Hi, everybody. Tim Moran, the communications analyst here at Oppenheimer. My pleasure to have Dave Schaeffer, the CEO of Cogent. Sorry for the brief delay here. Dave, it's been an interesting year in many ways. We'd like to get down to the details, but I'd like to just maybe start at a really high level, if that's okay with you, and I know you've done a few of these here this week. We're hearing that the demand for AI infrastructure, for both networking and compute, is incredibly strong at the moment. There's a shortage of GPUs, shortage of compute, that AI tokens, I guess the utility that people are using them for, they're getting very good returns. So they want a lot more compute. There could be other drivers of this, obviously. But you're right in the thick of things.
You're talking to all the people that are basically building out this compute and networking. Can you give us a sense of what you're hearing and seeing out there for overall demand?
Yeah. Sure, Tim. First, thanks for inviting me to your conference. I'd like to thank investors for their time and Oppenheimer for this venue. Cogent is in the middle of that infrastructure revolution for two distinct reasons. One, we are the largest carrier of internet traffic in the world. AI is possible because of the internet; AI will be distributed over the internet, and it will be integrated into the internet. We have actually seen internet traffic growth begin to accelerate. In the most recent quarter, year-over-year growth accelerated from 12% in the first quarter to 16% in the second quarter. That acceleration in traffic growth follows a pattern of many other applications, whether it be streaming or gaming, where a particular application drives demand. We are now at the very beginning of agentic AI being embedded into everything. That will change the internet two ways.
It will create more traffic, and it will also change the directionality of traffic, as users will have to upload data to then be tokenized and run through LLM for inference. The second part of the AI revolution impacts our wavelength business. That is optical transport. Oftentimes, the development of an LLM is in a different location than where the data is being stored. Because the GPU represents about 60% of the total capital cost in that data center, you want to make sure that those GPUs are being utilized 24 hours a day, seven days a week at peak efficiency. While the internet is cheaper on a per bit-mile basis, it inherently has no defined latency and will require the processor to buffer those bits. The premium that is charged for wavelengths of about two and a half times as much per bit-mile is easily justified.
While there has been a historic demand for private line or wavelength services from international carriers, regional carriers, content distributors, cloud providers, we are now seeing a surge in that demand from AI training, either by hyperscalers or by neoclouds. You are correct, Tim; there has been $1 trillion of capital deployed to date. There is another $6 trillion that has been announced for deployment. Much of this demand is still being developed. There is an acute shortage of power. There is an acute shortage of data center space. There is a shortage of processing. At the end of the day, what you are trying to do is produce the most number of teraflops at the lowest amount of power input and the lowest amount of cost.
We are seeing that proliferate to a lot of new data center construction, which then will have waves to link them back to historical storage sites. I think for all aspects of Cogent's business, these trends are a positive.
That is a great overview as usual, Dave. Thanks. I guess, and I know it is not your exact expertise, but ultimately, we have to pay for this AI infrastructure. Do you think the applications or new AI-based agentic revenue streams are there, and what are you using yourself or seeing out there to pay for this infrastructure?
I am a firm believer that AI is as transformational to society as the internet has been. I think it will become embedded in everything we do. I think it will make us richer and more productive. I do have some real concerns about the way in which AI is being developed today, the ability to monetize that, and how their investors will get a return on capital. I think there are really four different points that are important. The first one is, today, we are developing AI in a very brute-force method using a lot of power. While the power efficiency of next-generation GPUs continues to improve, it is clear that for a long time we are going to be power constrained, so we need to be more efficient in our power utilization if we expect to continue to develop new LLMs.
Secondly, there is a bottleneck on the type of processor that allows companies like NVIDIA to have 85% gross margins and +90% market share. I am a firm believer in capitalism, and I think there will be a number of new entrants coming in to make the cost of that compute come down and be more power efficient. The second question is, the LLMs that are being developed will eventually become fungible with one another. If you look at the original Turing test, it was. Can you determine if a computer or a human is answering the question? If you cannot decide, it is probably artificial intelligence if it is a computer. I would actually argue that a correct test would be, if I ask a question of 10 LLMs and get 10 different answers, I have not yet converged on artificial intelligence.
If I ask 10 different models the same question and get the same answer, but did it through different models, I believe we will have true artificial intelligence at that point.
You mean AGI, artificial general intelligence, or—
That is correct, Tim.
Yeah.
Yeah. True cognitive intelligence, when you can ask a question across multiple models and get a consistent answer.
How far away do you think that is, Dave?
