed AI training, thousands of GPUs repeatedly exchange gradients and parameters in tightly synchronized collective-communication steps. The slowest link in that exchange can stall the whole job, so east-west and data-center-interconnect (DCI) latency directly affects how efficiently expensive accelerators are used. As Nokia notes in its work on scale-across networking for AI, the distance between data centers in a distributed cluster is effectively capped by latency, and fiber propagation delay becomes the dominant component as those distances grow. Lowering delay per kilometer is what lets a cluster spread farther without breaking its latency budget.
Key Use Cases of HCF in AI Data Centers
HCF is not a single-purpose product. Its value changes depending on where in the network it sits.
- Metro and regional DCI (the clearest fit today). Interconnecting AI data centers a few kilometers to tens of kilometers apart is where lower propagation delay pays off most, because the fiber is long enough for the per-kilometer saving to add up. This is the primary target of current deployments.
- Cross-data-center AI clustering (scale-across). By cutting delay per kilometer, HCF lets operators place cooperating clusters farther apart while staying inside a latency budget, giving more freedom in siting facilities around available power and land.
- GPU cluster and backend fabric interconnect. Inside the training fabric, latency and bandwidth both matter. HCF can contribute here, though much of this layer is served today by short-reach parallel optics and high-density structured cabling such as 800G MPO cabling for AI data centers.
- Storage-to-compute links. Frequent high-volume transfers between storage and compute nodes benefit from high capacity, one area where HCF's low loss and wide usable spectrum are promising, though intra-building distances are short enough that latency gains are modest.
-
Hollow-Core Fiber vs Single-Mode Fiber: Key Differences
The honest comparison is not "HCF replaces single-mode fiber." Single-mode fiber (SMF) remains the mainstream medium across data centers and long-haul networks, and it is mature, low-cost, and supported by a complete ecosystem. HCF is a specialized option for latency- and capacity-sensitive links. The table summarizes the trade-offs at a high level.
| Attribute | Hollow-core fiber (HCF) | Conventional single-mode fiber (SMF) |
|---|---|---|
| Light path | Air/near-vacuum core | Solid glass core |
| Propagation latency | Lower (roughly 30% less delay per km, per Microsoft and Nokia) | Baseline |
| Reach at equal latency | Extended; vendors cite roughly 1.4–1.5x | Baseline |
| Attenuation | Recent research below the silica limit; still maturing in production | Well-characterized, very stable |
| Ecosystem (connectors, splicing, amplifiers) | Emerging | Fully mature |
| Relative cost | Cost premium today; falling as production scales | Low, commodity |
| Best-fit role | Latency-sensitive metro/regional DCI and scale-across links | General-purpose, the default choice |
If you want the fundamentals of the incumbent medium before weighing HCF against it, our single-mode fiber overview covers the standard G.652/G.657 fibers most networks run on today.
Commercialization Status: From Pilot Links to Early Deployment
HCF has moved from laboratory curiosity to early field deployment, mainly driven by hyperscalers. Microsoft acquired the University of Southampton spin-out Lumenisity in 2022 and has been deploying HCF across Azure regions, including metro data-center interconnect links. On the performance side, researchers from Microsoft and Southampton reported a record-low loss of about 0.091 dB/km, published in Nature Photonics and below the long-standing silica limit, in a pilot that reportedly involved more than 1,200 km of fiber carrying live traffic. Industry trials by other operators and vendors have shown similar latency reductions on metro routes.
What this adds up to is a technology in early, targeted deployment rather than mass adoption. The signal-loss and latency results are real and independently reported; broad rollout across the industry is still ahead.
Limitations and Deployment Considerations
HCF's advantages come with practical constraints that a network planner has to weigh honestly:
- Cost premium. HCF still costs more than commodity SMF, though the gap is narrowing as manufacturing scales.
- Ecosystem maturity. Connectors, splicing techniques, test tooling, and amplification for HCF are still developing, and skilled technicians are fewer. Reported record losses do not automatically generalize to every production route.
- Standardization. Specifications still vary between suppliers, so interoperability and qualification need care.
- Link selection. The latency benefit scales with distance, so HCF earns its cost on longer metro and regional links, not on short intra-rack hops where propagation delay is already tiny.
- Energy. Lower loss can mean fewer repeaters and amplifiers on a route, which can reduce energy per bit, but the size of that saving depends on the specific link and should be evaluated, not assumed.
How Network Planners Should Evaluate HCF
A practical way to decide whether HCF belongs on a given link is to work through a short checklist: How long is the link, and is propagation delay actually a meaningful share of its latency budget? Is the route latency-critical, such as synchronized training traffic between sites? Do the transceivers, amplifiers, and connectors you plan to use support HCF? Does the cost premium buy enough reach or latency headroom to be worth it versus adding SMF capacity or upgrading coherent optics? For most links, established SMF paired with high-capacity coherent pluggables and well-planned data center connectivity solutions remains the pragmatic answer; HCF is the tool you reach for when the latency budget is the binding constraint.
Frequently Asked Questions
Is hollow-core fiber better than single-mode fiber?
Not universally. HCF offers lower propagation latency and extended reach at a given latency, which is valuable on specific links. SMF is cheaper, mature, and better for the vast majority of general-purpose connections. "Better" depends entirely on the link and its latency budget.
How much latency can HCF actually save?
Vendors report roughly 30% lower propagation delay per kilometer, and Microsoft cites about 47% faster transmission than standard silica fiber. That applies to time spent in the fiber; end-to-end improvement is smaller because switches, transceivers, and protocol overhead are unchanged.
Where does HCF fit in an AI data center?
The clearest fit today is latency-sensitive metro and regional data-center interconnect, including scale-across links between distributed AI clusters. It can also play a role in GPU-fabric and storage-to-compute connectivity, though short in-building distances limit the latency benefit there.
Is hollow-core fiber ready for large-scale commercial deployment?
It is in early, targeted deployment, led mainly by hyperscalers on selected routes. Manufacturing scale, connector and splicing ecosystems, and standardization are still maturing, so broad industry-wide adoption is not yet here.
Does HCF reduce energy consumption?
It can, mainly by lowering signal loss so fewer repeaters and amplifiers are needed on a route. The actual saving depends on the specific link design and should be measured rather than assumed.
Summary
Hollow-core fiber is not simply a faster cable. Its real value in AI networking is reducing propagation delay on long optical links, extending reach within a fixed latency budget, and giving operators more freedom to distribute AI clusters across sites. Those benefits are strongest on metro and regional DCI, weakest on short in-building hops, and always weighed against a cost premium and a still-maturing ecosystem. For the foreseeable future, HCF and single-mode fiber are complementary: SMF carries the bulk of the traffic, while HCF targets the latency-critical routes where every kilometer of delay counts.
As AI clusters keep scaling across sites, choosing the right fiber and interconnect for each link is becoming a core design decision. Hengtong works with operators on fiber optic cabling for data centers and can help match the medium to the latency, capacity, and reach requirements of each part of the network.