How Fiber Innovation Is Scaling AI Data Centers

Badri, I know that speed of deployment often competes with reliability, it's kind of a balancing act. So how do you help customers move faster without compromising performance and things like long-term scalability?

Two things. Preterm assemblies, that is fully, even for the very highest fiber counts, 6,912 fiber counts or 3,456, pre-terminated assemblies are the high-quality, fast-moving solution. Splicing is the next best thing. But both of these require a lot of, let's say, support from a company like us, technical support, engineering support. Basically, we offer that to all the installers, the installer community, as well as the end customer.

Tell me a little bit more about those fiber solutions that you just mentioned. What is special about them?

First thing is the fiber solution itself is, there are two parts to it. One is the fiber and the cable, which traditionally you're familiar with. What wasn't really familiar in the past was the pre-termination, putting the connectors, the end pieces where the customer just plugs the terminator fiber into their patch panels, into their data center racks and so on. So that's become the working horse style for selling connectivity, the data center. Even if it's a telco build or a data center AI build, you've got to have a pre-terminated connector. And then the smallest form factor, the better. So right now, we are MMC-focused. MMC is one of the smallest form factors available, and we are certified by US Connect, who is a licensing partner for this. It gets you 3X the capacity, which means the speed in the same amount of time, and they get significantly more capacity without getting into more physical space or larger real estate builds.

I know there's something called hollow-core fiber that's super cool and multi-core fiber. Both of those are getting a lot of attention as AI networks scale for various reasons. Where do you see each one of those, hollow-core and multi-core fibers, fitting into the future of data centers? What would need to be true for them to move from labs and trials into mainstream deployments?

Multi-core is the same fiber we've known before, but just more of the same inside the same size. Let's say you've got a standard single more fiber, it's 125 microns. Now you've got four transmission channels in the maybe even seven or whatever. Why that's important is everything stays the same in terms of transmission that people are used to the last 30 years in terms of the transmission standards, the ITU, the IEC, all of the ratifications that went on, that gets preserved. What gets changed is that you're able to engineer a product that now has 4X the capacities in space shield capacity. So you can think of it like this. If there's a cable that has, let's say, 1,728 fibers in it, you put four cores into each fiber, now you have 6,912 transmission pathways inside that. And then where it gets interesting is, you are utilizing the duct much more efficiently, almost four times more efficiently.

And if you look at the way that the builds are coming up, a lot of the cable is actually not glass, it's mostly plastic. And so, you've got the space-efficient and cost-efficient solution and you utilize the same glass preform four times in one factory run almost. So you get this cascade of savings all the way from the draw tower when you make the glass all the way to the cable and the weight that you save on it. Just to give you a reference, 6,912 standard cable is about 800 kilos a kilometer. It's almost a ton per kilometer. And this would just one fourth of that you could get efficient. So that's a big savings in terms of time, in terms of freight costs, in terms of build capacity and so on without changing how things are working in terms of network design today, just get things done faster.

hollow-core moves it a little differently. What hollow-core does, it solves a slightly different problem, latency. How long does it take to get light from one place to another place? That's a big breakthrough. We're working on that. We're introducing both solutions or we have. However, the thing is, if you want to build multiple data centers and find them that you don't have enough space in the same city or in the same campus, you've got to situate them further away, but you don't want to lose latency because AI compute is latency-sensitive, you've got the ability now to build data center interconnections using hollow-core fiber. And that is promising to help the construction problem in one go, the latency problem the other way.

But it also has another third thing that most people will talk about or they don't talk about a whole lot, is you can put a lot of energy into that fiber, which traditional solid glass core fibers, basically, they have a transmission limit once you put lots of energy, optical energy into it. And that has to do with the way glass and light interact. But with air, that thing goes away. So you can put a lot more energy, a lot more wavelengths, a lot more photons into it and you won't see a network error performance degradation. So that's two different things, but they're both going to be very important for this AI growth.

How do we move those from labs and trials into mainstream deployments?

hollow-core and multi-core aren't that new in the sense that it wasn't invented in the last two to three years or four years. It's actually invented at least 15 years ago or even more. The structure itself for multi-core I think was in advanced research even in Bell Labs times when they were quite active in the fiber domain back in the day. The ability for us to do that has been with us for some time. It's just that the incremental research cascades into more and more, let's say, efficient manufacturing technology, and that gets scaled when you get more demand and more customer traction. And now, with the AI use case, you've got this big pull, so we hope that that is. hollow-core is the same way. There used to be the structures of hollow-core, photonic crystal fibers and various types of band gap engineered fibers and different structures. I would say, gosh, at least 15 years going back.

I think the one that we are talking about today was published in 2014. That's 12 years ago. The one that's exciting people to the particular hollow-core structure that, the kind that people are ... the anti-resonant nodeless structure. But the photonic crystal engineering or science goes back 14-plus years. So you'll see literature going back all the way.

I want to round out our discussion with one more question. AI data center interconnects now require cables with thousands of fibers and getting there meant moving from a flat ribbon to intermittently bonded ribbon cable. What made that kind of evolution necessary, and what did it change for the operators who are actually building this AI infrastructure?

We've been doing at it for now eight, nine years or so selling what we call intermittently-bonded ribbon. Other people have different names for it. The original motivation was to build large scale fiber to the home is where many of telcos took up this particular structure, especially in Europe. The invention of this goes back, again, another 12, 14 years. 2010, 2012 is when it started becoming a lab reality. And then from there, it started scaling. I would say around 2017, 2018, demand started picking up when people ran out of duct space in the telco. In the AI thing, it's the same, but just a gigantically much more demand for that available rack space or the cable tray space. I think what it changed basically is the fact that you could use the ... The fact that a fiber is a circle and the cable is a circle, if you took a flat ribbon, you are basically putting a literally square peg in a round hole type of metaphor.

You end up wasting all that space around the rectangle inside a circle. But when you take circles of when you roll them into bundles, the fiber itself is a circle, then you make a bundle of bundles of bundles, and then you've got basically gigantic grouping of fibers that fit perfectly into the thing. So now, you've got utilization of your cable just goes up significantly. That is very attractive for the end customer because they're just getting whatever they need, getting fulfilled without so much more extra cost in cable, raw materials, plastics and other stuff. And it's also lighter weight. Everything sort of just rolls out. If you just use the space more efficiently, I think that's where everybody started noticing that this structure is more important to push through and that's kind of what happened. That's what we did.

Badri, thank you so much for sharing these insights with us. I can't wait to see where this space evolves.