Welcome back, everyone. Thanks for being here. I'm Rishi Jaluria. I cover software here at RBC. I'm delighted to have with me the CFO of Akamai, Ed McGowan. Ed, thanks so much for being here.
Thanks for having us. Appreciate it.
Yeah. Look, I think Akamai is obviously a household name, but there's been a very clear evolution in the business as you become more of this AI infrastructure provider, moving away from the traditional, or let's say diversifying beyond the traditional CDN that's the roots of the company. Can you maybe talk a little bit about that evolution? We'll get into some specifics from here.
Yeah. So, it's funny. We've always been in some form of computing from the early days, if you go back to the initial founding. And we sort of refer to that as sort of functions as a service. So think about doing compute in, say, JavaScript to help a website do basic things like create a waiting room for, say, a really popular event where instead of seeing a 404 or some type of error message or spinning dial, you sit in a nice waiting room. Or maybe it's doing A/B testing where it's showing an ad to one person and looking at the results. So, computing has always been part of what we've been doing.
I would say probably as part of the security business grew, you obviously start to look at enormous amounts of data, store data, compute on data, and we're using third-party clouds for that, and it was becoming incredibly expensive. The expense rate was growing faster than revenue. As a CFO, that's a bad thing. At the same time, we were getting requests from customers to do more than just what we were doing with functions as a service. They said, "Hey, I want to run my own code on your platform, and here's what I'm looking to do."
Private Relay is a great example of Apple coming to us in the early days, and we said, "Sure, we can do that," but our platform wasn't originally designed for that type of thing, but we can do something custom. We did that, and then we had a couple of other customers come to us for similar type requests, and we said, "All right, there's enough demand signals, A, from ourselves, and B, from our customers, that we should probably think about getting into this business." We acquired Linode back in March of 2022.
Yeah. That was a great platform to start with. It's very developer-friendly. You could spin up a virtual machine in literally minutes. We said, "Okay, we need to make this enterprise-grade for ourselves, also for our customer base." Here we are today. We've more than 4x the revenue that we acquired and obviously signed $14 billion of large deals this year, and it's really sort of the breakout year for us in terms of getting to a massive scale. We see it as an incredible business for us with incredible demand, and we're leveraging more and more of that platform ourselves as we go.
Yeah. Maybe let's hit on the $14 billion that you've seen so far year to date. Can you maybe talk a little bit about the current demand environment and where are you in your ability to take on more of these deals andâ
Yeah.
âstrike while the proverbial iron is hot?
Sure. I think the way to think about this business right now and the environment we're in is you start with power. Power is the most important, scarcest element today. You look at what sort of power can we add over N number of years? What does that look like from a capacity perspective? You start there. The demand signals we're getting from our customers are insane. I've never seen a pipeline like what we have today, where we actually have to turn away business.
You've heard AWS say they're sold out through 2028, and the reason for that is it's the lack of power in the market. 2027 right now, if someone came to me and said, "I want to do an X billion dollars of deals and I want it in 2027," we couldn't provide that. No one can. There's just not that sort of power availability. As you get into the 2028, back half of 2028 to 2029, not an issue, but there is a real tight constraint at the moment. That's where you sort of start with the business.
We're seeing demand from anything from our largest customer soon to be, when they start generating lots of revenue for us, which they're starting now, to startups, to enterprise customers. We're seeing kind of an interesting thing with enterprise customers where they're coming to us saying, "I want to build my own model, train my own model with my own data because I'm spending a fortune going out to some of these third-party models, and then create a way to go for questions that can't be answered off of my own data to somewhere else." It's almost like a DIY sort of thing.
which is interesting. It might be I need X number of, say, B300s or RTX PRO 6000s or whatever it may be for GPUs, and from there, it is building up applications that are requiring a lot more compute than, say, the normal website. There is demand signals from all over the place with our enterprise customers for doing multiple different things. Then there is the foundation models. Obviously, they are looking for capacity any way they can get it, to interesting new innovative AI companies that have really interesting ideas that are looking for capacity, et c. It is really coming from all over the place, and it is the biggest pipeline I have ever seen in our history. It is insane.
Yeah. It has been a super interesting transformation. As you sign these really large AI infrastructure deals, there has been a lot of debate in the market around unit economics and IRR. On the most recent quarter, you kind of gave us the math to get to, I believe it was 13% IRR. Can you maybe explain what are the assumptions underlying this? What are the different scenarios? What is giving you confidence? Maybe help us get a little bit around that.
