Relativity raises $22 million, wins $40 million hyperscaler order

As AI systems scale, the distance between GPUs can now span multiple buildings and campuses. A signal that takes roughly five microseconds to travel one kilometer through conventional fiber could take three and a half microseconds through Relativity Networks’ hollow-core line. The Orlando company has now raised $22 million and secured a $40 million follow-on order from an unnamed hyperscaler.

The customer recently tested Relativity’s ChronoCore optical networking technology between two data centers, according to Data Center Dynamics. That test matters because AI computing is no longer confined to one room: operators are distributing GPUs across campuses to reach available power, while trying to make those sites behave like one synchronized machine.

Hollow-core fiber changes the path taken by the light. Instead of sending it through solid glass, the cable guides light through an air-filled center surrounded by glass tubes. Relativity says the result is transmission 30% faster than conventional fiber; an investor quoted by Data Center Dynamics puts the advantage at roughly 47% over solid-core glass fiber. The two reports use different comparisons, but both point to lower latency.

The technology is being made with Prysmian. ChronoCore fiber is produced in Eindhoven, the Netherlands, cabled at Prysmian’s plant in Claremont, North Carolina, and supported with couplers, fiber characterization and installation training in Orlando. The partners have produced a 24-fiber, 10-millimeter cable that was tested in standard microducts, a step toward real-world deployment.

So what changes in practice? Data Center Dynamics says conventional fiber has typically limited data centers to within 60 kilometers (37 miles) of power providers or one another, while hollow-core fiber extends that range to 90 kilometers (56 miles). That gives hyperscalers more room to connect sites where electricity is available. The order confirms commercial interest, but the hyperscaler is unnamed and the sources do not disclose its deployment scale or operating results.