Cogent Communications Holdings, Inc. (CCOI) Earnings Call Transcript & Summary

Always pleased to have Dave Schaeffer from Cogent here with us. He's kind enough to come every year and great to hear the latest from you, Dave. So thank you again.

Michael, thank you for hosting me. As always, thank BofA for a great venue and thank investors for taking time out of their day to hear what we want to talk about.

Outstanding. So I had a number of questions that I want to go through. I will leave a few minutes at the end if there are any questions from the audience. So when I mentioned that, there should be microphones going around, just raise your hand and we'll take time for that. So Dave, I think you would admit that growth has been disappointing and below your own projections for the last year or so due to a number of factors. So -- can you just remind us maybe what went right and what went wrong or maybe not wrong, but didn't meet expectations in the last 12 or 18 months? And then why you believe that's going to change in the next 12 or 18 months?

So let me start by laying kind of the backdrop. Cogent had done no acquisitions between June of 2005 and May of 2023. In that 18-year period, we had a compounded average growth rate of 10.2% a year. Our corporate segment actually was outperforming that growing at 11.2% a year over that 18-year period. And because the majority of our sales, 76% of our revenues were on-net services, we were experiencing consistent margin expansion. Our EBITDA margins expanded at approximately 220 basis points a year.

So the on-net, it's about 90% incremental margin to growth, right? Correct.

No, the actual total margins had expanded. So we went public in June of '05 with 0 EBITDA margins. When we acquired Sprint, our EBITDA margins were 40.5%. We were acquiring a business that was in decline and burning cash. The business we acquired for the 3 previous years had declined at 10.9% a year, and it represented 42% of the revenues of the combined company. So we expected our total revenue growth to go negative. We also knew that the company we acquired was operating a number of unprofitable products and network services and had a negative 60% EBITDA margin. We outlined a plan to take costs out -- and more importantly, we had a program to repurpose the physical assets of that business that have been sitting dormant since 2015 and turn that old TDM voice network that was built at a capital cost of over $20 billion in 1980s into a modern wavelength network to serve international carriers, regional carriers and content distributors. We did not anticipate the tailwind of AI. So that was a bluebird. I will not profess. I was Carvoyant in September of '22 when we signed the deal, we knew that AI was going to burst upon the scenes and create a $7 trillion capital investment cycle. So what went right? The ability to take costs out were actually better than we expected. We actually raised our cost reduction targets, and we achieved those cost reductions quicker than we expected. What went wrong was we lost more of that revenue than we expected. Now in a way, that's a blessing in the skies because that revenue was negative margin. Today, the Sprint customer base is margin 0. So it doesn't help, but it doesn't hurt. It has shrunk from 42% of our total revenues to only 15% of our revenues. We today have a business that is at least breakeven in those Sprint customers. The second part of the transaction was the establishment of a new business, a wavelength business. We actually reenabled the Sprint network on budget and on time. We did not exceed our goals, but we did not fall short. We also made the election after the transaction was announced to invest an additional $100 million into the physical buildings that we acquired and convert those former switch sites to data centers. We had a program to divest of 24 of the 125 that we did convert. We chose those based on their location, size and proximity to major markets. We then completed that. There was a great deal of skepticism that the company already was spending money on building a wave network and then committing an additional $100 million into what was viewed as a speculative endeavor. We, I think, proved that investment to be correct. We sold 10 of those 125 facilities for $225 million to $24.1 million net because there was very little transaction costs associated with that divestiture. We did that in-house and we used those proceeds to reduce debt on our balance sheet. So I think on the data center monetization, it was not originally anticipated. We did spend more than we expected, but we got more than we spent back. And we still have 14 data centers that we plan to divest of using those proceeds to further delever and we will be left with the remaining footprint that while we view it as not monetizable through sale, we should be able to lease out that capacity.

And Dave, just quickly, on the 14 remaining data centers, do they look like the 10 that you already sold, meaning location, occupancy? Should we expect similar valuation per megawatt for the remaining 14?