I think we're within a decade of getting that, maybe less. It's highly dependent on, can we train the models more efficiently? If we can, I think we can get there pretty quickly. The third question is, how are companies going to monetize that output to get a return on their investment? There, I don't see a clear answer. I think it's very different than the dot-com boom that you and I both lived through, Tim, where private equity and venture put in capital, and it dried up. Here, we've got very well-capitalized, somewhat monopolistic companies deploying huge amounts of capital and taking advantage of bonus depreciation and the tax attributes of being subsidized in that development. So I think we'll continue to see capital pour into this sector, but I believe the ultimate ROICs on that capital will be very low.
It's hard to predict if they're below what they are, the cost of capital for those companies, because some of those companies effectively have a negative cost of capital based on their rate of equity appreciation. But I do think we will see a lot of capital deployed with no discernible return. Third, the real value creation is at the inference level. At the inference level, that will mostly be distributed, much like a caching model. I think inference will occur in three phases. Either inference from cache tokens immediately responding to an answer, inference from centralized or origin tokens, like an origin server and a CDN, and responding with an answer, but over a slightly longer interval.
Third, there may be situations, I actually think the majority of situations, where users have to upload data to answer the specific question that they are asking, either from their phone or from their corporate networks or PCs. I think with that, we will see agentic AI doing a lot of that uploading, doing that in the background, and delivering the results appearing to be instantaneous, even though it did take some time. Because that process was running in the background, it will appear instantaneous. But it is very unclear how you will charge for that or get a return. My guess is that it will just become embedded in other things that we use, just like operating systems are embedded or the internet is embedded. That does not mean the end user does not get a lot of value.
It just means that it is very hard for the service provider to extract that value. There, I would look at the roadmap of the internet. You had very well-capitalized companies in AT&T, Verizon, MCI, Sprint, and CenturyLink that literally had trillions of dollars of enterprise value. The internet came along; it destroyed that value, and all of the benefits accrued either to the end users with free services, or it accrued to applications that sat on top of the internet, companies like Netflix , Uber, or Google. We have yet to see what the application layer on top of AI will look like, but I would suspect all of the value will be created at the application layer or will be transferred to the consumer. So, the $7 trillion that I mentioned earlier is great for society but probably not so great for investors.
Dave, which industries did you say got largely impacted negatively by the internet?
Telecom.
Yeah.
Cable, any service provider where the application and the network were delaminated, where the bit volumes increased and the revenue per bit declined precipitously. Because the internet was network agnostic, we had intermodal competition and access for the first time. Remember, it was against the law for anybody else to interconnect to the PSTN. It was not only a de facto monopoly; it was a de jure monopoly sanctioned by the government. You went to jail if you tried to build a second telephone network. The internet allowed companies like IDT, VocalTec, and Net2Phone, all these early companies, to circumvent that. We will see the deployment of AI that will, I believe, destroy the software industry as we know it, allowing that to become much more agile, much more specific, and more importantly, much cheaper. You will no longer need millions of lines of code.
If you think about what software has evolved from and to, originally, software was about computation and tabulation. Then, for the past 50 years, software is all about process codification. With AI, you no longer need centrally codified software. Each user can actually create software for themselves, and they do not even know they are doing it. They are just using AI in the background to do that. With that, we see business processes become much more efficient. No longer do you have to fit your business into the oracle view of the world. You now have your view of the world, and you use the tools to make the business work the way you want. Now, it will all be free, or virtually free, compared to a SaaS model or a perpetual license model.
The percentage of GDP that goes to software development will fall precipitously, yet software will become more important than ever. Just like the internet is more important than a telephone network ever was, yet the percentage of GDP that is spent on communications has fallen precipitously.
That is great color, Dave. Maybe getting back into your business a little bit more. Thank you.
Oh, okay. That is the purpose of why I came here.
Well, just two seconds. What are the top one or two other industries? Only because I have to do a call this afternoon to a whole bunch of people about this very topic. What are the other two industries you think are most at risk?
I think any professional services business that relies on proprietary data that sat behind paywalls and firewalls will be broken down. Accounting, legal, and engineering—you will no longer have to go to someone like yourself, a trained civil engineer, to calculate what size beam to put in when you are adding an addition to a building. You could probably take a picture of the building, send it to AI, it will do the calculation and deliver the spec.
Well, to your point, my civil engineering friends are able to do 10, 20-times more work than they were able to before.
