Good morning. I am Jon Greenberg, Managing Director in Investment Banking at Morgan Stanley, here with Dave Schaeffer from Cogent, CEO, President, Chairman, and Founder, I believe. Quick disclosures for important disclosures. Please see the Morgan Stanley Research Disclosure website at www.morganstanley.com/researchdisclosures. If you have any questions, please reach out to your Morgan Stanley sales representative. Good morning, Dave.
Good morning, Jon. Thank you for hosting me. I'd like to thank the investors for taking time out and getting up early. Always thank Morgan Stanley for a great venue. I think you get the first. This is the first time I've been interviewed by the son of someone who interviewed me almost 25 years, 28 years ago, when your dad was covering research.
We were just waiting for me to become of age, right?
Right.
Welcome back.
I'm showing my age, one of two things here. But let's jump into it.
Let's get into it. 2024 felt like the year of artificial intelligence, or at least its big party coming into our industry. Maybe starting there, how do you think AI is going to impact Cogent? Maybe starting on the customer perspective, but then getting into the OpEx perspective.
You know, I think AI is, like many other buzz phrases, often used and seldom fully understood. I think it does make sense to take a moment and do a little bit of definitional work around it. I think it's one of the technologies that is transformational, and it is just beginning. It is very poorly named, and it's not what I think most investors think of when they think of artificial intelligence. Artificial intelligence, no matter which large language model you use, kind of always functions the same way, where it takes large data sets, poses a question to those data sets, and then discerns the pattern that the program uses to come up with the correct answer. It does that repeatedly and eventually takes those patterns and projects them into new back sets.
The primary application of that inference will be the transformation of the way software is produced and impacts people. Initially, software for the first 30 years was all about computational efficiency. Software then for the next 50 years was all about process codification and improvement. What we are doing now is substituting capital and machines for people, coming up with very agile code that can be applied in very specific situations, very customizable and very flexible. I think that is going to transform the world. For Cogent, we benefit three ways from AI. First of all, the raw material to create AI is information collected over the internet. In the internet's 30-year history, it has stored about 800 ZB of data, and it is adding to that pool now at the rate of about 200 ZB a year.
The information that is transmitted and collected on the internet now has value when it did not. That is a positive for Cogent as the world's largest internet carrier. We carry approximately 25% of the world's traffic. The second benefit is to our optical transport business or wavelength business. Typically, the large language model training occurs at locations that are distant from where the data resides, simply because of the power intensity of those processes. As a result, you need to move the bits between their storage point and the location where that training is happening. The best way to do that is with either wavelengths or dark fiber. While the internet is cheaper, because the internet is latency indeterminate, it means your GPU utilization would decline if you were moving that traffic via the internet.
It's actually a positive driver for a relatively new part of Cogent's business. Then the third leg will be the ultimate inference phase, taking those patterns and projecting them into smaller data sets for very customized workflows and processes. That will all be done over the internet. I think it will drive a new wave of internet traffic growth, much like video or social media had driven previous waves of growth.
What do you think the time horizon is on that? Because there's a ton of momentum right now, it feels like, on the power component of this in terms of these large data center campuses for learning, which is obviously going to create its own sense of fiber demand. Then the inference stage, which I think is more where you were focused on and pushing that more toward the edge, feels like that's still many years to come.
I think we're at the beginning of the process, not even out of the first inning. I think there is a rapid evolving process of improving large language models. There was one thesis that once you trained through the full 700 ZB or 800 ZB of data, you were done. The reality is each model comes up with a different neural network. I think that process is going to be ongoing and continuous as those new large language models are evolving at the rate of about a new version every four months. Secondly, we have only started using the inference tools in very narrow subsets like ChatGPT or Grok. I think there's going to be much more widespread embedding of that inference in every process and service that we use. I think this could actually be an existential threat to the traditional software industry.
Just as the internet destroyed the telecom industry, I think it's likely that AI will do the same to traditional software. There are two large questions that have yet to be answered. The first of those is, is the economic value of the output greater than the cost of producing the inputs for that output? I think the answer right now is unclear, but probably no. I think the cost will come down quickly. The second and probably more important question is, is there a business model to capture that economic value? I think the internet is a great example of how that value gets distributed to end users, and it's very difficult for service providers to figure out how to monetize it. I think the same thing is going to happen with AI.
We have talked a lot about the demand side in terms of your business and your ability to generate revenue off of this. On the cost side, you spoke a lot on your earnings call about your sales to install cycle and the ability to provision customers quite quickly. You think AI is supportive of that longer term and your ability to keep addressing the customer?