We have never sold a data center prior to this. We had no real view on how well we will do other than we felt that we would do much better than the $100 million invested. We ended up getting $4.3 million a megawatt. The 10 facilities we sold were about 53 megawatts getting to the 225. The remaining facilities are 54 megawatts, so about equivalent. And at least from our perspective, work relatively similar. I think the expectations should be about the same. Each of the facilities has some different characteristics, whether it be physical site or availability of incremental power almost all of these facilities are empty with no real revenue associated with them. We have the added benefit that when we divested of the first 10, we reduced a negative drag on our EBITDA by $7 million a year. We were effectively paying the real estate taxes, personal property taxes and maintenance of facilities with no revenue.

And these are old wire centers, right? They were not purpose-built data centers, old wire centers, you converted with $100 million to make them into data centers. So they're different maybe than a traditional data center.

That's correct.

Correct.

What these were if you look at a typical facility, it was 15 miles from a downtown was 6 or 7 acres with a 40,000-foot building with 4 to 5 megawatts of inbound power full of old Class 4 and Class 5 telephone switches. There were 23,800 cabinets of telephone switch gear that we removed from the footprint and disposed of. We had to convert these facilities from negative 48-volt DC power plant, which is how a switch site operates to AC120, which is how a data center operates. We had to refurbish the air conditioning, the fire suppression systems, the security systems and do some cosmetic refurbishment to then make these basic data centers. These are still facilities that have only about 100 watts per square foot and are not today able to support high-density liquid-cooled cabinets. Now they are modifiable to that, but they are smaller, much more distributed than the types of data centers that are in the headlines today that people are protesting where people are trying to stop the construction of a 1 gigawatt million square foot facility on 1,000 acres of land. These are much more innocuous. And they, I think, will have a great deal of applicability for inference. In fact, I square, the purchaser of the first 10, did not disclose to us during the purchase process, what their ultimate plan was. But subsequent to buying them and now standing up a management team to start to commercialize these facilities, they disclosed to us that their ultimate plan is to make these inference facilities and kind of have token repositories distributed around the country where they can then be providing inference as a service to Neo clouds and hyperscalers. They ultimately tend to invest $1 billion at least that's what their public statement is. $225 million of that came to us. Now to what else went right and wrong. I want to finish the answer to your question. Cogent's core business actually accelerated. 84% of our revenues come from selling Internet-based services. That business had been negatively impacted by the pandemic. It had seen its growth rate fall from that 10.2% down to 5% pre-acquisition. In the 3 years since the acquisition, the core Cogent business has actually seen its growth rate reaccelerate to about 9%. That business over a 3-year period grew 29%. The acquired Sprint business rate of decline accelerated and declined in totality at 69%. So there is both good and bad post transaction in our 2 customer bases. Probably the biggest area of disappointment has been in our wavelength business, our brand new business. We knew that it was going to take close to 1.5 years to 2 years to do the conversion. We actually started on that conversion prior to the acquisition after the deal was announced. And we completed the enablement of an initial footprint of 800 data centers by December 31, 2024. So about 18 months after the deal was closed, right on plan, right on budget. Where we were disappointed as we hoped to be able to sell more wavelength services on an interim basis as we were bringing these facilities online. We sold about 1,000 waves in that interim period, not the 3,000 that we had hoped to. Then once the entire footprint was enabled, we have continued to grow that business. Last quarter, it grew 62% year-over-year. It grew 9.2% on a sequential basis and has performed well, but off of a small base. We today have 3% market share in the North American intercity wavelength market. Our ultimate goal is to get to 25% share. We were hoping to do that by midyear '28. I don't believe sitting here in the fall of '26, that is realistic. I think the ultimate goal is realistic, but the pace to get there is going to take longer. And I think the 2 reasons why our ramp in revenue growth was slower than we anticipated were, one, our need to build credibility in this new product. I think we had earned and have a great deal of credibility in our core product suite, but I think many customers did not associate Cogent with wavelengths and were comprehensive about our ability to deliver the quality that we represented. I think we've solved that problem. We have now sold wavelengths to 548 unique customers. We've delivered service in 608 of the ultimate 1,137 locations that we brought online. I think the second challenge has been the supply chain challenges in the entire technology and telecom ecosystem that we did not anticipate. Now on the positive side, we get the tailwind of a new incremental use case for wavelengths that is AI training, connecting a data source to a training facility and doing it with defined latency. While that is more expensive and harder to administer than using the Internet, it does allow for more optimal utilization of the GPU capacity. Since 60% of the capital cost of a training facility is in the GPUs, maximizing their efficiency is critical. And with that...