Yeah, so there was a debate. It was crazy. I sort of learned a lesson about be careful what you disclose, because people went into a dark place saying we were not making money, which was not true. So we wanted to just clear the air and say, "Look, this is the calculation for whether it is ROIC," which I do not think is the best way to look at this, or IRR. You have got to keep in mind, with a contract like that is so foundational, if you will, the leverage you get off of that is massive. So we included in our cost all of the costs that you would need to fit out a data center to connect it to your backbone.
So there is router costs, switching costs, there is fiber costs, all that stuff. Also, when you build out a platform, you need to create overhead, right? If someone wants X number of million vCPUs or GPUs, whatever it is, you need to provide overhead on top of that. So think of that as sparing or additional capacity. So we included that in there as well. As we add new customers to these data centers, and remember, these are big footprints for this particular customer, they get to benefit, or we get to benefit, actually, because that cost is going to be shared across many different customers.
We included all of that in our analysis when we did the calculation. We were also conservative in the sense that we just assumed that the revenue runs through its contract. We actually contemplate, at the end of that contract, to continue to sweat the asset, right? There is a pricing mechanism that says, "Hey, look, if the assets are still viable year eight, nine, and 10, well, we can continue to use that." What we see in our own network today for CPUs is we are getting seven-plus years with a failure rate in the low single digits.
Yeah.
For seven-year-old machines, we have some stuff that is 10 years old and better. We said, "Let us just assume just the contract length." If you have any assumption for future, obviously the IRR goes up dramatically. Obviously, there are time-based assumptions in there as well, which is pretty conservative. You can debate where that comes out. But also keep in mind, for someone who is going to do a deal of that size that has that much benefit to you are going to get the best pricing we offer to anyone.
Yeah.
Think of that as the floor in terms of IRR.
Yeah. Okay. Maybe let us hit on the CPU discussion. I think Tom had said 90% of the servers deployed are CPUs that has better revenue per megawatt versus your GPUâ
Yep.
âheavy configurations. How should we be thinking about what that optimal long-term mix of CPU to GPUs look like? As AI inferencing matures as a category, do CPU deal economics structurally improve your competitive positioning?
Yeah. It really all comes down to demand. So, one of my customers actually gave us a good analogy where they said, "Think of the GPU as the brain," right? So you train the brain, and then the CPUs are the hands, the feet, the tooling that you need. All of these systems, and there's some statistics out there that I read that says for every one GPU, you're going to need eight CPU or something. I don't know if that's true or not, but there's some math.
But we're certainly seeing that you do need enormous amount of CPU to make these things work, right? You can envision sort of the next generation of websites where you go to a website today and you do a simple search for things, you can think of a travel agency or whatever, to now having a virtual travel agent or virtual contractor for a site like a Home Depot or something. That's going to require an enormous amount of compute, right? That's powered by CPU. You don't need GPU to run that.
I give Tom a lot of credit. He said this years ago, and people laughed at him, said, "Oh, you can't do that. You can't do AI with CPU," et c. But there's a lot of applications that CPU is just fine, right? But we're going to go where the demand is. If we see demand for various types of GPUs, not only NVIDIA but others as well, to the extent that there's demand there, we'll obviously provide it. But I do think that CPU, in the long run, has a long, long way to go.
Yeah. Maybe let's talk about the financing kind of state of the union.
Sure.
You just closed a $2.5 billion offering last week. I think Fitch put an investment-grade credit rating on there.
Correct.
I think you said business is now fully funded for these committed dealsâ
Yep.
âwithout having to go to the debt markets through 2028.
Yep.
Can you walk us through kind of how this updated financial positioning is setting up the company for successâ
Yep.
âover the next several years?
Yeah. There was some debate about we needed to have additional funding. We are funded all the way through 2028, and really, it just comes down to how fast you want to grow in 2029.
Do we refinance some debt? We have debt due in August of 2027 and November now of 2028. But you're pretty much self-funded at that point. If you're growing faster, obviously you may want to tap some leverage to grow faster. It really just comes down to that in terms of how fast you grow versus how much leverage you put on. Obviously, I showed you guys with our schedule how much EBITDA we're adding.
So the deleveraging is massive over the next two years because, A, we're paying off debt, and B, our EBITDA margins are expanding dramatically. In terms of investment grade, we're investment grade with Fitch, we're investment grade with Moody's. We did get downgraded by S&P, and they just took the assumption that you're going to continue to refinanceâ continue to grow really fast and need more leverage and stay at an elevated level, where the others looked at it as, "Look, this is really financing a project.
Yeah.
So you take on additional leverage, and you delever really quickly. So you get back to. It's almost like doing M&A where you have an assumption of big synergies that take a year or two toâ
Yeah
âcome up with. So it was just a difference of opinion of how you model that, basically. But yeah, and the other thing I was amazed at was just the demand was very. Every deal is oversubscribed, but this was dramatically more oversubscribedâ
Right.