Absolutely. But that just means, as a society, we need less civil engineers than we did otherwise because they became so much more efficient. So those professional services are another area. Medical falls into that. So every one of those intellectual endeavors now becomes much more efficient. We have talked about software; we have talked about professional services. Now let us talk about businesses. Businesses are basically the amalgamation of labor, capital, entrepreneurship to fulfill a need in society. Businesses will become much more efficient. The barrier to creating a business will be lower. Your business will be more equal to a larger business. So early on, it was that big businesses benefited because they had access to technology that small businesses could not get. The PC changed that. AI is going to change that even more.
The innovation pyramid will probably be inverted, where the real innovation will come from the smallest businesses, not the big businesses. If you think about the innovations we use today, they all derive from governmental programs that went to big corporations, that got to smaller corporations, and eventually to end users. With the democratization of these technologies, you will see every user innovate and create a pyramid that is inverted, where the last adopter of new technology will be the government, not the first. That is transformational for society.
Dave, you are probably the largest internet backbone provider. You have a very large fiber network and wavelength network. You would think you are going to be a major beneficiary of this, and obviously, you also have a very large corporate fiber network. How does Cogent benefit from this longer term, and when do we start seeing it in the numbers yet?
Listen, we have had a tough year as you introduced me, Tim. From my forced stock sale to the decline in revenue as a result of the acquired Sprint business to the increase in capital intensity to repurpose those assets and the increase in net leverage. All of those things have degraded Cogent's enterprise value. On a going-forward basis, the underlying trends that you have described will shine through and will have us return to top-line growth. Remember, for 18 consecutive years, organically, we grew at 10.2%. For three years post-Sprint because of the runoff in that business, we have declined for three years. Even though we have been able to grow our EBITDA and take costs out, investors want to see top-line growth. Two, we became a new entrant in the wavelength market. That is a relatively concentrated market. In absolute terms, we are doing pretty good.
We grew 9.2% sequentially last quarter, 68% year-over-year. The problem is, it is only 6% of Cogent's revenues, a $60 million business, so it is not big enough, fast enough. Investors are very frustrated because they hear about all this demand around us, yet they do not see the wavelengths installed. Well, if you are going to spend $7 trillion, you have spent $1 trillion, but you got $6 trillion more to spend, you cannot use the wavelength till the data center is built, till the power is in place, till the GPUs are there, till the servers are there with the adequate memory. So there are a number of preconditions that have slowed down the customer acceptance of our services. At the end of the day, every business wins for the same reason. They deliver more value than their competitors.
Cogent became the biggest internet provider in the world because it delivers the best value in internet. More bits in more locations connected to more networks at a lower price. It's not any more complicated than that. We have the ability to convert that into cash more effectively than others. Listen, AT&T used to be a big internet provider on the backbone. UUNET was the biggest. It's owned by Verizon, doesn't even really even exist anymore. Lumen has shrunk that business. Cogent has flourished because it has a better monetization engine. On the wavelength business, we have true structural advantages: more sites, faster provisioning, higher reliability, unique routes, lower price. It is a more concentrated market. There's a market we're new to. While I'm disappointed we're not bigger in the wavelength business, we've gone from 0% to 3% of the market in six quarters.
That, I think, is relatively impressive. Yeah, it's not enough. We need to do more. As long as we give on a route-by-route basis a better solution than our competitors, we'll win the business. Just like when you go to one of 1,953 data centers around the world and plug in for internet connectivity, more ASes connect to Cogent. We're in more data centers, and we deliver those bits cheaper than anyone else with equal or better reliability. That's why our traffic growth is double the rate of the internet. That's why we are gaining share. In the wavelength business, we're only 3% of the market. It's a hell of a lot easier to gain share when 97% of the market is not your customer. Whereas in the transit market, 25% of the world's traffic already flows on Cogent.
Great color. On the wavelength side, what's it take to gain more share at this point? I think a few years ago you said, look, you're going to price whatever it takes to kind of get up to that 20%-25% share within a five-year period. Right? Are you at the right price point, or are there other things you need to do to really accelerate that share gain?
If I thought further price discounting would move the needle, we would do it. We have the flexibility to do that. We are not there because a lot of this demand is not yet actionable. Again, you've got these three legacy segments that are growing, but at a modest rate, and we're capturing share. We have now sold wavelengths to 548 unique service providers or customers, and we've sold them in 608 locations. We've got our foot in the door. If you're Anthropic or OpenAI or Nebius or CoreWeave, you don't need wavelengths to a data center that's not finished. You don't need wavelengths to a server that doesn't have power. You don't need wavelengths to a facility where the servers aren't installed because Supermicro or Dell or whoever's making the box can't get the memory chips. There are just a number of constraints.
