Hey guys, all right, we're just gonna kick off the final session of the day. Thank you for sticking around. Me and Dave Schaeffer, who's the Founder and CEO of Cogent, are the thing standing between you and the wine selection that's sitting right outside. If you guys wanna kinda sneak out and grab one, you know, for this session, I'm not gonna blame you. I'd like to say thank you to Dave for being here and joining us and doing the meetings.
The funny story is when I moved back from Hong Kong with JP Morgan, the very first banking deal I ever worked on was with Dave Schaeffer, a company called PathNet, where Richard Jalkut was the CEO, the former head of NYNEX after the Bell Atlantic merger at the time. I've known Dave a really long time. You know, to kinda tie some of the themes that we've been talking about together over the course of the day, one topic I wanted to talk about was, Dave, you've been around a long time, and I mean that in the best possible way.
Oh, that-
I mentioned this before in another conversation, but you know, we were there for you know, the rise of the Internet, and you know, a lot of people got rich, but there wasn't a lot of value created, lasting value created in the connectivity sector. Then we got to the cloud, and I think here again, it was a lot of value created in the cloud itself, but you know, for all the connectivity that it required, there wasn't a lot of lasting incremental connectivity value created. Now we've got this big conversation about you know, the rise of AI. You know, there's a whole new class of hyperscale and inference data centers that are being built in all sorts of crazy locations, and they need to be connected.
You know, for a guy who's kinda been around the block a couple times, the question is, you know, Is it gonna be different this time? Are we gonna have lasting value created in the connectivity space? Or is it just gonna be another, you know, it's gonna be the software guys and the people who are leveraging the AI technology that really accumulate all the value.
First of all, I wanna thank Dave for inviting me, congratulate him on his new gig.
Thank you.
Thank investors for hanging around this late in the day to hear what we have to say. I think I'm gonna answer your question in two parts, Dave. First of all, AI is real, and it is transformational, and it will probably change the world to an equal extent in the way the internet changed the world. The internet is real. You are absolutely right. All of the companies that tried to build value as ISPs have mostly failed, and in fact, most of that wealth was transferred away from legacy telco providers and distributed to consumers and society. $2 trillion was invested in the bubble of telecom in the late 1990s and early 2000s, and society got a great return on that.
The service providers who delivered legacy services watched their value dissipate in that, you know, telecom services peaked at about 8.5% of the S&P 500. Today it's probably about 1.5%. Telecom services peaked in 2000 as a percentage of GDP at about 4.5% of GDP. Today, it's about 1.1% of GDP in OECD countries. Yet, we are using more telecom services today than we've ever used. The internet was extremely deflationary. It delaminated the application from the network, and it became very challenging for service providers to capture that value. Most of that value was distributed to the end users. Some of that value was accrued by new business models that sat on top of the internet.
Companies like Netflix and Meta and Google and Amazon could not have existed without the internet, and they captured a portion of that value, but the real winner has been society. The migration from premise-based computing to cloud is another example of an improvement in operational efficiency, the transference of value to the end user, a reduction in the cost of business operations, and new business formation. Some of that value is captured by the cloud provider, but most of it goes to the end user. Now let's bring that forward to your question about AI. AI will create value. There are really two fundamental questions to focus on. The first of those is the economic value of an AI output equal or greater in value than all of the costs of the input to create that output? The answer today is no.
Companies are losing money on AI outputs. However, the cost of those inputs is falling precipitously. The inputs really come in two flavors. First, tokenization, which is falling at about 85% a year. The second is the building of large language models. Now, the CPU or GPU cost of those models is falling quickly, but the models are both getting bigger and more complex, and they are being modeled against larger and larger datasets. The only reason AI is possible today is because of the internet. Now, over the course of the internet, the internet generated probably something like 20,000 zettabytes of data. Now, the vast majority of that data just disappeared. It was never stored anywhere. About 1,000 zettabytes was stored and is the raw material for these large language models to be built on.
