SILICON VALLEY — Inside the Amazon Web Services networking lab in Mountain View, California, on the 9th, rows of server racks draped in yellow optical cable stretched across the research floor. The cables ran through transceivers, which convert signals between electrical and optical form, linking trays about 50 centimeters wide within each rack and connecting racks standing side by side. On a table next to the servers, equipment measured how much transmission loss the cable prevented over distances of more than a kilometer. It was the first time AWS, the world's largest hyperscaler, or operator of large-scale computing infrastructure, had opened the lab to Korean media.
Optical fiber sits at the center of that network. AWS researchers showed reporters hollow core fiber, or HCF, billed as the next generation of the technology. It is an opaque strand as thick as a human hair, measured in micrometers, or millionths of a meter, that runs inside the optical cables carrying data between servers in a data center. From the outside it looks no different from conventional fiber. Inside it is another thing entirely. Where standard fiber is filled with glass or another solid, the center of an HCF strand is empty. Light travels quickly through that core, confined by a cladding of high-purity silica. The fiber measures 125 micrometers across, with a core diameter of just 8 to 9 micrometers.
Filled with air, HCF is the key technology for tying together distant data centers while cutting network latency. The idea rests on the lower refractive index of air compared with a solid such as glass, which lets light pass through faster. Switching the transmission medium from glass to air lifts speeds from about 200,000 kilometers per second to roughly 300,000, and sharply increases bandwidth, or the volume of data carried per second. That 300,000 figure is the speed at which light travels in a vacuum. Conventional fiber carries a transmission latency of 5 microseconds, or millionths of a second, per kilometer; HCF runs at 3.3 to 3.5. In practical terms, HCF extends reach by 50% over standard fiber and improves latency by 30%. Given that the continental United States spans 1,700 to 1,800 kilometers from east to west, shaving 1 to 2 microseconds per kilometer counts as a substantial gain.
The AI industry tends to name graphics processing units and high-bandwidth memory as its bottlenecks, but AWS argues the network matters as much as the silicon. "Take Project Rainier," said Stephen Callaghan, principal network development engineer at AWS, referring to the company's massive AI infrastructure effort. "There are places handling more than a million chips. A problem in any single part of the network stops every GPU." Callaghan added: "The GPUs have to move through every task in sequence, like an army marching in lockstep. Our principle has been to keep the network running at its best so customers can think about nothing but the GPUs."
HCF is drawing attention as a way past the physical distance limits of data centers. Constraints on the sheer volume of AI chips, on the electricity needed to run and cool the facilities and on industrial water supply are pushing individual sites to grow larger and to sit farther apart. AWS has broken ground on a new subsea cable linking the western United States and Japan, targeted for service in 2029 — a measure of how much securing low-latency, high-bandwidth networks across physical constraints has become the central challenge.
AWS operates 123 data center clusters across 39 regions worldwide. The optical fiber cable it has laid undersea, underground and overland to connect those sites, and to wire each of them internally, runs to 20 million kilometers. That is enough to circle the Earth, at 40,000 kilometers around, 500 times, or to make 25 round trips to the moon. Working with several manufacturers, AWS built its own HCF procurement system and began deploying the fiber in its data centers in 2024. It now links more than 10 of them. "The scale of AI is becoming enormous, and everything runs through the data center," said Matt Rehder, vice president of network engineering at AWS. "Our goal is to use HCF to extend reach without adding latency."
Copper, which carries electrical signals, was once the main material in the network cables connecting data centers. As the race to develop AI has intensified, optical fiber, which moves data as light, has taken its place in the spotlight. AI chips that handle training, such as GPUs, and the memory that speeds up their calculations have long been cast as the bottleneck; more recently, network speed and reliability have emerged as another one. The realization, in short, is that however many AI chips and memory modules a data center packs in, none of it matters if the network linking the servers breaks down. Data Bridge Market Research, a global market research firm, projects the worldwide optical fiber market will surge from $7.36 billion in 2024 to $15.45 billion in 2032.
AWS rivals have entered the same race. Microsoft began deploying HCF in Azure, its data center cloud, in 2023, and last September said it would work with Corning, the U.S. fiber manufacturer, and Heraeus Covantics, a silica materials company, to expand HCF production. Microsoft acquired Lumenisity, a spinout from the Optoelectronics Research Centre at the University of Southampton in Britain, in 2022, then built what it called the world's first advanced HCF manufacturing plant in the U.K. In response, AWS said on the 8th that it would partner with Qualcomm, the mobile chip company, on 1.6-terabit-per-second optical communications technology to raise data transfer speeds inside data centers.
Nvidia, the world's largest AI chip company, has concluded that securing enough optical fiber is essential to running its GPUs reliably, and signed an advanced fiber supply agreement with Corning in May. Corning agreed to increase its U.S. optical connector production capacity tenfold and expand fiber output by 50%. Nvidia secured the right to invest as much as $3.2 billion in the company. In March, Nvidia had invested $2 billion each in Lumentum Holdings, the U.S. optics and laser component maker, and Coherent, an optical transceiver manufacturer, signing multiyear non-exclusive purchase agreements with both.
Nvidia Chief Executive Jensen Huang has taken a personal hand in locking down optical equipment, attending the groundbreaking in June for an expansion of Coherent's plant in Sherman, Texas. Optical transceivers convert electrical signals into light and back. Having concluded that AI chips and memory alone — GPUs and high-bandwidth memory — can carry performance only so far, Nvidia is cutting copper wiring and adding optical transceivers in next-generation accelerators such as Vera Rubin.
For all its gains in reducing signal loss and raising speed, HCF has limits to overcome: complicated manufacturing and cost. Conventional fiber optic cable runs a few dollars per meter, while HCF costs about $600. Because the strand is hollow, performance degrades when it is bent or struck. Some also argue the bandwidth gains will not justify the expense.