AI today is a small driver of our total revenue streams. 89% of our revenues come from selling internet-based services. Now, some of those services do result in the data collection for AI, but it's commingled with many other services and processes. For example, if you're using Uber to get a car, you're using that particular application, riding over the internet, using Cogent's backbone. At the same time, they're collecting data for their next generation of AI and training. The application and the data collection are commingled. We have been growing our internet business consistently and gaining market share due to the network optimization that we did 25 years ago and have continued to perfect to result in the lowest cost interface routed bit mile. I know that sounds like a mouthful, but that's really what someone is using when they use the internet.
We had the opportunity to repurpose the former Sprint Voice Network, and 2024 was a year of transformation where we completed the repurposing of that network to sell optical transport or wavelength services. These are point-to-point committed services that are typically sold in three speeds: 10 gig, 100 gig, and 400 gig. We can offer those now in more locations across North America than any other provider. We have unique routes in 90% of the cases. We can produce very accurate GIS mapping for those routes. We can provision those routes much more quickly. Because we have no real cost basis in the Sprint network, we can price aggressively. With the acceleration of AI training and the nearly $2 trillion that has been committed to capital programs for that over the next five years, the use of this optical transport will only increase exponentially.
I think we're in a position to capture a large portion of that growth.
Great. Let's get into business performance. And 2024 was a pretty transformational year. The Sprint integration is largely complete at this point. Maybe can you just talk about 2024 to 2025 and EBITDA and how we should think about the different levers between organic growth versus operating leverage versus the remainder of the synergies to extract and then the wind down of the subsidy payments?
Yeah, I think I almost have to go back to 2022 to fully answer that question, Jon. I'm not trying not to answer your question. Cogent was a standalone company that had not done any acquisitions for 18 years. In our early formative period, we did a number of acquisitions, stopped in December of 2004. For the next 18 years, we executed on integrating those assets into the world's largest IP network and became the largest carrier of internet traffic. Our business was doing about $600 million in revenue, $260 million of EBITDA, with top line growing in mid-single digits. It actually decelerated because of the pandemic, and our margin expansion had slowed to only 100 basis points a year, sequentially from our average of 200 basis points a year. We then had the opportunity to acquire Sprint's business from T-Mobile.
This is the original long-distance wireline business that originally was being sold to MCI in 2002 for $129 billion and blocked by the Justice Department. Ultimately, T-Mobile paid us $700 million to take that business. When we took that business, we really acquired two very different things. We acquired a customer base of large enterprise companies that were buying a myriad of products, of which VPN services based on MPLS and DIA were the majority, but a lot of managed services. That business was burning $1 million a day, negative cash flow. We also acquired a physical network that was built at a capital cost of $20.5 billion. We were paid $700 million to take that business.
Our revenue stepped up to $1 billion in revenue while $940 million in 2023, and our EBITDA stepped up to $352 million, in large part because of the subsidy payments that we were getting from T-Mobile. Over the next 18 months, we worked feverishly to reduce the burn on that acquired enterprise base. By the time we held our earnings call in February of 2025, we had gotten that negative $300 million of burn down to zero, so effectively breakeven. That was through product rationalization, moving traffic from off-net to on-net where practical, and eliminating non-core products and services that were uneconomical. We still have some work to do. That business will eventually be about a 20% positive margin business, but with little or no growth associated with it, based on the nature of the customers and the long-term product substitution of internet access for VPN services.
The real upside came in our ability to take that long-distance network, 19,000 route miles of intercity fiber, 1,200 route miles of metropolitan fiber, and 482 data centers that comprise 230 MW of power and 1.9 million sq ft, and repurpose those assets. We've been working feverishly to turn that into an optical transport network. We connected it to our metro network. Today, we're in 880 data centers across North America, where we can provide a wavelength from any data center to any data center and do that with a 30-day provisioning window. We are repurposing those physical buildings. We've identified 104 of those buildings that we've converted into data centers. Finally, we identified 23 of the largest facilities where we have 100 MW of excess power and 1 million sq ft that we're trying to market and sell as data centers.
We have been working diligently on that process. The network reconfiguration is complete. The data center configuration project is substantially complete. By the end of second quarter, all of those data centers will be converted. We are in the process of marketing those assets and trying to see if they fit for AI or other applications.
Where do you see CapEx settling out once you're through with the data center repositionment?