Now, that data is being stored at a much higher rate, and it now has real economic value. That's why maybe Grok has value coming out of Twitter. They have the raw material of the data of those Twitter users to give a proprietary advantage to xAI. Every company looks at its underlying data as a scarce resource that will be valuable in building models. Now, the models are evolving very rapidly. They will be distributed, and people will then make inferences against those models. The second part of AI is taking a model and applying it to new data, and that'll happen in three flavors. The easiest case, you ask a question, the data that you're asking a question of is at an edge site, the model is applied to it, and the answer comes back to you almost instantaneously.
The second iteration is that the question you're asking needs to have your proprietary data that doesn't exist anywhere except inside of your network or on your dev device. That's gonna actually radically change the internet. The internet started as a completely symmetric network. Sending and receiving was equal. Over time, the business models have changed the internet, and today it is primarily a download network. You send a few request bits, and you get a Netflix movie back. You send a few request bits, and you get a TikTok video back. Very asymmetric. Now, we're gonna see a big shift in traffic patterns at the edge, where data needs to be uploaded to be run against those large language models and spit out answers.
The third iteration of an answer will come with both of those datasets are insufficient to get the right answer, and you need to go to a massive data farm that has maybe tens of thousands of zettabytes of data to be crunched and will come back in a slightly elongated answer and timeframe. I think we're still in that developmental stage. The other wild card in this equation is the way in which the actual training is done. It is very power-intensive. If we look at the developed world, it has been built for 60 years on a simple model of 2% GDP growth, 1% conservation improvement, and therefore a 1% growth in power availability. The model worked well with 5% slack in the system.
Traditional data centers, non-AI, have used up, in the past three years, today, about 2.2% of all power that is produced in the developed world. The second piece of this equation is the surge in power demand coming from AI. It is already 5% or, excuse me, 2.5% of the load on the grid. Combined, the two are at 4.7% or almost 5%. The slack is gone out of the model, so people are building data centers wherever they can find power in really remote places. We can't just keep throwing power at the problems. The models have to get much more efficient, or the way in which we're building the models has to be based on a different processor technology. That question is not yet answered.
The real issue will be no one has a good monetization model for AI. It is highly likely that the companies that are spending the money, which have deep pockets and huge access to capital, won't have the protected moats to extract economic rent and will then be forced to effectively commoditize that service, and the winner again will be both the consumer and the companies that figure out business models that sit on top of AI. My expectation is the $1 trillion that's going into training facilities probably is gonna have a return on capital far below the average cost of capital.
Wow, okay. We didn't talk about the connectivity part of it, but I'll let Bora kind of maybe take the next 17-minute question.
We'll see if it's 17 minutes. Really, I think, Dave, thoughtful summary, and it's nice to kind of watch how you transition across the different assets and into where we are now. I guess one, maybe two questions on my mind. Where do you see connectivity benefiting from all this demand? I mean, you run one of the largest B2B connectivity businesses in the U.S., so what sort of green shoots are you expecting from all this data center demand? And will we get out of that kind of conundrum where, you know, additional connectivity or additional bits hasn't really translated into additional returns for the sector?
I think there are three benefits. First, as I mentioned, the raw data that is generated over the internet now has value, so that means more data is gonna be collected and stored, and that will be helpful to the core transit business. We're the largest carrier of transit traffic globally, carrying about 2 EB a day of traffic, about 25% of global internet transit traffic. Secondly, there will be a need for significant fiber investment at the edge as coax plant is unable to give the upload speeds that are gonna be required. We will see significant pressure on mobile networks, where the cost per bit mile is generally about three orders of magnitude higher than on a fiber network. There will be a continued push to proliferate fiber to the edge.
That's not a guarantee that the people that are building that fiber are gonna make a return, but society will win by that proliferation. Then third, over the next decade, it is likely that the training facilities will not be co-resident with where the data is stored. They'll be built in some unusual places where land and cooling and water and power are available. As a result of that, there will be a tremendous need for transport services, Layer one. In delivering the internet, the internet does three things that other networks can't do. It's completely resilient. It was designed to survive a nuclear attack. two, it packetizes data, so the transport layer is extremely efficiently utilized. three, it can sit on top of any other network. It's a chameleon, so it's everywhere. It's application agnostic, it's flexible, and that is why the internet has succeeded.
