Let's get started. Good morning. Welcome to day one of TD Cowen's 53rd annual TMT Conference. My name is Greg Williams. I cover cable, wireless, and telco here at TD Cowen. I'm joined for this session by Dave Schaeffer, CEO of Cogent. So, Dave, thank you for joining us.
Thank you, Greg. Thanks for hosting me. Thanks, TD Cowen, for a great venue. Most importantly, thank investors for getting up early and being in the room.
Sure. We want to start off with probably the biggest topic since the print was the dividend growth cut. You know, when I look back at Cogent, you started a regular dividend, and it was a 2-cent growth every quarter. Then it went down to 1 penny growth every quarter. That seemed to be the permanent cadence. Now you are up to a half-penny growth. Is that a permanent cadence like we saw in the past? Or do you feel like when you get back down to deleveraging, possibly sell some data centers, you bring that back up to a penny growth? How do you think about that?
Yeah. Cogent has returned about $1.7 billion to investors. We initially did that exclusively through buybacks. We bought back about 10.8 million of our outstanding shares, or a little over 20% of our float. We implemented a dividend in 2010 and have now consecutively grown that dividend for 52 sequential quarters. Actually, the initial growth rate in a dividend was a penny a share. It then was increased to 2 cents and then ultimately 2 and a half. We reduced that during COVID to a penny a share. In recognition of the increase in leverage that we have had due to the integration of the Sprint assets, we have reduced that growth rate to half a cent a share. Our leverage is beyond kind of the comfort zone that we have targeted. That is because of the step down in subsidy payments that we receive from T-Mobile.
We anticipate delivering relatively quickly. On an LQA basis, our leverage peaked in Q3 of 2024. On an LTM basis, it will peak in Q3 of 2025 and then begin down. We may further delever through the divestiture of some non-core assets, although that is not critical to our delevering. I think in use of free cash from operations, the first objective is to bring leverage back down to opportunistically buy back shares. Last quarter, we bought back about $10 million worth of stock and will continue to try to take advantage of market fluctuations. Finally, I think we will be in a position to re-accelerate the growth rate in the dividend. I view the step down to a half-cent a share growth as temporary during this period of delevering.
Got it. You mentioned share buybacks. With the stock down here at this point, how do you think about capital allocation?
You know, we think that supplementing the dividend with buybacks does make sense. We are also focused on leverage. You know, we have chosen not to further slow down the dividend growth rate and take a more balanced approach. You know, I do think we are committed to continuing to return capital to our equity holders.
Got it. Part of the deleveraging could be the sale of data centers, which you've messaged for quite some time. I feel like the data center sales timing has shifted over time. You know, originally the impression was you'd have 23 data centers ready by the June timeframe. Recently, that tone seemed to have shifted. Now it's 24 facilities, not 23. You said on the call it could be closer to six months from now. The way I look at that, I mean, that could take us to November. What's changed where we thought we'd see transactions maybe this summer to maybe now this fall?
Let me step back to the announcement of the acquisition of Sprint. Prior to acquiring Sprint, Cogent operated 55 data centers, two of which were fee simple owned and 53 of those were in leaseholds. Those facilities represented 634,000 sq ft and approximately 69 MW of power. When we announced the acquisition of Sprint, there were 482 fee simple owned buildings comprising 1.9 million sq ft and 230 MW of power. Our initial plan was to take 45 of those 482 facilities, put a 10,000 sq ft data center with lower power density in a portion of those buildings, and leave the remainder of that asset fallow. In early 2024, it became clear to us due to AI that the power that we had was a scarce resource and we could monetize that. In April of 2024, we tested the market by identifying 23 initial data centers.
You are correct, we've added one more based on customer feedback. We went out initially to 115 counterparties. That has since grown to about 150 counterparties that we've talked to. We conducted tours of the facilities and analyzed the feedback that we got from customers. In June of 2024, we made the decision to spend $100 million over the next 12 months implementing a number of the recommendations that customers wanted to see if they would either buy or lease these facilities on a wholesale basis. We still have about 50 active parties talking to us. Four of the parties that have given us LOIs have progressed to contract negotiation. We anticipate all of the 24 facilities being in a marketable state with the changes that we are implementing as a result of customer feedback completed by the end of June.