We spent about $150 million of extraordinary CapEx, about $50 million on the network integration and about $100 million on the data center repurposing. On a steady state, Cogent's CapEx should be around $100 million a year. That is for direct capital expenditures, both for maintenance and modest network expansion. We will also spend about $40 million in capital principal lease repayments. These are leases on our balance sheet that have about $500 million of value, and we amortize those or pay those off at about a rate of $40 million a year. Probably another $10 million or so for new capital lease inclusions for expansion opportunities. All out the door, about $150 million of capital in those three different places on the cash flow statement. For EBITDA, our EBITDA was flat year-over-year.
We went from $352 in 2023. In 2024, we were $348.4. We will have similar EBITDA this year. Our revenues will also be flat at roughly $1 billion. As these new growth initiatives kick in, we shall resume growth and margin expansion. We have laid out a plan to be at $1.5 billion in revenue, $500 million in EBITDA by mid-year 2028. That is driven heavily by our growth in the wavelength market. The wavelength addressable market globally is about $7 billion. It is approximately $3.5 billion in North America, split into a $1.5 billion metro wavelength market and about a $2 billion inner city market, which is the primary market we are focused on. We believe because of the superiority of the network, the uniqueness of the routes, we can capture a 25% share, kind of mirroring our success on the IP market.
You touched on the uniqueness of the routes, but the waves market is a competitive market. There are other operators out there that have substantial network that's targeting that customer segment. Can you talk about what your competitive advantage is in terms of targeting those customers?
The suppliers of wavelength have changed over time. Originally, the two suppliers were AT&T and MCI. Back then, they withdrew from the market 20 years ago. Level 3, Global Crossing, Williams, many of the companies that make up Lumen began to dominate that market. Zayo, as an aggregator, became the second largest player in that market. All the companies that focus on the wavelength market sell a myriad of other products. They have not architected a network that is optimized for wavelength service. They have fewer endpoints available. Each of those major competitors has about 350 endpoints as compared to our 880. Two, the way in which they provision a wave is custom on each wave. It comes out of a network that is already in service delivering other products, typically requiring six field visits and about three months to provision.
We stepped back and took a very different approach. We had a dark fiber network that was built 40 years ago, had been sitting dormant for a decade, and we could architect it any way we want. What we did was built a network that can provision a wave more agile and more ubiquitously. It is virtually impossible to pre-provision the capacity because of the number of permutations between facilities. With 880 facilities, you have over 10 to the 2000th power number of possible combinations of wavelength paths. No way you can build those in advance, but you need to have a foundation that can deploy those waves quickly. In building a wavelength optimized network, you have two very different design goals. The first goal is to use your long-haul spectrum as efficiently as possible.
The second is to minimize the number of optical to electrical to optical conversion points that are incremental capital expense and incremental points of failure. We took those two design goals and built a network that achieves the optimal deployment in a way that our competitors do not. We understand in every business we operate, we have competitors. In the IP transit business, there were 200 global competitors, all larger than Cogent. Over a 25-year period, we caught our way to the number one position by having a more efficient network, a lower price point, less capital efficiency, and an ability to provision quickly. We are taking those exact same learning lessons that allowed us to become the primary provider to hyperscalers around the world and project that into our wavelength market. While we expect competition, we think that our solution is superior in many dimensions.
You have an increasingly diverse set of assets that you can monetize these days. We've talked a bit about the fiber. You also have the IPv4 assets, and then, of course, the data centers. Maybe shifting to IPv4 for a moment. You've monetized them in two ways. One is price increases, and another is the IPv4 securitization. Can you just talk about the future of that collection of assets, how you see the pricing trends going, and is there more securitization on the come?
Yes. There are three assets that do not really show up on Cogent's balance sheet: the dark fiber that we have, the buildings that we own that we acquired for a dollar that we are converting to data centers, and IP addresses that we have acquired over the years. We have 38 million IPv4 addresses. For a general investor to go, "What is that? How does that have economic value?" Morgan Stanley did help us securitize that through an ABS. In that first meeting, we had to explain to the bankers why this had value and why it was an asset that warranted being able to raise capital against it. When the internet was first architected, there were three foundational technologies: TCP/IP, which is how two computers talk to one another; BGP, how two networks communicate; and the third, IPv4 as a unique addressing scheme.
At the time, it was thought there would be plenty of addresses with 4.3 billion addresses available on IPv4. It is 2 to the 32nd power addresses and arranged in hexadecimal sequences that are unique to each location on the internet. As the internet became very successful, it became obvious that was not going to scale. We needed a new technology, IPv6, which has 2 to the 128th power addresses. It alleviates the constraint, but it has never been widely adopted. As a result, the addresses that we own, 38 million, it is the third largest assemblage in the world, has economic value. We began leasing those out. Markets developed to buy and sell those addresses. In fact, Amazon and Microsoft have spent collectively nearly $5 billion buying addresses over a decade. We chose not to sell, but rather lease.
