Analysts Don Kellogg and Roger Entner are joined by Bill Long, Chief Product & Strategy Officer at Zayo, to discuss the rapid growth of AI hardware, Zayo’s strategy for capitalizing on the fiber infrastructure needed to support it, and what the next one to two years could bring for the industry.

00:00 Episode intro

00:24 Is AI in a bubble?

02:26 Orders for dark fiber

03:25 The shift from training to inference

04:50 400G and new fiber requirements

06:24 The technical ecosystem needed for operations

07:50 Rapidly accelerating hardware demands

11:18 AI as tech support

12:02 Zayo's unique business model

14:42 Predictions for the next 12 to 24 months

15:26 Episode wrap-up

Tags: telecom, telecommunications, wireless, prepaid, postpaid, cellular phone, Don Kellogg, Roger Entner, Bill Long, Zayo, Sampath, AI, fiber, data centers, lit fiber, dark fiber, training, inference, 400G, cloud, Nvidia, ChatGPT, Verizon, AT&T, cable, enterprise

Don: Hello, and welcome to the 312th episode of The Week with Roger, a conversation between analysts about all things telecom, media and technology by Recon Analytics. I'm Don Kellogg, and with me as always is Roger. Roger, how are you doing?

Roger: I'm great.

Don: So Roger, this week we're joined by Bill Long. Bill is the chief product and strategy officer at Zayo, and actually, I believe, Bill, you've been on the podcast once before, so welcome back. How are you doing?

Bill: Great, great to talk to you guys.

Don: Yeah, great having you.

Roger: You just got another friend of the show, as you see. So, pretty exciting times. You are at the epicenter of where telecom and AI meets, right?

Bill: Absolutely.

Roger: There's this whole discussion — is this a bubble or is this serious and will stick around? How does this look like from your perspective?

Bill: Yeah, I mean, if you just to get wonky with you for a minute, if you take a step fully back, we think there's long-term durable demand. One is because if you look at, you know, TSMC chip fabs are reserved five years out — you can see that the chip fabs are reserved out. Those chips have to go somewhere, they go in the data centers, the data centers need to be connected on a five-year horizon, we feel very good about the demand. But even if you take a further step out and you think about the economic value of a token that that token provides versus the cost of generating that token, there's this huge discrepancy where, you know, a poll request with a developer doing the work would have an economic value or an economic cost of $2,000. When an AI can do it, it can be $2. So there's this huge spread between economic value created and economic cost required. And the only way that you bring those back into balance is with more infrastructure. So we think that both when you look at sort of the TSMC chip fabs, reservations out five years, it gives us confidence. But larger, the economic value of a token versus the cost of the token, that gets solved with more infrastructure. That means long-term demand is in a really good spot.

Roger: Yeah, and Jim Patterson, who has a terrific weekly or bi-weekly newsletter, the Sunday Brief — he's now the CEO of a company that builds enclosures. They are like booked out for three years.

Bill: Yeah.

Roger: He sold everything for three years. And so we see all the same thing — you as the connectivity provider for long haul, for big data center things. And you're not only working in lit fiber, but also in dark fiber.

Bill: Yep.

Roger: What do you see there?

Bill: Yeah, I mean we've seen a huge increase in demand. I think if you look at the training runs, they're requiring bigger and bigger power. It's very hard to find multiple gigawatt pods of power that are available. So doing a single training run over larger geographies, those training pods need to be able to talk to each other. And we've seen a dramatic increase in long haul fiber demand because of that. You go back 18 to 24 months, and a large fiber order for us was 8, 10, 12 fibers. We're now seeing regular orders for 432 and even 864 fiber counts. So just order of magnitude increase in demand for long haul fiber.

Roger: Wow. And since it's dark, there's really — that's planning for the future, right? That's not even what they're seeing now.

Bill: Yeah, that's right.

Roger: Everybody, from Jensen Huang to Sam Altman to Amodei from Anthropic, are talking about the shift basically from training to inference. How does that materialize on the connectivity side?