How does it maximize the efficiency of the GPUs?

So, every large language model works on the same basic architecture. They each have variances. They each do it slightly different, and they are evolving. But what they do is take a large set of data that was collected over the Internet, literally exabytes of data. They tokenize that data, breaking words into recognizable tokens that GPUs can process. They then pose a question to that set of data that they already know the answer to. And the GPUs crank through that data multiple times and develop a pattern upon which the predicted answer and the actual answer are matched. They do that recursively, repeatedly. And over time, they build a neural network that can then be applied to a new set of data. Most of the data storage exists in preexisting data centers and major data collection points, Northern Virginia, Ashburn, San Jose, Greater Chicago market, Greater Dallas market, Miami market, Amsterdam in the Netherlands, the U.K., mostly in England, Frankfurt. And as a result, there's a couple of dozen locations globally where most data sits. Most of the training is occurring in locations that are remote from those locations. There are locations where there is power and space available. We're seeing new data centers being constructed in some very unobvious locations, Southern Louisiana, Central North Dakota, Western Wyoming, Western Texas. And they're going to those locations because power, land, water are available. So in order to train, you have to take the data out of the data center where it's stored, move it to where the GPU is and then return it back after the training is complete. And that process is best done over a wavelength because a wavelength is a deterministic point-to-point service. On a per bit mile basis, it's actually about 2.5x more expensive than using the public -- but if you use the public Internet to transmit that information, you would have to buffer that information or you would have to be able to degrade the productivity of your GPUs. So all of the AI training is today done using wavelengths as an incremental demand set to what we expected from regional networks, from international carriers and content deliverers. And this trend is going to continue for multiple years, creating a huge amount of demand. Now what that training demand has done, though, is tax the entire supply chain, whether it be power generation, whether it be power transmission, whether it be land, whether it be cooling systems, switchgear, fire protection. At every level, this surge in demand has created shortages. There's a tremendous memory shortage where prices have gone up by a factor of 5. Telecom equipment, which would typically be orderable and deliverable in 90 days, has now got a 3-year delivery window. You've got a new class of buyers at a traditional telecom equipment supplier, 90% of their revenues were coming from service providers like Cogent or AT&T or Deutsche Telekom. Now 60% of their demand comes from 1 of 4 major hyperscalers. This is a shock to the system that has created a number of supply chain constraints. That impacts us 2 ways. One, it's raised prices for equipment that Cogent and other service providers buy. It's created telecom equipment inflation, which is unheard of for 50 years, price performance had consistently improved. This is the first time in the digital age where price per unit of productivity is going up. That will get reverted and we will return to the Moore's Law type of price declines, but that's probably going to take a couple of years to work through all of the systematic constraints.

Can I ask you, Dave, before I forget. You mentioned 3% share today, hoping to get to around 25% over time. And obviously, AI has been unexpected, right, positive incremental. How much is AI training spending on wavelength today in general -- in general, how much are they spending today?

Today, AI training, which is coming both from hyperscalers and Neo clouds, represents about 1/4 of the total market, but it's the fastest-growing segment. It is not always easily measurable because the major hyperscalers also operate traditional cloud distribution businesses, and they typically buy wavelengths through a centralized purchasing mechanism and don't always disclose to Cogent or any other supplier, what the ultimate use of that wave is. So when a Microsoft or a Google or a Meta or an Amazon buys a wavelength, they don't put a little check mark and see.

Yes, you don't know exactly how they're using it. You can't differentiate. Is your market share of that 25%, is that in line with your 3% share? Or is it higher?

I think it's a little higher with that group of customers for 2 reasons. One, for the existing hyperscalers, we had long-standing 15-plus year relationships and have been able to leverage those. Secondly, for the Neo clouds, our ability to have such a broad footprint and rapid deployment schedule is differentiated.