You know, it is difficult for us to say how quickly we will move from contract negotiation to a binding agreement with a firm deposit. You know, of the four parties that we're speaking to today in contract discussions, I think three of the four are for purchase. One is for lease. One is for the entire footprint. Two are for multiple facilities, but not the entire footprint. One is just for a single facility. We are continuing to accept LOIs. There are tours going on almost on a daily basis. We are just looking to maximize value. You know, we are not experienced in transacting on these. You know, it is difficult for us to predict when an LOI will actually go to a binding contract.
In my view, you know, while these are still in discussion until there's an earnest money deposit off with no contingencies, they're just discussions.
Got it. You sort of answered my next question of narrowing down the 150 participants down to the four LOIs. It sounds like you're entertaining any and all options, whether it's the entire portfolio, a portion of the portfolio, lease some, sell some. You're still in that sort of exploratory phase and seeing the LOIs come in. Is that right?
Yeah. There have actually been more LOIs. It has been probably close to 20 LOIs. Of those, four have progressed where there is enough substance and maybe enough congruency in our valuation thoughts that it made sense to start talking about a binding contract. There are other parties that are still, I think, in price discovery and are not yet at a hurdle rate that we feel is acceptable.
Okay. And, you know, it would be a taxable event if you sold the data centers versus if you leased them. It might be more favorable from a tax perspective. Is that? You know, I think you'd try to want to avoid paying taxes, but I'm sure you're in an NPV on both leasing versus selling. But, you know, I guess the question is, help us with your thinking about whether you want to sell them versus lease them. And, you know, could it be 70-30 sale, lease, or 50-50?
You know, I think what we want to do is maximize the value. You know, there's clearly an advantage to the certainty of a sale, and there is a tax consequence associated with that. While Cogent has not been a significant cash taxpayer to date and still has a significant number of NOLs in Europe, its U.S. NOLs are almost completely depleted. So if we received a significant windfall from the sales, absent the capital we've invested in modernizing the facilities, the remainder would be viewed as a capital gain and would be taxed as such. You know, if we did lease them, we would be very focused on the credit quality of the tenant. And we would probably look to securitize that stream of rental income as a way of raising cash.
Because ultimately, the goal is to allow leverage on the balance sheet to come down and then accelerate the return of capital to equity.
Right. And speaking of leverage and your desire for leverage to come down, there's a fear out there on valuation that you might not be able to fetch the $10 million per MW price given the pressing need to sell these things. Can you help us, you know, ease those concerns about potentially selling lower than the message $10 million per MW?
We have never sold a data center. As we tried to come up with a valuation of those data centers, we looked at comparable public trades and data that was available. What we saw was a lease rate of about $1.4 million per MW per year on a triple net basis and sale prices of about $17 million a MW. Now, there are significant differences in these facilities. These facilities were not purpose-built. Their power densities are lower, and they are not as large. Now, on the positive side, they're very distributed. It does have value to some purchasers having a wide distributed footprint. $10 million sale price or $1 million as a straw man have had offers ranging far below that to actually full ask price. Again, we have to vet the creditworthiness of these counters.
You know, until there's a deal done, I can't tell you what we're going to get for them. I can tell you that they are non-core to Cogent. As we look at our ability to fill up the data centers, we have increased Cogent's data center footprint to approximately 180 facilities with about 210 MW of power across that footprint, a total of about 1.92 million sq ft. These are both the leased and owned facilities, both core and edge locations. You know, we've carved out 109 MW in the 24 facilities and 1 million sq ft as surplus or something that we cannot expect to fill up. The remainder of that footprint, I think, fits with the general Cogent data center model of leasing out one or two racks at a time to retail customers.
Got it. Wanted to switch gears and talk about the Waves business.
Okay.
It's obviously very top of mind. Cogent's off to a slow start on the Waves business. You know, this time last year we were talking about it. You are looking to at least increase the backlog to 10,000 orders and then install 4%-5% of the backlog a month, which is 400-500 circuits per month. Like this time last year, you're saying by January you'd have 500 installs a month. You know, help us with what really happened. I mean, you mentioned that your sales funnel ended up being stale in the fourth quarter and cleaning that up. Were you busy cleaning up a stale funnel? You know, what really happened between last time we spoke to on the stage today on the Waves business?
