For telecom services and communications infrastructure here at RBC. Welcome, everybody. I'm pleased to spend the next half hour with Dave Schaeffer, CEO and founder of Cogent Communications. Welcome, Dave.
Hey, thank you for hosting me, John. I'd like to thank RBC for a great venue, and maybe most importantly, thank all the investors in the room for taking some time to hear about Cogent.
I'm going to ask some big picture questions, get a little bit more detailed onto wavelengths, talk a little bit about asset modernization, legacy business, and then wrap up with a couple of balance sheet-related questions. Before that, why don't we kind of dial back? I think within the past day, there's been some developments on your front around share repurchases. Maybe you can tell us what's the latest update since the quarterly earnings call.
Well, Cogent has had a program of returning capital to shareholders since 2006. We've returned about $1.9 billion, and we've done that through a combination of dividends and buybacks. We made the decision on the last earnings call to reduce our dividend and indicate that we would not begin increasing it until we reach four times net leverage. We're currently at about 6.6 times net leverage, so we have a significant amount of delevering that'll come through a combination of growth in EBITDA as well as asset sale and monetization of non-core assets. We also initially paused our buyback program, and the only thing that we announced yesterday after the market closed was the board authorized us to potentially resume it. There's not a guarantee that we're going to be in the market buying back, but we now have that flexibility with about $105 million available under the authorization program.
It is just an acknowledgment that that additional degree of flexibility will give us the way to maximize return of capital.
Any audience questions before we kind of move on to the topics that I outlined? Turning first to big picture, I guess interested in exploring your views on which enterprise verticals are seeing different demand patterns, accelerated demand for high-speed data versus, say, the rest of the market. What are you noticing from basically your net central business as well as your retail business?
First of all, I'm going to start with our end user business, which is a combination of corporate and enterprise customers. Cogent's corporate business is focused on businesses with multiple sites that typically have one or more locations in one of the 1,875 multi-tenant office buildings that we're connected to. That business had grown historically at about 11% a year organically from going public in 2005 up through the beginning of the pandemic. That growth rate went negative and has subsequently returned to growing at about 3%. That business is driven predominantly by on-net sales, with roughly 60% of the revenues and 80% of the connections being on net. That business is growing at a slower rate for two reasons: one, the impact of the pandemic, and two, we did acquire some corporate customers from the Sprint acquisition.
The majority of the customers that we acquired with Sprint, however, were enterprise customers. The totality of the Sprint business was declining at 10.6% a year prior to the acquisition, and it was virtually all off-net services, and much of the revenue was coming from non-core products. We actually accelerated that rate of revenue decline to over 24%, purging the non-core products, moving some off-net customers to on-net, and then finally having some locations that we have just discontinued service. We are through the majority of that revenue decline. The rate in that enterprise and corporate acquired customer base decline has moderated to low single digits. What we have seen across all of the customers has been three key trends. One, a need for greater bandwidth.
The average on-net connection has grown in three years from 550 Mb up from when the company was started at 100 Mb to today about 950 Mb. Two, there's been a decline in VPN services as many customers just use the internet. Three, specific to the acquired Sprint base, we have groomed locations where we could not use fiber in countries we are licensed in. Pivoting to the second major customer segment, which is our net-centric base, we sell services in 58 countries, 1,900 data centers around the world. We carry a quarter of the world's internet traffic, approximately 2 EB a day of traffic. That business continues to grow. Traffic growth was up 9% year- over- year and up 5% sequentially. Prices do come down.
That business is also becoming more global as the U.S. percentage of global internet traffic has declined, and the rest of the world, particularly some of the more developing countries, are growing much faster. As a result, our net-centric revenue growth has been about 8%. Blending that together, the total legacy Cogent business is growing at about 5%. The acquired Sprint business is today declining at about 2%.
As you maybe just kind of one-off on CapEx then, so what level of investments in the network do you anticipate in order to kind of support some of these trends around speed and capacity?