AI has a different need. The GPUs represent most of the capital, so therefore, you want them to run very efficiently, and as such, you don't want any buffering and any discrepancies and latency. You want a Layer one connection between them. Two, the concept of TCP/IP doesn't work in AI training because by its very nature it has a series of handshakes that require buffering. There will be a tremendous need for old-fashioned, more expensive per bit mile, Layer one connectivity. That'll be dark fiber, that'll be wavelengths.
It seems like your thesis is that the compute may exist in even more remote rural areas so that there's affordability, I'm assuming, is kind of the rationale. It's more available, it's more affordable, and we need more of that. Is this the second layer that you expect the mid-haul and kind of long-haul routes to be developed, new construction to get there? Or how will they be connected?
I think there'll be three ways in which new routes get deployed. You will use existing inventory that was put in 20, 30, 40 years ago that sat fallow for decades and will be used. Secondly, we'll see fiber pulled through empty conduits that were waiting for an application to justify pulling fiber. The third part will be new greenfield builds to remote sites. The problem there is, there is so much monopsony power among a handful of hyperscalers that they are able to force those builds at very low IRRs, typically low single digits. They don't want you to go bankrupt, but they don't want you to make any money on them. As a result, we're seeing companies today deploy capital below their cost of capital. It's part of the reason why we've been extremely disciplined about where we deploy capital.
Dave, maybe pivoting a little bit to the business, and there's a lot to unpack here. You know, I was just looking at the stock price chart, and a year ago today, the stock was about $60. Today, it's about $19. I was looking at the first quarter result. Looking at the first quarter of 2026 result model. Doesn't look like the business has gone down 70%. You know, is there something wildly different about Cogent today than a year ago, other than some of the things we know about the forced sales from the banks and other things?
You know, companies operate over a long arc, and stocks move in short movements. Our stock reaction has been far more violent than the operations justify. You know, we went public in 2005 at $6 a share and then had organic compounded growth for 18 years of 10.2% with an average of 220 basis points a year of margin expansion and a policy to return $2 billion to shareholders. The stock appreciated actually 15x during that period. We acquired Sprint. Sprint was complicated, it was messy, and we were paid $700 million in cash over a 54-month period by T-Mobile. That's a highly unusual transaction. I think part of the stock sell-off has been the noise around the complexity of that transaction, the fact that our total top-line growth went from positive to negative.
That was not a surprise. It was not something that we didn't anticipate. If we disaggregate the growth in the nine quarters that we have owned the Sprint assets, Cogent's core business grew 27% during that period, but the acquired Sprint business declined at 64%. A lot of times, investors don't wanna unscramble that egg and see what is going on.
You foreshadowed all that, right? I mean, that was part of the
It's why we got $700 million. Listen, Mike Sievert and T-Mobile were not dumb. They knew what they were selling us. They knew they had to subsidize it, but they also knew they could not turn the fallow Sprint network into a valuable asset. That was our primary reason for doing a transaction. The second part of it has been the pace at which we've been able to grow the revenue in a new business, the Wavelength business. We grew that business 100% year-over-year, but it's off of a small base. We grew it 19% sequentially, quarter-over-quarter. Investors wanted more and faster. Also, with the work that we had to do to repurpose both the buildings and the Sprint network, our leverage ticked up. It went up from 3.5 times to 6.6 times.
I think our stock reaction is in part due to that increase in leverage, the decrease in our growth rate. Now we're back to the point of inflecting to positive top-line growth. It'll probably be on a multi-year basis, somewhere between 6% and 8% going forward, not the 10.2% that we did for 18 years. When we did the acquisition, our EBITDA margins went from 40.5% the quarter before the acquisition to 1% the quarter after. Now we've clawed back and got those margins to 22%. We need to get them back to 40%, and we will. The way we do that is through the sale of on-net services.
