Microsoft and University of Southampton researchers have built a hollow-core fiber that loses just 0.091 dB per kilometer at the standard 1,550 nm wavelength, while propagating light up to 45% faster than conventional glass-core fiber. The achievement, detailed in new technical disclosures and covered by IEEE Spectrum, marks the lowest attenuation ever reported for a hollow-core design and puts the technology a step closer to real-world deployment inside hyperscale data centers.

Light at Record Speed Through Air

A conventional optical fiber guides light through a solid silica glass core. Light slows down as it passes through glass—about 1.44 times slower than in a vacuum—and inevitably scatters off tiny density fluctuations in the material, a phenomenon called Rayleigh scattering. To get around both limits, researchers have spent decades trying to replace the glass core with air.

That is what hollow-core fiber does. Instead of a solid center, light travels through a hollow channel surrounded by an intricate microstructure of thin glass capillaries. The new design, called a double-nested anti-resonant nodeless hollow-core fiber (DNANF), uses two concentric rings of these capillaries arranged around the air core. Each ring acts as an anti-resonant reflector: it bounces the optical signal back into the center, confining more than 99.99% of the energy in air. Because light in air moves nearly at vacuum speed, propagation latency drops by about 45% compared to standard single-mode fiber. And because the light barely touches the glass, Rayleigh scattering and other glass-related loss mechanisms almost vanish.

The research team, led by Francesco Poletti, chief scientist at Microsoft’s Azure Fiber and a professor at the University of Southampton, measured a minimum attenuation of roughly 0.091 dB/km on multi-kilometer spools. That is well below the practical floor of about 0.14 dB/km for the purest silica fiber. The fiber also maintains a loss below 0.2 dB/km across an unusually wide 66 THz spectral window, and within an 18 THz subband the loss stays under 0.1 dB/km. Such broad-spectrum capability could ultimately allow operators to pack far more wavelength channels into a single strand than is possible in the crowded C-band used today.

The Economics of Lower Attenuation and Higher Speed

The headline numbers translate into concrete operational advantages, especially for cloud providers and anyone moving massive datasets.

Fewer amplifiers, lower power. At 0.091 dB/km, a 1,000-kilometer span accumulates about 91 dB of loss, compared to 140–160 dB for conventional fiber. Since amplifiers are needed whenever the signal weakens too much, the reduced loss means fewer amplifier sites, less electrical power, and simpler repeater stations. For long-haul or subsea routes, the savings compound quickly.

Latency that matters. A 45% reduction in propagation delay shaves roughly 1.5 microseconds off every kilometer. That may sound trivial, but for distributed AI training—where thousands of GPUs across multiple data centers must stay synchronized—it can shave milliseconds off each iteration, directly accelerating model convergence. High-frequency trading firms, which already pay premiums for the shortest fiber paths, would notice the difference on any route.

Wider spectrum, more capacity. The 66 THz low-loss window opens the door to transmission bands outside the conventional C- and L-bands. Operators could ultimately multiply per-fiber capacity several-fold, provided the rest of the photonic ecosystem—amplifiers, lasers, transceivers—evolves to support those wavelengths.

Tamed nonlinear effects. Because almost all the optical power stays in air, nonlinear penalties that cap launch power in conventional fiber are greatly reduced. This could allow higher power per channel and denser packing, further boosting throughput on a single fiber.

From Lab Spool to Live Traffic

Microsoft hasn’t waited for academic papers to turn into products. In December 2022, it acquired Lumenisity, a University of Southampton spin-out that had been commercializing hollow-core fiber from its Romsey, UK facility. Since then, the company has integrated the technology into Azure’s networking roadmap.

Public statements and trade reports now describe live pilot deployments. In one disclosed project, Microsoft connected two Azure data centers in Europe using hybrid cables that bundle 32 hollow-core strands alongside 48 conventional fibers, spanning routes over 20 km each. A figure of 1,280 kilometers of hollow-core fiber “deployed and carrying live traffic” is cited by company representatives, though some sources round this to 1,200 km. The discrepancy likely reflects differences in internal reporting, but the larger point stands: Microsoft has enough confidence in the technology to light it with production workloads.