With the acquisition of Sprint, we spent $100 million of extraordinary CapEx in converting former telephone switch sites into data centers. We spent $50 million physically interconnecting the Sprint network to our network. We anticipate our capital spending on a go-forward basis to be roughly $100 million a year, and then supplementing that will be about $40 million annually in principal payments on capital leases. Really just two different places on the cash flow statement for CapEx are about $140 million in total. That includes normal wear and tear and replacement technology refresh and capacity augmentation.
Talking about kind of expected market growth and the TAM for wavelengths, can you give examples of types of tasks or workloads where waves are best suited to your product versus some of your other traditional services on net?
Yeah. Cogent's primary business, and its only business for its first 20 years of delivering service, was selling internet-based services and VPNs that sat on top of the internet. The internet is the most ubiquitous, easiest to use, and lowest cost way to move bits. However, there are three characteristics that certain applications require that the internet is not good at, whether it's high security, large packet size, or most importantly, predefined latency. When we had the opportunity to acquire Sprint Global Markets Group, it was really two acquisitions. One, the operating business, the cash flow associated with it, which was negative, and the customer base, which was unprofitable. Secondly, we acquired a dormant physical network that we have now repurposed to sell wavelengths in over 1,000 locations.
The wavelength customers, on a per-mile basis, pay about two and a half times as much per bit mile as they would pay using the Internet, but they get these incremental features that are critical to their applications. What we have seen is three historical customer bases and one new customer base. The legacy base were companies that had regional networks that needed to interconnect islands of traffic. Think of cable companies or fiber overbuilders. Secondly, international carriers that needed to extend their network to serve multinational customers. Finally, content distribution networks that could either be third-party CDNs or could be part of a hyperscaler's business, a more traditional AWS or Azure-type application. There is now a fourth use case that is the fastest growing and also sufficiently large that is causing the entire market to grow, and that is AI training.
AI requires huge amounts of power and processing that may not be co-resident to where the data that has been collected sits. Because the GPUs in those facilities represent half of the capital cost of the facility, you need to run them at high efficiency. Using wavelengths allows you to not buffer that traffic. That incremental use case has been a tailwind to the wavelength market.
Any behavior around demand for wavelengths as models evolve? They're still getting bigger, but is there a kind of a trajectory where that might no longer be the case and as we pivot towards more inference workloads?
I think there's two different parts to that answer, John. First of all, large language models are developing rapidly, and the same data can be trained multiple times with different outputs and better refinement and greater efficiency. I see training continuing for a very long time. Secondly, the output of those training efforts is then distributed around the perimeter of the network and is available for inference. That inference occurs by taking smaller sets of data, running them against the outputs of these large language models, and making a specific recommendation. Almost all of that input and output will occur over the public internet. The internet is a winner two ways from AI. The data that is being collected now has economic value for training, and then two, the outputs of that training are now valuable to end users that will drive greater internet bit usage.
I think the large language models will continue alongside a higher and higher percentage of inference.
As you look at your target for wave market share, what portion do you see coming from taking existing customers from traditional on-net services to wavelengths versus bringing on net new traffic?
Our initial expectation was that the market was not growing. We had lived through a period where we built the world's largest internet backbone predicated on the idea of market growth that never materialized. I think going forward, we are seeing aggregate growth in the market in the 5%-10% annually in revenue terms. The market is growing faster than that in bit volume as wavelength sizes increase. A wavelength is sold with three dimensions to price it. One, the size of the connection. Today, the market is one of three sizes: 10 gig, 100 gig, or 400 gig. Two, the physical length of the path. The longer the path, the more you pay for it. Third, the length of the contract. The longer the duration, the lower the price.
We are seeing a migration away from 10 gig to 100 and a smaller migration from 100 to 400. We sized the addressable market two ways. One, we did a top-down analysis where we looked at four independent research analyses that sized the global market at $7 billion, North America at $3.5 billion, and of that $3.5 billion, $2 billion of it was addressable by Cogent with its inner city network. The second way we got comfortable in our market share expectations was bottoms up. We went to our 300 net-centric salespeople. They covered the entire universe of wavelength buyers and asked them to build a customer-by-customer projection of what they thought they could do over the next five years.