The reason the Sprint business was in such bad condition, they weren't utilizing the network, they were selling a variety of services in which they had no core competency, and they were negative gross margin. They were in the worst possible situation. They had negative revenue growth and negative gross margin. That's why we were paid to take it over. We've gotten the acquired business to break even. Now, the acquired business is a lot smaller. It's declined by 64%. It went from 42% of the combined company's revenues to today 30% of the company's combined revenues. Now it is small enough and stabilized enough that the underlying growth will show through. During that period, with negative top line, we actually grew our EBITDA margins over a two-year period over 1,000 basis points a year. 800 basis points just last year.
That's almost unheard of. It all came from grooming, through cost-cutting, through product optimization. Now, going forward, we'll have some of that still ahead of us, but we'll also have the high operating leverage of on-net services. 80% of our sales last quarter were on-net. The day we acquired Sprint, we were 76% on-net, 24% off-net. Sprint was 93% off-net, 7% on-net. The combined company, first quarter out of the box, was 47% on-net, 48% off-net, and 5% non-core. The combined company at the end of Q4 2025 was 61% on-net, 38% off-net, and 1% non-core. It is that improvement in on-net that's gonna drive the improvement in our margins. I think the stock will recover as investors dig in and do the work to understand what's actually happening and not just look at the headline.
You picked up on a theme that's been relevant throughout the day, starting with the balance sheet and the leverage. Tell us a bit about ways to delever, especially around kind of creative asset sales. I know there's interest in the switch sites and the central offices and monetizing those and refurbing those. That's an idea that, you know, really it's been tested a lot. We've seen some successful cases of it. How good of an idea do you really think that is for operators, and how much opportunity is there? Also, conduit you mentioned too. What else might be the next wave of interesting kind of disposals?
We did three things to monetize unusual assets to help us fix the balance sheet. The first of those, which was truly novel, is we leased out a portion of our inventory of IP address space and then securitized that with an asset-backed securitization. No one had ever done that. We actually talked to six banks about it. Four of them thought I was a lunatic when I walked in and said, "What's an IP address and how do you securitize it, and what's the tangible asset?
You should have had us do the market study. We would've helped with that.
Bank of America was one of those.
We were able to get a few banks to understand it. We marketed it and raised $380 million off of that asset through an SPV that is a ring-fenced entity with an asset-backed securitization. The second thing we focused on were the buildings. We acquired 482 fee simple owned technical buildings that totaled 1.9 million sq ft and had existing 230 MW of power going to those buildings. One problem was they were full of old telephone switches. There were 23,500 cabinets of old telephone switches still in place. To give you a sense of what that looks like, these are eight-foot-tall cabinets, 23 inches wide, and if you lined them up, it would go about 8.5 miles. That was a lot of gear to dispose of.
We cleaned out those facilities. The second problem was all of those facilities were built like central offices. They're -48 V DC. Computers run on AC 120. We chose to go in and convert the power systems in a subset of sites. Third, we said that some of those sites are too big and have too much power for us to be able to effectively fill them up with our go-to-market model. We are selling 24 of the facilities that have 109 MW, so the larger facilities, 1 million sq ft. We currently have 10 of those under letter of intent to be sold. That conversion started in June 2024 and was completed in June 2025. We are looking at pockets of value to monetize.
We do not have conduit to monetize because Sprint network was the first fiber optic network, but it's direct buried in armor. We do have dark fiber. We have sold a limited amount, but our primary way of monetizing that fiber is to use it for the Wavelength network.
I'm just gonna say thank you, Dave, for playing cleanup here with us today. There's 45 seconds left. I'm never gonna get a whole question in, so,
I won't get any answer.
I will invite you to kind of be part of our little gathering outside before you go, so.
Well, I actually gotta go to my next group of meetings on site.
Oh, you have another group.
A small group.
Okay.
Okay. Well, for everyone else, we do have a small happy hour happening. It's probably already started, and we appreciate you. Thanks, Dave, for having you know, like you said, playing cleanup hitter. We appreciate it.
Hey, thank you all.
Thanks so much.