We do have the capability, meaning we have the field services and provisioning team to be able to install 500 orders a month. Just to go back and refresh memory, when we announced the Sprint transaction, we said it would take between a year and a half and two years to repurpose the Sprint network to sell wavelengths. We closed that transaction in May of 2023. We set a target of having 800 data centers wave-enabled where we could provision a wave on demand with an any-to-any data center connectivity by the end of 2024. We ended the year with 802 data centers. In that intervening period, we had manually provisioned approximately 1,000 waves. Those waves that we did provision were done much the way other providers do it, as opposed to a more automated process. We had built a funnel.
The majority of that funnel sat and waited, and we were unable to deliver. We did deliver 1,000, but probably 80% of that funnel eventually fell out. Some of that funnel carried over. As we entered 2025, we had built a newer funnel in the fourth quarter of 2024 and continued to build that funnel in Q1 of 2025 and continue beyond that. We know that customers are accustomed to a three to four-month installation window with other vendors and have been skeptical around Cogent's ability to do this in ultimately two weeks. Today, we are doing it in 30 days. What we saw in the first quarter is we actually grew our install base by roughly 18% sequentially or better than 200 waves. Almost all of those waves were installed at the end of the quarter because customers were not ready.
Not because Cogent could not install them, but we also have to wait for customers to be able to accept those services. Now, with our IP-based services, we have a 20-year experience, and we know that about 4% of our on-net funnel installs every month, and about 50% of that funnel ultimately installs. We allow those customers in transit services three opportunities to extend their install. At the end of that, we then force bill them based on a firm order date. We have not done that yet for Waves, in large part because Cogent did not have the ability to deliver. It is unfair to go back to the customers and say, "We're ready now. You have to take them whether you're ready or not." I think over the next several months, we will continue to build credibility with customers.
We will continue to build a funnel, and we will continue to increase the pace at which we are installing. The ultimate goal is to get the orders to install ratably over the quarter and to meet our 500 wave capability. If we, in fact, see that we are getting greater demand than that, our expectation is we can increase that service delivery capability. Today, we install about 3,000 IP ports a month, and it did not make sense for us to add more than about 500 waves a month of service delivery capability until we've proven that out. You know, I think in terms of pacing, you know, we've been pleasantly surprised by the aggregate size of the market. AI is an incremental use case that we were not anticipating when the deal was announced.
I think another positive surprise to us is how much customers dislike their current supplier choices.
Can you talk about the cadence as we almost approach June here? You mentioned that the end of last year saw an acceleration. Did that cadence continue?
Yeah. I think we are not yet at our full capability, but we've seen better improvement in units installed in both April and so far in May, each month doing better than the pacing we achieved in the first quarter.
Got it. I'm a little confused. If you're provisioning waves in days, 30 days and getting down to two weeks, you know, why build such a big backlog of 10,000 orders? It feels like to me, you know, you order it, by the time you get another order in the backlog, you're already provisioning and installing.
If customer behavior acted that way, we would not need a backlog of that size, and we would be immediately adding to our service delivery capabilities. Again, I'm drawing experience from our transit business where we know, one, the install is even quicker than our wave target, and two, customers generally do avail themselves of multiple extensions from the firm order date. In a transit order, we take the order, we commit to a date of 17 days. We actually average nine days to install. The majority of customers do not accept those orders initially on the install date because the data center is not able to give them the cross-connect. They may not have their equipment, but they've got an expectation that Cogent will install quickly, and they typically push those out in a few week increments and eventually take that order in 30-60 days.
In the wavelength market, customers are accustomed to 50% of their orders never installing with the other vendors. For those orders that do install, it takes three to four months. Cogent, as a new entrant to that market, really is coming in with some pretty bold claims saying, "We're going to do this and deliver in 30 days," and the customers say, "I'll believe it when I see it.
Right. So you're not the bottleneck then if there's?
We are not the bottleneck, whether it be number of locations or speed to deliver. The only limitation that we currently have, and we've been clear about that, and it will be gone in the next probably 60 days, is in certain data centers, we were in 883, 100% of that footprint, we can deliver 10 and 100 gig waves in every direction. In about 80% of the facilities, we can deliver 400 gig waves in every direction. In a subset or 20% of the data centers, so call it roughly 150 or so data centers, there are certain directions where we do not have 400 gig capability. Since 400 gig represents less than 5% of the market, this is not an issue in terms of our ability to meet customer needs.