Those two different views of the market triangulated to a $500 million business for Cogent out of a $2 billion market, equating to a market share in wavelengths in North America equivalent to our global transit market share.
The competitive environment is changing. It's static. How would you kind of describe the competitive set?
It has changed over time. It was initially dominated by AT&T, then AT&T and MCI. Those two players withdrew from the long-haul market, and companies like Global Crossing, Williams, Broadwing, level three dominated that market. Today, they are consolidated into Lumen. Zayo has emerged as the second largest player in the market, and Cogent today has about 1.5% market share as a new entrant and really a unique network that differentiates us from the right of way that's shared by the others. We have five pillars of our competitive advantage: more coverage, more data centers, faster install, unique routes, higher reliability, and lower pricing. It is those five competitive advantages that will allow us to get market share against those legacy providers.
Turning to asset monetization, what's the update on data center sales for the two kind of initially discussed assets as well as the rest of the portfolio?
When we originally acquired Sprint, we acknowledged that there were 482 buildings owned fee simple. They contained 1.9 million sq ft and 230 MW of power. These were all telephone facilities. None of them were data centers. Our initial plan was to take 45 of those facilities, spend virtually no capital, and put in a low-density 1 megawatt, 10,000 sq ft retail co-location facility. After about a year and a half of owning the asset, we realized that there was latent value in the power that was coming to these facilities. They could be repurposed as true data centers. We went out to counterparties, surveyed the market, got a list of features that those potential acquirers would want, and then decided to invest $100 million of capital to repurpose these facilities on a much broader scale.
We ended up between June of 2024 and June of 2025 completing that project. We have 24 large facilities with 1 million sq ft and 109 MW of power in them. We began marketing them. As of this most recent earnings call, we announced we had entered into a letter of intent to sell two of the facilities. We are negotiating multiple other letters of intent and eventually hope to divest of the entire footprint. We think we will raise significant capital. The two facilities that are under LOI represent $144 million of proceeds. That's far in excess of the $100 million that we have spent converting them, leaving the other 22 facilities all to generate incremental profit as we dispose of them.
Based on the location of those data centers, the amount of capacity, power density, other attributes, what are the best use cases for portions or the entirety of that portfolio?
Yeah, I think this portfolio is not well suited for AI training. The power densities are too low. The sizes are too small. I think there are three suitable use cases: retail colocation, so more a traditional corporate data center type application; two, high-density cross-connect inter-networking activities; and then finally, because of the geographic dispersion of this footprint, they are well suited for inference. I think the counterparties that are in negotiation are intending to use these facilities for some combination of all three of those applications.
Anything more on the monetization theme more broadly, maybe non-data center?
You know, we have a couple of other assets that are surplus to Cogent's operating business. We have excess IPv4 address space. Rather than sell it, we have been in the process of entering some wholesale leasing agreements. Today, we generate about $65 million of revenue off of IPv4 leasing. That's up from $20 million four years ago, and we see that business continuing to grow for us. We may elect to divest of some of the addresses. We have nearly 38 million addresses, making us the third largest holder in the world. Finally, we have excess fiber in our network, and we have selectively sold a few routes. We may continue to do that. I think there is a surplus asset monetization story, but it is not a long-term driver of value. The key driver of value is building an operating recurring revenue business.
The use of proceeds from data center and other asset monetizations is?
I think it's a combination of strengthening our balance sheet, delivering on a net basis, potentially buying back shares, and then ultimately expediting the window in which we should be able to resume a much larger dividend payout.
Maybe getting to margins. What's keeping margins from returning to pre-Sprint levels as legacy Sprint and off-net kind of get de-emphasized as part of the mix?
