Hollow-Core Fiber: Below the Rayleigh Floor

Scope

This article explains why the loss of conventional optical fiber has been stuck for four decades, how hollow-core fiber finally moved past it, and what else changes when the light travels in air: latency, nonlinearity, dispersion, and the link budget. It closes with the practical difficulties that an air core introduces and the deployment state as of late 2026. Background: the hollow-core fiber entry for the device, Rayleigh scattering and propagation loss for the mechanisms.

The floor under solid fiber

The first low-loss fiber, in 1970, attenuated about 20 dB/km. Twenty years of materials work brought that below 0.2 dB/km, and then progress effectively stopped: the best solid-core silica ever made sits near 0.14 dB/km at 1550 nm, and the gap between a 1990s production fiber and the all-time record is a fraction of a dB.

The stall is not engineering fatigue; it is a floor built into the material. Silica glass is a frozen liquid, and the density fluctuations that were present at the glass transition are frozen in with it. They scatter light by the Rayleigh mechanism, with the familiar $\lambda^{-4}$ dependence, so scattering loss falls as the operating wavelength moves toward the infrared; but silica's phonon absorption rises steeply beyond 1600 nm, and the sum of the two mechanisms has its minimum near 1550 nm at roughly the value the record fibers achieve. Purity solved the absorption problems of the 1970s; nothing solves the thermodynamics of the glass itself. As long as the light travels in silica, 0.14 dB/km is approximately the end of the road.

The conclusion, obvious in retrospect and pursued for twenty-five years, is to take the glass out of the path.

Guiding light in air

The first serious attempt, the photonic-bandgap fiber of the 2000s, surrounded an air core with a periodic cladding lattice and reached about 1 dB/km before surface modes at the core boundary stopped further progress. The design that broke through is different: antiresonant guidance, in which a ring of thin nested glass tubes surrounds the core and each membrane acts as an antiresonant reflector. No bandgap, no periodicity; the membranes are simply the wrong thickness for light to enter them, at every wavelength except narrow resonances. In nested designs of the Southampton lineage (NANF and its double-nested successor DNANF), the overlap between the guided field and glass is parts in ten thousand, and the loss mechanisms of solid fiber are suppressed by roughly that factor.

The result arrived in 2025: a DNANF with 0.091 dB/km attenuation, published in Nature Photonics by the Microsoft and Southampton team, below the best solid-core fiber ever made and with a low-loss window far broader than silica's. The historical objection to hollow-core fiber, that its elegance came with a loss penalty, is gone; the record now belongs to the air core.

The latency arithmetic

Light in glass travels at $c/n_g$ with a group index near 1.468, which is 4.90 µs of one-way delay per kilometer. In an air core the group index is within a few parts in ten thousand of 1, and the delay is 3.34 µs/km. The hollow path is 47% faster, or equivalently saves 32% of the time: 1.56 µs per kilometer, 312 µs of round trip per 100 km of route.

Who pays for microseconds is a short list with deep pockets. Financial traders paid for it first, and hollow-core links have carried exchange traffic for years over short routes where every microsecond of round trip is priced. The new buyer is the AI datacenter operator: synchronous training traffic between buildings and between metro regions turns propagation delay into idle accelerator time, and a third of the light-speed penalty is recoverable by changing the cable. This, together with the loss result, is why Microsoft has deployed thousands of kilometers in Azure and announced in 2026 that manufacturing moves to outside partners at scale.

Nonlinearity nearly vanishes

The Kerr nonlinear index of air is roughly three orders of magnitude below silica's, and with the field barely touching glass the fiber's effective nonlinear coefficient falls by a similar factor. Stimulated Brillouin and Raman scattering, the processes that cap launch power in solid fiber, weaken to near irrelevance. Two consequences follow. Launch power ceilings rise, which converts directly into link budget or into more WDM channels at the same amplifier count. And the nonlinear penalties that coherent transmission spends DSP effort managing shrink toward the linear regime, which simplifies the accounting for long links. High-power beam delivery, an application solid fiber handles badly, comes along as a side effect.

Dispersion

An air core has almost no material dispersion, and antiresonant designs put chromatic dispersion at a few ps/nm/km near 1550 nm, several times below the 17 ps/nm/km of standard G.652 fiber; the chromatic dispersion entry covers what that parameter does to a link. For short-reach IMDD systems this multiplies the dispersion-limited reach; the Datacenter Link Budget Explorer makes the UI-spread arithmetic concrete for conventional fiber, and a hollow-core channel starts that calculation several times ahead.

The span budget, worked

Take an 80 km span. At a deployed-grade solid-fiber figure of 0.18 dB/km (cabled, with margin) the fiber loss is 14.4 dB; at the hollow-core record's 0.09 dB/km it is 7.2 dB. Against a 20 dB budget the loss-limited reach doubles, from about 110 km to about 220 km, before amplification is needed. Field numbers will land above the record laboratory value once cabling, splices, and repair margin are included, and honest planning uses those cabled figures; the structural point survives the derating, because halving the per-kilometer loss halves the amplifier count of a long route or removes in-line amplification entirely from metro spans that needed it.

What an air core costs

The difficulties are real and mostly at the ends and joints. The mode of a hollow-core fiber does not match standard single-mode fiber, so interconnection needs mode-field adapters or engineered transitions, each worth some tenths of a dB and each a potential source of the paired reflections that become multipath interference. Termination and connectorization remain specialist work, which is why deployments favor factory-built cable assemblies. Monitoring changes character: an air core generates almost no Rayleigh backscatter, so the OTDR trace that solid-fiber operations lean on is faint and reads differently, and backscatter-based sensing along the route is largely unavailable. Contamination control at open ends matters more than glass practice suggests. And manufacturing volume is years behind solid fiber, which is what the 2026 move to outsourced production is meant to fix; until volume arrives, cost per kilometer rations the technology to routes where microseconds or decibels are priced highest.

These are engineering costs of the kind deployment programs retire steadily; the loss penalty, the objection that blocked the technology for two decades, is not among them anymore.

Where it stands, late 2026

Hollow-core fiber has crossed from research program to infrastructure: record loss below solid core published and reproduced across a broad band, thousands of route-kilometers carrying Azure production traffic, manufacturing scaling through outside partners, and finance links as the long-standing commercial precedent. The open questions are volume economics, long-haul and subsea qualification, and how quickly the interconnect and monitoring toolchain matures around an air core. The physics question that held the field for twenty-five years is answered.

References: Petrovich, Poletti et al., DNANF attenuation below 0.1 dB/km, Nature Photonics (2025); Poletti, "Nested antiresonant nodeless hollow core fiber," Opt. Express 22, 23807 (2014); Kapron, Keck & Maurer, Appl. Phys. Lett. 17, 423 (1970); Microsoft Azure networking blog on hollow-core deployment and manufacturing scale-up (2025-2026). The hollow-core fiber entry summarizes the device; the Datacenter Link Budget Explorer carries the span arithmetic.