Bill: Yeah, what's interesting is, of course, we're seeing a lot of demand for still training in the bulk fiber, long haul fiber for that. But when we're talking to a lot of the neoclouds and folks who are turning up data centers that we're providing connectivity to, they want to make sure that the physical infrastructure that's being architected can work for training and for inference. So if they're deploying GPUs today that are going to be used for a training run, they want to make sure those same GPUs in the future can be used for inference. So what that means is you need to not only allow a huge gigawatt-scale data center to talk to another gigawatt-scale data center — if you want to be doing inferencing, you need to be able to connect to where those inferencing requests are going to be coming from, where the data resides. So the things that matter for inferencing are: do you have the architecture to connect to the rest of the world? Make sure you're in an interconnection node that can connect to the rest of the world, that the scale can support the sort of multi-hundred gig connectivity that's required, and that you can do it over private networking, not just over the open internet. So people want the flexibility, and architecturally, it means you need to design — when you're deploying these data centers — design them for the future, not just for one training run and done.

Roger: Yeah, I know. And we're reading the press, an announcement from both AT&T and from Charter that 400G is becoming the standard for metro fiber for them, in response to AI. How does that materialize on your end, and does that mean that 400G is becoming the operational standard?

Bill: I was speaking at a data center conference in Germany and I said, we're much closer to 800 Gs, if not to terabit, right? It's no longer a G, it's a T.

Roger: Right.

Bill: Yeah. We're seeing demand for waves go through the roof. We're standing up whole systems for customers now. Even though the span speed is still at 400 gigs, it's 400 gig times many, many times. So we've turned up over the past quarter multiple 20-terabit systems, and what that looks like though is that in order to make it — to meet the demand that you need for that, it needs to be fiber rich. So if you're planning for 10, 20, 30, 400 gig kind of circuits, that will meet some of the demand. But really what you need is you need to be fiber rich such that you can meet the multi-tens, hundreds terabit type of connectivity requests that are coming in, and it's really only dense, fiber-rich infrastructure-focused providers who can do that. It's not the onesie-twosies type of over-the-top waves providers who can meet where the big demand is these days.

Roger: Yeah, and it's, you know, us — one of the largest, if not the largest provider here — are at the forefront of it. So when we look at the data center side, we have like 20 T's now. How does this filter down from the data center to long haul, to metro, to then individual enterprises? Do they need already dark fiber to be AI ready, or how does this then percolate to the entire system?

Bill: Yeah, I think the way that we think about this ecosystem is kind of like a two-by-two matrix, where you have training and inference for consumer and enterprise. And there are different pinch points that are arriving there. So we've talked a lot about training and sort of the requirements for long haul dark fiber on training. And so when you get to enterprise inference as an example, the way the architecture is changing there is, where your data resides versus where the intelligence that you need to apply to use that data on are in two different places. So your data may reside in the cloud, it may reside on premise, versus the intelligence that you want to use can sit somewhere else. So one of the demands for enterprise inference is you really need to have an architecture that can allow the data to connect with the intelligence in a way that's not going over the public internet, that can scale and is highly performing. So it basically looks like cloud on steroids, where when you're connecting to cloud, you had to have those same things, but the volume of traffic, the security profile of that traffic, the performance that requires, and the ecosystem that's enabling it are sort of putting another zero behind each of those elements, and that's what you get with AI.

Bill: Well, the answer is it was going to be. We've talked a lot about this long haul dark fiber, and I mentioned where in the old days 10, 12 fibers was a big fiber order for us, but we now are seeing orders for 432 fibers. When we did our long-term planning, the capacity we'd planned on taking 10 years to use up was gone in 18 months. So we recently announced a large, complete overpull of our full nationwide backbone deal we did with Nvidia, where we basically sat down and looked at our fiber map and showed the whole ecosystem — we showed all hyperscalers, we showed Nvidia, we showed others — that if we don't solve long haul fiber connectivity, we're going to be out of fiber. It was going to be a constraint. We've now, with the announced launch that we have that we're doing, multiple 24 different routes connecting the major hubs — long haul is taken care of. But that also had pile-on effects, and we needed to have the fiber, so we had to do a deal with Corning to make sure that we had the fibers. It takes lit equipment, so we've done deals with equipment providers to make sure we have the waves equipment that we're going to need to do it. So it affects the whole supply chain. And then you have the pile-on effects where there are spot use cases — one interesting use case that we're talking about now is that the whole internet was basically built to replicate and distribute the same bits, meaning if we're all streaming the same Stranger Things episode, that content is cached deeply in the network. As you have more bespoke or personalized AI content, the metro network starts to break. So we've solved for aggregate demand in a long haul network, but as each of these spot-by-spot use cases comes up, it's creating pinch points at different points in the network.