I think it's threefold. One, we acquired a business that was 93% off-net versus Cogent's business, which was only 25% off-net. We have moved traffic on net. We have groomed away some locations that could not be brought on net, and we've converted all of the access customers to fiber. The second point is we had a number of non-core products that we acquired, many of which were gross margin negative. We went from an annual run rate of over $60 million in those non-core products down to a run rate now of about $5 million in those products, and there is still a little more grooming to be done. Third, we've had general costs that we've been able to take out of the Sprint business and improve our margins. To just remind investors, Cogent went public in 2005 with 0% margins.
Over the next 17 years, we grew those margins at a compounded rate of increase of 220 basis points a year and got up to about a 40.5% EBITDA margin. When we acquired Sprint, we acquired a business with negative 80% EBITDA margins. That business was 40% the size of the combined company. It was a meaningful drag on margins. Our EBITDA declined from a quarterly run rate of about $65 million to about $4 million. We have been able to rebuild that margin up to $48.3 million last quarter. In nine quarters, we've effectively grown the underlying EBITDA $5 million a quarter, mostly through cost cutting. To help us offset that, we received a stream of payments from T-Mobile of $700 million. Today, we're still receiving those payments at the rate of $25 million a quarter, and that will continue through Q1 of 2028.
As a result, our reported EBITDA of $73 million for the quarter was almost identical to where Cogent would have been if it had not acquired this business. Going forward, the growth in on-net services, whether it be IP-based services, address leasing, and maybe most importantly, wavelengths, will allow us to continue to grow our EBITDA. We project going forward returning to a 6%-8% top-line growth rate on a combined basis. That is below where Cogent was pre-acquisition, and we expect to have at least 200 basis points a year of margin expansion. We anticipate that the EBITDA margins that we had pre-acquisition without any subsidy payments of roughly 40% will be achieved again through the margin expansion and the roll-off of the subsidy payments.
Last question from me, and maybe the last question of the fireside, but what's the plan to tackle upcoming debt maturities? You've got two notes due in June of 2027.
We have a great deal of flexibility. We're still 18 months away from those maturities. We have about $400 million of incremental capacity in our secured borrowing capacity. Today, we have only one issuance out. Two, we have incremental capacity in our ABS and incremental borrowing there. Three, we may tap the high-yield market for unsecured debt. Fourth, we may look at doing an asset-backed securitization of our fiber assets. We have not done that, and that's been a very popular theme in our space. Finally, we can look at raising capital at the holding company level, most likely through a convert, but probably with equity at these points that would not be feasible. So I think we have a lot of optionality.
We have a year and a half runway, and we have adequate cash on the balance sheet, coupled with the slowdown in the outflow of cash due to the reduction in the dividend.
We have time for one question, please.
Yes, sir.
Hi, David. Quick question. You sold, even a letter of intent to sell, two data centers for $140 million. What do you expect the two closings to actually close? Expect that? Second to that, you mentioned another 22 data centers that may be useful for inference. What do you think the timeline is on adoption?
The party that we've entered into an agreement with is someone that we feel is quite credible and has the cash to close. They have completed their first round of diligence and posted a deposit. That deposit goes up as they convert to a binding agreement and then goes up even further once the diligence items are completed. I would anticipate closing sometime in Q1. That's what both parties are targeting. To the second question around the other remaining facilities, you know, we're in conversations with multiple parties, many of which have provided LOIs, which we have not countersigned due to their inability to show us proof that they can get to closing.
I think that many of these facilities, because they were literally peppered across the country as class four switch sites for long-distance termination, are distributed across the population of the United States, typically along rail tracks, typically in an industrial park 10- 15 miles from downtown. I think many of these facilities will be ideally suited for the lower density computing necessary for inference. So, you know, since we have no core competency in that area, I think it would be best for Cogent to allow others to do it, and I'm confident that we'll be getting some additional offers from parties that can move forward.
Thank you for your time.
Hey, thank you very much, John.