HKT’s hollow-core fiber AI ‘superhighway’ highlights the tech’s promise for ultra low-latency DCI networks, but analyst CRU Group says costs, ecosystems, and production mean at-scale adoption is years away.
In sum – what to know:
High costs – Hollow-core fiber cost runs between $3,000-5,000 per fiber kilometer; parity with standard fiber is unlikely to be achieved in the next 10 years.
Slow production – Yield for hollow-core fiber sits at 10% versus standard fiber, and the surrounding ecosystem still lacks common standards.
Narrow adoption – Only hyperscalers and Chinese carriers are deploying hollow-core fiber, chiefly for AI cluster interconnect and financial exchange purposes.
Earlier this month, HKT announced plans to deploy a 3.2Tbps AI data center interconnect ‘superhighway’, linking Hong Kong’s Lok Ma Chau Loop and the Tseung Kwan O data center cluster using hollow-core fiber technology. Claims of up to 30 percent lower latency than standard fiber make hollow core an attractive proposition.
But the move by HKT, formerly Hong Kong Telecom, should not be considered a sign that hollow-core fiber is ready to be deployed at scale – according to Ahmed Ali, lead for optical fiber and cable at London-headquartered commodity research and analyst firm CRU Group.
Why the economics don’t work (yet)
Hollow-core fiber routes light through an air or gas-filled core, reducing latency and attenuation by replacing the medium light travels through rather than the distance covered. That provides a distinct architectural advantage for hyperscalers, as it allows them to site data centers further apart, balancing proximity to users against proximity to power, while maintaining an ultra-low-latency and high capacity link.
The biggest problem according to Ali, is that the technology is currently hard to scale. “The manufacturing yield is about 10 percent, which is very low compared to standard fiber. The other issue is that the ecosystem itself – the connectors, testing and splicing equipment – is not developed yet,” he says, with every supplier’s kit currently bespoke to its own product. That combination keeps cost at roughly $3,000-5,000 per fiber kilometer, a figure that is unlikely to come down anytime soon.
“We’re not anticipating the cost to come down to the level of standard fiber, at least not in the next ten years,” Ali said. “It will remain a premium solution.”
Latency at any cost
That price tag confines real deployment to a narrow band of buyers willing to pay for ultra-low latency, such as hyperscalers running AI training clusters, and increasingly, a smaller niche group looking to connect financial stock exchanges. Publicly disclosed hollow-core builds tend to run between 10-40 kilometers, occasionally stretching toward 200 kilometers, although Ali noted the longer distances are typically trials rather than commercial deployments, since manufacturing continuous long-haul hollow-core fiber remains technically difficult.
Geographically, the US leads on the strength of its data center buildout alone, with the three incumbent Chinese telecoms operators also experimenting with hollow-core deployments. Back in June, China Telecom, YOFC, and Dekoli claimed to have delivered 51.3 Tbps speeds over a distance of around 206.5 kilometers (128 miles) without signal regeneration. Hollow-core also isn’t the only next-generation fiber technology in play.
Several adjacent technologies are also being explored to improve efficiency and bring the benefits of low latency to different use cases. Multi-core fiber (multiple light-carrying cores in one strand, adding capacity rather than cutting latency) and co-packaged optics (pushing optical connectivity closer to the GPU), for instance, are parallel efforts aimed at delivering greater capacity and throughput for a greater number of use cases, from data centers to subsea cables and 6G connectivity.
Future outlook
Hollow-core’s value proposition is currently limited to scenarios where the latency gain clearly justifies the “bolt-on” cost. But rather than replacing copper or traditional fiber with hollow-core, Ali believes that when the current AI infrastructure buildout cycle matures, we could see hollow-core being integrated into data center network design from the outset. That, he believes, could be likely sometime after 2030, if manufacturing costs fall and standards mature.
For now, HKT’s build sits at the leading edge of a niche rather than the start of a mainstream shift. While the technology clearly works, until the ecosystem behind it catches up, it will remain beyond the reach of all but a handful of hyperscalers.
