Invest1 publisher3 min readPublished
AWS swaps glass for air to stretch its data-center links 50% further
Air at the core of an optical fiber carries light faster than glass does, and AWS is using that to put more distance between the buildings that share a single AI job. The gain it reports is measured in kilometers.
The Investor · Invest desk
What happened
- AWS showed Korean reporters hollow core fiber at its Mountain View networking lab, a fiber that holds air at its center inside a ring of high-purity silica instead of being filled with solid glass.
- Because light bends less in air than in glass, the switch of medium raises transmission speed 50%, to 300,000 kilometers per second from 200,000, and increases the volume of data carried per second.
- AWS has been deploying the fiber in its data centers since 2024 and is now using it to link more than ten separate data centers to each other.
- AWS says its optical transponders for dense wavelength division multiplexing carry 73% more bandwidth and use 35% less power than its earlier products.
- Project Rainier, AWS's ultra-large AI infrastructure effort, includes sites handling more than one million chips, according to AWS network development engineer Stephen Callaghan.
Compiled by The InvestorSomething wrong?How this is made
Why it matters
- constraint With power, land and water limits and local opposition blocking single large sites, the operative question for a new campus becomes how far apart the buildings can sit and still hold one job together.
- capability A fault anywhere in the network idles every GPU on a job, so lower latency per kilometer is what makes a training run across several buildings behave like a run inside one.
- contradiction AWS says its rivals are adopting hollow core too, which undercuts reading the fiber itself as an advantage AWS holds alone.
- precedent If the network is as much of a bottleneck as GPUs and high-bandwidth memory, buyers of AI capacity have a second spec sheet to ask for, and hyperscalers have a reason to publish latency per kilometer.
The number that decides where a building can go is latency per kilometer. AWS puts conventional fiber at 5 microseconds a kilometer and hollow core at 3.3 to 3.5 [3]. Over 100 kilometers that is 500 microseconds one way against 330 to 350, a saving of 150 to 170 in each direction [19]. Run it backwards and the reach number falls out: hold the latency budget fixed, divide 5 by 3.33, and the same delay covers about 1.5 times the distance, which is the 50% extension AWS quotes [20][4]. Its summary figure of 30% better latency is the conservative end of its own range, since 3.5 against 5 is 30% and 3.3 against 5 is 34% [18].
This is a siting argument before it is a performance one. Power, land and water constraints and local opposition are making one large facility on one plot harder to build, and AWS's case is that fiber between separate buildings substitutes for the plot it cannot get [9]. The company runs 123 data center groups across 39 cloud regions serving 245 countries and territories [5], and the fiber it has laid undersea, underground and above ground comes to 20 million kilometers, or roughly 163,000 per group [6][22]. The material does not break out how much of that is hollow core [25]. "Our goal is to extend reach without latency using hollow core fiber," said Matt Leder, senior vice president of network engineering at AWS [15].
A few hundred microseconds matter because of what runs across them. Callaghan described the dependency inside those buildings. "In those places, a problem in any single part of the network stops every GPU," he said [14]. "Our principle has been to keep the network running in optimal condition so that customers can focus only on their GPUs," he added [24].
The exclusivity is claimed one layer up from the glass. Satish Vangala, AWS's director of network product engineering, said AWS is the only company able to use DWDM optical transponders on long-haul networks [12]. That was said by an AWS executive in an AWS lab, on the first occasion the company opened that lab to Korean media, and the material carries no independent measurement of it [17].
My read is narrow: the fiber changes siting economics before it changes any benchmark, because what it relaxes is the distance a synchronous training job will tolerate between separate buildings. The counter-thesis may well be right. If hollow core is available to any buyer who wants to order it, every hyperscaler eventually gets the same radius, and the difference falls back onto the transponder layer, where the claim rests on Vangala's word. Data Bridge Market Research puts the global optical fiber market at $7.36bn in 2024 and forecasts $15.45bn by 2032 [16], which is 2.1 times in eight years, about 9.7% compounded [21].
What to watch
- A second buyer or a fiber maker publishing its own per-kilometer latency for hollow core, or a price per kilometer.
- A hollow-core kilometer count from AWS, or the names of the sites it has linked beyond the current ten.
- Any rival deploying DWDM optical transponders on a long-haul network, which would test Vangala's exclusivity claim.