AI Driving Fiber to New Highs and Lows
The AI revolution is pushing the fiber ecosystem to dramatic new highs and lows in the quest to pack as much compute power as possible into the smallest practical amount of physical space. Copper wiring is out and glass is in, to reduce physical footprint and power requirements while steadily improving reliability and drastically improving speed as needed.
The sheer size and scale of these projects is mind-boggling. Consider Meta’s Hyperion data center, a four million square foot facility being built in Northeast Louisiana that will deliver over 2 gigawatts of computing capacity when completed. This project will rival the surface area of the island of Manhattan and one of the reasons why Corning and Meta have announced a multi-year $6 billion for fiber.
Let’s first talk about the continually decreasing size of fiber packaging. Initial generation fiber started out at a diameter of 250 microns, including the actual glass and coating, but advances in materials have put fiber on a diet, slimming it down to 190 microns. Combined with new packaging, you can now put double the fiber in the same physical space.
But that’s only the beginning of packing more bit-moving power into existing conduits and new AI data centers. Through multi-source agreements (MSAs), multicore fiber technology is being standardized for large-scale commercial deployment, with four cores being put into a single fiber strand diameter.
Slimmer fiber packaging and multi-core essentially bring the carrying capacity of a single strand of fiber to a factor of eight when compared to first-generation fiber—which, by the way, isn’t going anywhere as it continues to progress to multi-terabit speeds with continued advances in optics and signal processing.
Fiber today provides connectivity at the server, rack, data hall, data center, metro, and long-haul areas, with the medium starting to inch its way from the server down to the chip level for faster speeds, lower latency, and lower power consumption. Most of the fiber densification action today is happening from the server through long haul, as companies move from hyperscaling to neoclouds packed to the ceilings with GPUs for AI model training.
The first layer where fiber is making its presence known in replacing data center copper is out-of-band management (OOBM). PON-based solutions are delivering a triple play of reducing the number of active switches by up to 90%, delivering 50% or more power savings, and simplifying operations by 80%. Eliminating switches means more room for GPUs in the data center.
Plugging in optical gear more efficiently has driven another MSA, this one for XPO, the eXtra-dense Pluggable liquid cooled Optics module. XPO will increase optical switch density by a factor of four through a combination of packaging and liquid cooling, enabling data center planners to pack 64 channels at 200G into a 1U open rack form factor, making still more room for GPUs.
The next challenge in density is connecting data centers together across distance. Pre-AI, a typical amplifier hut between two data centers was designed to support four fiber pairs, or “rails,” per rack, plenty of bandwidth. For these hyper-growth times, vendors have gone to hyper-rail solutions, packing up to 128 rails into a single rack.
Finally, we are also starting to see initial deployments of hollow core fiber. Hollow core is essentially a hollow tube within a fiber where light travels, traveling faster because it isn’t slowed passing through glass. Data center operators like hollow core for lower latency, enabling them to synchronize data sets across longer distances between regional facilities and because it can send fiber at higher signal power, translating to fewer amplifier huts.
All of these advances are happening in parallel to create the ideal conditions for AI advances in the years to come, but I expect many of them to show up across the fiber network ecosystem as prices drop and the need for bandwidth at all levels of society continues to steadily increase. Today’s AI-driven fiber advances will more than likely be next year’s telecom network devices delivering 6G and next-generation household speeds.
Images, Shutterstock