These early installations are not just proof-of-concept. They test real-world splicing, connectorization, and environmental resilience. Hybrid cables—mixing hollow-core and standard fibers—allow operators to trial the new glass without ripping out existing plant. That pragmatic approach mirrors how coherent optics and new fiber types often enter the network: alongside the old, not as a wholesale replacement.

The Manufacturing Gauntlet

Producing hundreds of kilometers of hollow-core fiber with consistent sub-0.1 dB/km loss is a different challenge from making a few pristine spools in a lab. The geometry is unforgiving. Wall thicknesses, tube diameters, and the precise nesting of the anti-resonant rings must be uniform over tens of kilometers. Contamination—moisture, particulates—inside the hollow core can ruin performance. So can microbends introduced during cabling or installation.

Industry observers note that much of the fabrication remains manual. Unlike the highly automated drawing towers that churn out conventional fiber, hollow-core production still depends on skilled technicians. Achieving the yields and throughput needed for mass deployment is the overriding barrier.

Competitors are pushing hard, too. China’s Yangtze Optical Fibre and Cable (YOFC) announced a 21.7 km hollow-core fiber and has cited attenuation as low as 0.05 dB/km in certain tests. Linfiber Technology reported at this year’s Optical Fiber Communications Conference that it has drawn a continuous 47.5 km length at 0.1 dB/km. These vendor claims highlight rapid progress but also underscore the need for independent verification—most figures come from company presentations or PR, not from peer-reviewed studies of production-length reels.

Splicing and connectors remain immature. Joining two hollow-core fibers without introducing large losses or letting air and debris into the core requires specialized equipment and training. Field splicing loss, robustness to bending, and long-term reliability all need to be proven before carriers will deploy HCF outside controlled environments. And the amplifier ecosystem built for the C-band must be reimagined to exploit the wider spectrum that HCF promises.

What IT Architects and Network Planners Should Do Now

Hollow-core fiber isn’t a drop-in replacement for the planet’s five billion kilometers of installed silica fiber. But for specific, high-value use cases, the value proposition is becoming too strong to ignore. Here is where to start.

Identify high-impact corridors. Look for latency-sensitive links where a 45% speed-up could change system design: AI cluster interconnects, synchronous replication paths, financial exchange connections. Also scan for long-haul routes where amplifier count and power costs are significant. On those spans, even moderate per-kilometer loss savings can tip the total cost of ownership.

Run controlled pilots with hybrid cables. Request hybrid trunk cables that include a few hollow-core strands alongside standard fibers. Measure splice losses and attenuation repeatedly—preferably with independent test equipment—and monitor stability over weeks, not hours. Use the pilot to validate vendor claims about loss, latency, and the viability of connectors.

Demand production-grade guarantees. Any vendor offering HCF today should be able to specify attenuation versus reel length with real statistical distributions, not just best-case lab samples. Ask for documented environmental qualification: temperature cycling, humidity exposure, crush resistance, and microbend performance. Until such data is publicly available, treat deployment kilometer figures as provisional.

Model the full system, not just the fiber. The per-fiber cost of HCF is likely to be higher initially. But total cost includes amplifiers, power, real estate, and perhaps even DSP complexity (lower dispersion may ease receiver design). Run the numbers for your specific routes and traffic growth projections. Also stay engaged with standards bodies and industry consortia that are beginning to address connectorization and splicing interfaces—the sooner a common ecosystem forms, the faster costs will fall.

What’s Next

The 0.091 dB/km figure and the 45% latency advantage are real technical milestones, not theoretical projections. They push hollow-core fiber from a laboratory curiosity to a credible candidate for next-generation data center interconnects. Microsoft’s early deployment of more than a thousand kilometers shows strategic intent and a willingness to learn by doing.

But the road to commoditization runs through manufacturing scale-up, ecosystem maturity, and independent field validation. Competitors like YOFC and Linfiber are advancing similar designs, which could accelerate standards and drive down prices—or fragment the market with incompatible approaches. For enterprise architects and service providers, the immediate task is to run limited, measured trials in high-value corridors and to track the development of the surrounding connector, amplifier, and qualification ecosystem.

If those pieces fall into place, hollow-core fiber could become one of the most consequential infrastructure technologies of the next decade—lowering the latency and raising the capacity ceiling just when distributed AI demands exactly that.