We have had the ability to increase pricing as those addresses are a scarce resource. We took that cash flow and securitized it. When we initially did this, we had $32 million of annualized cash flow in March of 2024. By the end of the year, that cash flow was annualized up to $47 million. We are looking to do some additional securitization against that incremental cash flow.
How about the pricing trends that you're seeing?
Yeah. I'm going to discuss sale and leasing. On the sale side, there are two public exchanges for the purchase and sale of addresses. Prices have increased from the first trades in 2011 at $4 to a peak at the end of 2023 of $60 an address. Today, the trading range is between $48 and $52 an address and fairly stable. Low volumes, but lots of small trades. On the leasing side, we were unique. We were the only people leasing. Few cable companies had small leases to retail customers, but not an organized program. Amazon and Microsoft had been a rapacious buyer for 11 years and then began leasing at a price that was substantially higher than Cogent's lease. At the time, Cogent's average lease was $0.29 per address per month. Microsoft and Amazon both, coincidentally, launched a price point of $3.60 per address per month.
With that pricing umbrella, we were able to go into our base and raise prices. We have taken prices up by 40% from $0.29 to $0.49. We have continued to sell out incremental addresses. Today, we have leased out 13 million of our 38 million inventory. We are continuing to evaluate. Does it make sense to sell off some of that excess inventory, or does it make sense to just keep moving forward with additional leasing and perhaps additional securitizations?
For others out there who, like me, have heard a lot about IPv6, but it doesn't seem like there's much out there. It's more talk than action. Can you talk about the threats that exist to the extent that you see them on IPv6?
There are three challenges to IPv6 being widely adopted. The standard came into play in 1998. In 27 years, it has only gained 7% market share. The limitations are all equipment needs to be IPv6 compatible and can for the path of the packet travel. That is becoming more common. It is still not the case today, but probably over the next five years, I think that limitation will go away. The second limitation is probably the most important. That is when you use the internet, you want to reach all of the endpoints. If an endpoint is only visible on IPv4, IPv6 cannot see it. As a result, if you elect to use IPv6, you today only get 7% of the public internet, and 93% of it remains dark to you.
Given that choice, almost everyone says, "I'll pay an extra $0.50 a month to be able to see the whole internet." The third issue, which is the ultimate process of transition, is very problematic because it has to happen all at once. There is a fair amount of labor involved in going to each device and changing the number. It is kind of like the Y2K problem on steroids, where every device needs to be touched and upgraded to V6, but it all has to happen instantaneously. Because of the complexity, ubiquity, and importance of the internet, I do not think that is practical today. It is probably another couple of decades before we see widespread ubiquitous use of V6.
Another hidden asset on your balance sheet that you talked about, the data center central office portfolio. How do you think about monetizing those over time?
Yes. Some of those buildings are just too small or too remote a location to, I think, have great marketable value. We have converted 104 of them to data centers. 55 of them are smaller data centers. 49 are larger. In those 49 larger data centers, we identified 23 of them that have substantial surplus power beyond what we are going to need. There is actually 109 MW of existing in-place power today coming into those facilities with 1 million sq ft of space. Now, when we took over these facilities, they were full of old telephone switches, TDM voice switches. We initially thought we would just clear out a corner and run a small retail data center. Our thinking has evolved. We have gone in and depopulated these entire facilities.
We have tested and upgraded the fire suppression systems, the generators, the battery plants, the perimeter security system. We have converted these sites from - 48 DC power distribution to AC 120, which is what data centers, particularly those running AI, would require. That process is ongoing. It should be complete in about the next four months. With that, we've actually gone out to the market to over 140 counterparties, probably had about 50 in active discussions that are interested in them. Literally, tours are going on every day. We hope in the next month or so, we'll be in a position to call for offers on those facilities. Even though they won't be completely converted, they'll be substantially enough along that we can ferret out who's going to own them or lease them. We've offered these facilities on two different economic models.
You can either lease them on a triple net basis for $1 million a megawatt, or you can purchase them outright fee simple for $10 million a megawatt. Both of these pricings represent about a 40% discount to the current market price.
Dave referenced a tremendous number of numbers during this conversation. For those reading the transcript, no notes. You were referencing no notes.
That's okay. It's been doing this a long time. I can make one mistake. My 59 was 49. That's okay. I caught myself.
Thank you very much for coming to the conference. I think there's a lot of exciting things to happen with Cogent over the next year. We are hopeful to see you back here in a year.
Thank you very much, Jon.
Thanks, Dave.
Thank you all very much.