Right. Throughput's not the issue. Maybe talk about the demand set. You said that AI is an incremental use case. What are the other typical use cases? What are the type of customers? What are the order types? It sounds like it's not 400 gig wave yet. Maybe you can talk to the AI opportunity. You know, is that probably more in the inference phase for you versus, you know, or sort of the training phase, if you will?
Today, the primary use cases are regional networks linking their networks together, international carriers extending their networks. Examples would be someone like a Frontier or a Metronet or a Ziply that have islands of traffic and want that traffic linked together into a homogeneous network and then connected back to major aggregation points. A second extension would be for a subsea or international carrier like a Telefónica, a Telecom Italia, a British Telecom, an Orange that would want to extend their network across North America. The second major use case is for content delivery and delivery of replicated content to multiple sites. That would include companies like a Meta and Amazon, a Microsoft, a Netflix.
Caching content.
Caching content, but it would also include companies like a Fastly or Akamai, hosting companies like a Rackspace or an OVH, and a fairly fragmented market of smaller data center operators and hosters. The incremental use case for wavelengths is actually for AI training today, much more than inference. That is extending the data center where content exists to the data center where the training is occurring. Typically, the training facilities are not co-resident where the data is stored. Today, there is over 900 zettabytes of stored data that have been collected over the internet in the past 30 years. It is that data that is used for building large language models. The training data centers tend to be large, very power-intensive, and remote. Wavelengths are used to move that data back and forth.
While a wavelength is about two and a half times more expensive than using transit to move that data, because over 50% of the cost of that training facility is in the GPUs, the training data center operator wants those GPUs to operate at full efficiency without buffering. For that reason, they purchase wavelengths. Dark fiber is also used for that purpose. It tends to be even more expensive than wavelengths, but there are certain customers who value that control and that requirement to go to very remote locations. For the last part of your question around inference, we're just beginning that. I think inference will do two things.
It will drive incremental internet usage, and it will occur across a much more distributed set of facilities that will probably use wavelengths to distribute certain large language model outputs to those facilities, and then also in those facilities have much more narrow sets of data in which the neural networks will be applied to come up with specific inferences that'll be delivered over the internet.
Right. You mentioned that customers are disliking their current wave provider, and you mentioned or alluded to the fact that provisioning for some of them could take three months, six months, or even never. You can do it in 30 days. However, when we speak to some of these competitors, you know, Zayo and Lumen own half the waves market. They are investing heavily in waves. You know, we have Lumen speaking later today and speaking about waves. I presume they are cutting down their provisioning time. They are pushing that service as well. You still have unique routes. You can price it better, but do you feel the competition can step up?
You know, I would hope they would. You know, I'm again, use our experience in the transit market. You know, when we entered the market, there were 200 providers, and the average time to deliver a high-capacity IP circuit was 90 days. Within a decade, the market has shrunk to 12 providers, and Cogent's provisioning time was at nine days when the market average was 30. You know, our competitors operate heterogeneous networks and have a more limited set of endpoints. I know one of our competitors, for the first time in their history, put out a press release and said they had 47 data centers where they could deliver a wavelength within 30 days. They had never made that claim before. You know, maybe that was in response to us. Maybe their business is just changing.
What I do know, though, is when you look at those 47 data centers, there were actually only 22 data centers with extended cross-connects to the other 25. So it was a much more limited set than their press release led you to believe.
Interesting. And then just on the revenue targets, my last question for the Waves. You know, when you merged with Sprint, you mentioned you'd reach a $300 million run rate in revenue by 2026. That's clearly off the table now with these delays. So maybe is there an updated target of when you hit that run rate? And then the longer-term run rate was $500 million in revenue run rate by 2028. Is that also subsequently delayed?
No, we still firmly believe that by mid-2028, we will be doing an annualized run rate of wave revenue of $500 million. It is true that the wave market is more diffused than we had originally expected. We've actually increased the number of data centers that we are targeting. We were at 883 at the end of last quarter, up from the 802 at the beginning of the year. That number will increase. That gives us over 10 to the 21st power number of wave path combinations. We also have reiterated the fact that by the end of this year, we will have some intermediate proof on our wave cadence and should be generating around $20-$25 million in quarterly revenue, getting us close to a $100 million annualized rate by fourth quarter of 2025.
Got it. With that, David, we're just about out of time. Thank you very much.
Thank you very much, Greg. Take care.