Yes, there is real AI training going on, but it's like building a house. You can move into a house before it's painted, and the painters can come and finish it. You can't move into a house before the toilets are installed because you can't use the house. We have a number of situations where there are critical path constraints for our customers that are not allowing them to buy the wavelengths that they want to buy. Those problems will be solved. Capitalism works. Greed is good. It gets people to do things, and we will see all of these bottlenecks, whether it be Micron or Samsung dominating memory, or NVIDIA on GPUs, or ASML on fab equipment, or TSMC on production capacity. Every one of these critical component supply chain limitations will go away. $7 trillion represents 25% of a year's global GDP. That's a lot of money.
That motivates a lot of people.
What's the main bottlenecks right now? What's the best guess when these bottlenecks start to loosen up? Is it 12 months from now, six months?
First, I don't think they're uniform across—
Yeah.
...all customers. I think they are subsiding, but not as quickly as we would like them. I would suspect we will be talking about bottlenecks two years from now, unless the capital markets take away the punchbowl. If the capital markets shut things down, it will not matter. But absent a constraint on capital, and remember, the capital markets are different here than they were in the dot-com boom, because Google or Apple can thumb their nose at the capital markets. They have trillions of dollars of cash on their balance sheet. Now, they want to be balance sheet efficient. They want to use off-balance-sheet vehicles. But if they are pushed to the wall, if BlackRock will not give them capital, they will just use their own balance sheet if they have to.
I do think these issues get resolved, and probably two to three years from now, it will all be in the rear mirror, but it will not be like there is a single-step function.
Fair point. But when do you think data center capacity and GPU compute capacity will really start to come online, that it can move your business?
I think it is happening, but at a slow rate, and if you just looked at compute capacity and graph that out, it looks like it is growing fast, but it is not growing nearly fast enough to keep up with the requirements of LLMs. We are falling behind. We have designed more models than we can actually train, and we need to catch up. Probably a couple of years. But then the constraint becomes power in the system. The developed world had built power infrastructure on a simple algorithm: 2% GDP growth and 1% conservation efficiency. So effectively, for the past 80 years, power capacity has gone up 1% a year. Traditional data centers, non-AI data centers, use 2.5% of global power availability. AI alone needs 3% of global power. There was not an extra 3% slack in the system.
You've got all kinds of wild schemes to add power, from on-site nukes to on-site turbines, and there's pollution issues and water availability issues. The reality is, the permitting for this is challenging. We're seeing a huge political backlash today on new data center builds. It's part of the reason why the 24 data centers that Cogent put up in the market to sell were so interesting. Even though they were old, they were small, they were in remote locations, they were here and now. You try to get a new data center built; even in West Texas, you're going to be fighting a lot of local politicians. I'm not going to say it won't happen, but it ain't happening quick.
Dave, just back on the internet side, the internet transit business. Do you think the volume growth will continue to accelerate off this? When and where do you think it can peak out?
Every industry eventually gets hit by the law of large numbers. The reason why, in 35 years, with a compounded growth rate of 23%, it hasn't hit the internet is because it reinvents itself. It's a different thing every few years. The next different thing is AI. The last different thing was streaming. So every time you think the internet is slowing down, people are smart. They come up with new innovations to sit on top of the internet and drive another leg of growth. We're today, globally, probably carrying about roughly 9 EB - 10 EB of traffic a day, move across the public internet, and that average bit travels about 2,800 miles. That's what the internet is. It passes through roughly four ASes, or networks, and about eight or nine routers. All of that is necessary to keep the price of bits falling. The average price falls.
Because of these supply chain constraints, prices are not falling as quickly as they've historically for equipment, but they will fall again. I remember meeting a CEO of a company that ultimately went out of business, and when he was trying to sell me his business, his key value proposition, he says, well, I have scarce routers. You can't get them. I said, John Chambers is a greedy guy. He's going to make as many routers as I need to buy." There is no such thing as a scarce router. There is no scarce memory. There is no scarce GPU. There could be temporary shortages, but not long-term.
Well, Dave, we're out of time. Sorry for starting late. We'll catch up soon because I got like 20 other questions for you. But it sounds like things are really about to turn around across the board.
Listen, Cogent's had a tough go. We've got balance sheets that we're fixing, we've got growth rates that are inflecting. We've been fortunate we've been able to demonstrate margin expansion and cost discipline. But we are dramatically undervalued relative to the worth of our business, in part because our growth rate decelerated, in part because we stopped giving a large dividend, in part because our leverage ticked up. These are things that are in our control. The macro questions that you asked, Tim, are the right ones, which is we're in the right place at the right time.
Absolutely. Well, Dave, thanks a million. Thanks, everybody. Take care.
Take care, all. Bye-bye.