How Hollow-Core Fiber Could Let AI Data Centers Scale Across More Sites

AI data-center design is becoming a geographic problem as well as a computing problem. When power, land, grid access, permitting, or local politics prevent operators from building one enormous campus, they may need to distribute GPU capacity across multiple sites. Relativity Networks is betting that a faster optical medium can make those sites function more like a single, tightly coupled system.

Its ChronoCore platform uses hollow-core fiber, which guides light through an air-filled center rather than conventional solid glass. Relativity says the approach can reduce propagation delay to roughly 3.4–3.5 microseconds per kilometer, compared with about 5 microseconds for standard fiber. 1618

The core idea: make distance less restrictive

Fiber does not eliminate the physical distance between data centers. It can, however, reduce the time an optical signal spends traveling between them. That distinction matters when a distributed AI workload repeatedly exchanges data between GPU pools.

Relativity’s stated use case is inter-data-center connectivity: its cable is designed to provide 24 low-latency pathways between AI facilities, while the company says hollow-core fiber can carry data roughly 1.5 times farther at a comparable latency. 1424

In practical terms, the technology is intended to give infrastructure planners more options. Instead of treating a single contiguous campus as the only viable home for a tightly connected AI system, an operator could potentially combine capacity from several sites where electricity and development conditions are more favorable.

Why AI makes the latency trade-off more important

A small amount of additional propagation delay may be unimportant for ordinary connectivity. It becomes more consequential when a workload is spread across facilities and must coordinate frequently. Relativity’s pitch is that lowering the fiber’s propagation delay can help distributed training and inference remain synchronized across a larger physical footprint. 110

The company describes this as a shift in which “geography becomes part of the computer”: the network is no longer merely connecting separate facilities, but helping determine how far apart those facilities can be while still behaving like one coordinated AI factory. 1718

That does not mean faster fiber alone turns multiple data centers into one cluster. End-to-end performance also depends on switches, transceivers, network topology, software, workload behavior, and every other source of latency in the path. Hollow-core fiber addresses one layer of the problem—the propagation time of the optical link.

What the 50% claim really means

Coverage of the announcement has described hollow-core fiber as roughly 50% faster than conventional fiber. 2 Relativity’s own materials put the figure closer to 47%, with ChronoCore at approximately 3.4 microseconds per kilometer and about 32% lower latency over the same distance. 18

Those figures should not be treated as interchangeable. A reduction from 5 to 3.5 microseconds per kilometer is a 30% reduction in one-way propagation delay. Expressed as a speed increase, it is about 43% under that simplified comparison. The exact percentage depends on which baseline and measurement the comparison uses.

The more durable point is not the headline percentage. It is that replacing a glass propagation medium with an air-core design could recover enough time to extend the practical distance between facilities.

Why the financing and hyperscaler order matter

Relativity announced $22 million in SAFE-note funding and a separate $40 million follow-on order from an unnamed hyperscaler. 23 A SAFE is an agreement that generally converts into equity in a future priced financing, so the funding is better understood as early-stage capital than as a conventional priced equity round. 2

The order provides a different kind of signal. Reporting says the hyperscaler tested ChronoCore for inter-data-center connectivity before placing the follow-on order, suggesting that at least one large cloud operator moved beyond an initial evaluation. 113

That is meaningful, but it is not proof of broad adoption. The customer has not been named, and the public reporting does not provide the contract’s deployment schedule, technical test results, or confirmation of how much infrastructure the order will ultimately support. The $40 million figure is therefore evidence of commercial interest, not a guarantee that hollow-core fiber will become a standard replacement for conventional fiber.

The infrastructure shift behind the product

Relativity is targeting a constraint that sits outside the GPU itself: where enough power and suitable development capacity can be found. Company materials and coverage describe operators spreading AI infrastructure across geographically separated campuses as they look for available electricity and workable sites. 1627

That changes the role of networking. In a traditional data-center buildout, the network primarily connects machines inside a defined facility. In a distributed AI buildout, the interconnect can influence whether separate facilities are close enough to participate in the same system.

Hollow-core fiber is consequently positioned as a geographic-enablement layer. It may allow operators to aggregate compute from multiple power-feasible locations rather than waiting for one site with sufficient electricity, land, grid access, and political support. Optics.org reports Relativity’s comparison of a typical roughly 60-kilometer constraint with a potential 90-kilometer reach for its fiber, although the usable distance for any specific AI workload will depend on the full system design. 24

What Relativity’s announcement reveals—and what it does not

The announcement points to a broader industry direction: as AI systems grow, infrastructure companies are optimizing not only compute density and internal networking, but also the physical geography of the overall system.

Relativity’s funding gives it capital to scale production, while the follow-on order indicates that its hollow-core approach has attracted at least one serious hyperscaler evaluation. 34 The technical proposition is straightforward: lower propagation delay could make geographically distributed GPU capacity more viable.

The unresolved question is operational. The industry still needs public evidence about manufacturing scale, cost, reliability, deployment complexity, and measured end-to-end performance on real distributed AI workloads. Until those details emerge, ChronoCore is best understood as an important infrastructure bet on a changing constraint—not as proof that fiber technology has solved the AI data-center geography problem.