Roger: And the traffic is becoming — it is not symmetric, but it's becoming more symmetric.

Bill: That's right. It's both upstream, downstream — as you're uploading your ring camera videos into the cloud to monitor things, as, you know, we have a great use case where there's a major retail company that has distribution centers and they have over a hundred million dollars worth of robotics in each of those distribution centers, and they're streaming the live feeds from those robotics up into the cloud. Where they used to have a 10 meg connection into their warehouse, they now need a diverse 100 gig. So it's just an order of magnitude increase.

Roger: It's insane. And even on an individual level, AI has become a wonderful troubleshooting and tech support alternative. It took me eight hours to increase the RAM on one of our machines from 128 gigs to 256 gigs. It took hours for that memory to be trained by the system, all of that. And I used ChatGPT as my tech support, took pictures of the motherboard and the different colors of that to walk through it. And I'm not the only one who does this. And so you have videos you upload, you have all of these things. So yeah, it's just multiplying. So what are you doing differently than anybody else in the market? People look at you, and I don't think they properly fully appreciate what Zayo is doing. On one hand, I want to apologize that I didn't say — why don't you introduce Zayo — but I felt like with our listener base, we're inside telecom, you're a known name. But I think most people don't appreciate you the way you should be appreciated. How should they look at you?

Bill: Yeah, I think, in the larger context, we are a digital infrastructure provider, not a network service provider. So if you look at the legacy of where most telcos came from, it was a legacy of providing voice service or video service or something, and then they had the infrastructure in service to that — providing voice or video or something like that. Zayo came at it from a completely opposite direction. We were an infrastructure-first company, and we offer services as just a mechanism to sort of access that underlying infrastructure. So what we're doing differently specifically for AI, because we are so infrastructure focused, we view this as a two-sided ecosystem, where the intelligence is the supply. So if you think about where that intelligence sits — in large data centers in cornfields in Iowa and distributed everywhere — job one is to provide the infrastructure, so the network, the fiber, to connect to the intelligence supply. On the flip side, the demand for that intelligence is with enterprises and consumers. And we are connecting the enterprise locations, whether that's headquarters locations, infrastructure they have in the clouds, branch offices, et cetera. And then consumers, whether that is through cell tower connectivity, connecting to cable providers or telecom providers — that's the demand. And then our job is to remove the friction from the demand to be able to get to the supply. So the way that we're approaching the market is infrastructure first, two-sided ecosystem, where we connect to the supply, we connect to the demand, and systematically make it easy such that you can access infrastructure at software speed. So that's very different than someone who was coming from offering wireless services first, like a Verizon or AT&T, or even a cable provider first, like Charter and Comcast. We are infrastructure first and bespoke by connecting to — we connect to over 2,000 data centers, and basically all the neoclouds and hyperscalers and AI labs are using us to connect. So that's where we're just a very different sort of provider out there.

Roger: Yeah, you have a different focus.

Bill: Yes.

Roger: You don't have retail customers.

Bill: Exactly.

Roger: You live and die by these large organizations and how to service them. So, maybe we can close our conversation by — how should they do things differently and get prepared, or what are we going to see in the next 12 to 24 months? Because I don't think anybody can look out much further than that.

Bill: Yeah, I mean, even 12 to 24 months feels like decades in the old world, the amount of action that's happening.

Roger: Yeah.

Bill: But it's really plan for agility. I talked about earlier how AI looks like cloud on steroids. I mean it — it starts with, make sure your architecture has the agility to connect to who you're going to need to connect to, not just today, but for the future, that you can do it at the scale you need it, with the security you require, and the performance that's needed. So it's really plan for the unknown instead of being surprised by it.

Don: Awesome, Bill. Thank you for coming. This was fascinating, and I'm sure this will be with us for a long time. And I'm looking forward to having you, or Sampath, coming on the show to give us an update.

Bill: Great. Great catching up.

Don: All right, thank you.

Roger: All right, thanks, gentlemen.

Don: Roger, we'll talk next week.

Roger: Talk to you next week.