Table of Contents
RCR AI TechTalk with Corning’s Brian Rhoney, vice president of data center market development
This year marks Corning’s 175th anniversary and its evolution from creating glass envelopes for Thomas Edison’s light bulbs to solving today’s critical density, space, and power challenges with advanced optical fiber, high-density cable, and specialized connectivity solutions. While chips scale in quarters, physical power scales in years, making density and compute escalating problems for hyperscalers, neoclouds, and all AI infrastructure providers. With average deployed rack density now hitting 27kW, and AI clusters as dense as 100 kW, advanced fiber optics and passive optical connectivity are enabling facility designers and engineers to replace copper and electrical signaling across many facets of modern data centers – inside the server chassis, all the way into the chip substrate. Corning’s innovations in fiber optics and passive optical connectivity have led to monumental supply agreements with Nvidia (including $3.2 billion to expand optical connectivity manufacturing and perhaps ultimately replace copper cables inside rack-scale systems; Meta in a $6 billion deal for fiber through 2030; and Amazon, for U.S. production of optical fiber, cable, and connectivity products in Amazon data centers.
Corning’s three critical layers:
- Between clusters, the ‘Scale-Out’ layer – high-density optical fiber, cables, and connectors that connect GPU clusters and server racks
- Inside the server, rack, ‘Scale-Up’ layer – upgrades capacity inside a single server or rack via miniaturized connectors that jam 3x more optical connections into server racks
- Inside the chip, the ‘Packaging’ layer – glass core substrates and a GlassBridge connector platform use wafer-level ion-exchanged glass waveguides to couple optical signals directly into photonic integrated circuits, helping bypass traditional copper limitations in HPC and AI data centers.
“We were one of the original companies on the S&P 500,” says Corning’s Brian Rhoney, vice president of data center market development. He notes that when it comes to data center glass that connects GPU clusters, “density is paramount, requiring smaller fiber that we then turn into smaller cables and connectivity solutions to interconnect large clusters in support of AI models.”
The fundamental building block for supporting the AI models is the glass fiber, Corning Contour Cable and Flow Ribbon Technology, which Rhoney says maintain signal integrity even under extreme bending inside cramped AI server racks. “Bend-insensitive products mean you can be a little more aggressive with the handling in deployment,” he says, describing how bend-insensitive fiber – originally made for tight outdoor underground spaces – solves the physical space problem inside AI data centers. “Originally developed for carrier fiber-to-the-home networks and data center interconnect environments in which fiber was put into duct space underground, we’re bringing that inside the data center, where cable trays are becoming a limiting factor with AI and the amount of interconnects that are required in the scale-out networks of the last few years and the scale-up networks that will iin the future interconnect and access GPU-to-GPU and GPU-to-memory communications within the data center.” He explains that the RocketRibbon cable with Contour fiber packs up to twice as many fibers into standard 1.25-inch to 2-inch conduits, which translates into a 40% smaller footprint and ability to tolerate extreme bending across tens-of-millions of server connections.
Where traditional ribbon cable has been flat, like a sheet of paper that would be stacked into round cables, the tightly twisted ribbon format is pliable, “to fill out the fullness of the cable itself, giving us 2x more density than traditional cables,which means we can put twice as many cables as we need in those cable tray spaces inside the data center,” says Rhoney.
Another cornerstone of Corning’s data center portfolio is the connector that goes at the end of the cable. “A cable is only good if it has a connector on it. Ultimately, we have to connect to electronics through connection points, which can be a bottleneck of space within the rack. Historically we’d use MPO or MTP-based connectors, and now it’s the MMC connector, which is 3x denser than the traditional multifiber connectors we’ve used for cloud compute in the past decade.”
To pack increasingly more fiber into GPU racks, Rhoney points out that Corning relies heavily on vertical integration. Beyond manufacturing its own core fiber, cables, and connectors—the latter through its US Conec joint venture—the company scales from individual components to complete network solutions. This includes manufacturing the physical metal hardware, such as high-density patch panels, designed to plug directly into active AI clusters from Nvidia and hyperscalers.
“Historically in the data center, we were focused on equipment connections like optical transceivers where the electrical signal inside the equipment [switches or servers] turned into optical signals, but now we go further for improvements to reliability, power, performance, and scale of optics in the data center. Where copper historically was the link, we now go inside the box with optical connectivity all the way to the chip set,” explains Rhoney.
As part of Corning’s new Photonics MAP, Corning extends the equipment-to-equipment connectivity – typically purchased by the end users like the hyperscalers –and extends it to the inside electronics by optically cabling all the way to the chip inside the switches and the NICs. Corning Glass Core Substrates and GlassBridge waveguides allow optical data lines to plug directly into the computing silicon, bypassing the physical constraints of copper electrical traces.
To keep up with growing connectivity demands throughout all facets of data centers, Corning is aggressively expanding through three new advanced manufacturing plants that will expand its U.S. optical connectivity production capacity by 10x and raw fiber volume by over 50%.
“Many of our end user announcements highlight expansion of current facilities as well as new facilities in places like Texas, North Carolina, and other locations because we try to collocate our manufacturing close to where it’s going to be consumed,” says Rhoney. “We are accustomed to growth, as was the case with cloud, so we are in the mindset of steady, sustainable expansions.”
What excites Rhoney about this time in the industry is what’s happening from an application point of view and Corning’s Springboard plan for Scale-Out and Scale-Up:
- Scale-Out, which connects standard node-to-node clusters using traditional networking protocols connecting different server racks across the data center floor;
- Scale-Up, which links GPUs directly together within a cluster so they can behave like one singular, massive super-chip, which is roughly 9x more bandwidth than the scale out.
Moving the “last mile” inside the box—running fiber optics right to the chip package via technologies like co-packaged optics (CPO)—is the engineering vision that represents a new frontier in optics. According to Rhoney, this will bring an order-of-magnitude change in the number of connectors required per GPU. “Between now and end of decade, we are transitioning from copper-fed connections to an optical connection.” He says the “last mile” in the data center will be inside of the equipment, all the way to the chip. With fiber inside the box, he says “designers will use optics in applications they maybe wouldn’t have thought of before.”