âthan I was expecting, which was a good sign. And obviously, the cost of capital is incredibly low.
Yeah. Maybe just to build on that, how should we be thinking about the timeline to return to being sustainably free cash flow positive? Then you've historically had very healthyâ
Yeah.
âfree cash flow margins, especially for a capital-intensive business.
Yeah.
What's the timeline to get back to healthy cash flow?
I don't want to put a number out there because it does depend on growth, right? If the growth is coming from CIS, obviously you put capital in the ground first before you get revenue. Just take your models and build them out and say, assume different growth rates, and you get there pretty quickly, right? You could assume at some level of growth in 2029, you could be there in 2028. If you want to grow faster, you probably push it off another year.
If you want to grow even faster than that, it might be another year. But once you get to a point, at some point, you get to sort of a normalized growth rate, say, like a double-digit growth rate at some point, and you say, "This thing is producing an enormous amount of cash." It's insane how much this produces in cash when you get to the other side of the build-out.
Yeah.
So we'll be back at that level at some point. I don't want to call a date yet because with the demand signals I seeâ
Of course.
âand investors want growth, so we'll see in terms of what things look like. But if you model this business out, you'll see that you do get to very healthy cash.
Yeah.
Free cash flow.
Okay. No, that is great. And so maybe the flip side to all the healthy backlog and commitments and everything is in the last quarter, you did see a little bit of revenue slipâ
Yeah.
âfrom Q4 into 2027, some delays that I believe you said have been resolved. How is kind of that tracking so far, and any kind of risk that we should be keeping out on the side that there could be further slippage like this?
Yeah, so I call it sort of spreadsheet risk. What I mean by that is we had three large customers that were starting all in Q4. You have to make a call and say, "When do we think we will be able to start taking revenue?" To start taking revenue for deals like this, you need, A, the shell of the data center to be built. B, you need the fit-out, which means the customer configuration where they might want to, say, have power from the ceiling versus the floor or have some privacy shield in front of their servers or something like that, and that takes several weeks to build. Then you get the equipment in. Sometimes equipment can come in that might be defective, so you send it back, you get it back, et c.
So it is really timing risk. You try to bake that into your forecast, and there is only 90 days or 91-day month, a quarter, I think, that you can generate revenue. If you are off by a few days, you are going to have less revenue, you cannot make it up. That is all that happened was we had a timeline that we got initially at the end of the year. We took a conservative approach, and as we got closer to Q4, we ended up having some delays that were longer than we had expected, so we just had fewer days of revenue, and that is it. Now all three of them are generating revenue today. Those three projects are generating revenue today, which is great.
The risk from then is much, much lower in terms of now it is really just about when do you start to see your supply chain ramp up as the customer ramps up. So much lower risk associated with that. With the big contract we just recently signed and we had had a call in September about, that obviously starts ramping up next year. We gave you a $200 million range, which effectively, it was 150 particularly on that, but just in general on CIS. We have given ourselves a lot of runway there that the risk is, sure, there is some spreadsheet math there that might be off by a few days or a month or whatever, but we have tried to take that into consideration. So execution risk, very, very low.
It is really the construction and the fit-out is not done by us, it is done by a third party. They have a schedule, we take a conservative approach, and we will get it. We rack and stack the equipment, becomes available for the customer, that is when revenue starts. It is really execution risk is relatively low. There is some spreadsheet risk, I guess, in terms of the date in which everything starts. You could take a positive view or a negative view. We have given you a big range on it, and we have tried to take that into consideration. I think, obviously, when you are talking about 365 days versus 90 days, you have got a little bit more room there, so the risk is lower. But sure, that is why we gave you the range.
Yes, absolutely. I do have a bunch more questions, but I did want to open up to the audience if there is any questions. If you do, please do raise your hand. We have a mic runner coming out just so everyone on the webcast can hear it. All right, I will jump right back in. I want to talk about an opportunity that Tom has been pretty upfront and forward on, which is really this opportunity in physical AI. Very timely, we just published a huge deep dive on the topic.
Yeah.
Can you expand, like, why do you feel Akamai is uniquely well-positioned for this opportunity? Tom has laid out several deals that have already been signed in this space. What are some use cases that you are seeing particularly interested in today?
Yeah, so, I will never do Tom justice, so take this as the CFO's version of talking technical. Performance matters, right? If you are like, say, robotics, for example, that really has a performance requirement that you cannot add latency to something. Like if you are a machine that is operating on a floor and it sees something and it needs to react to it versus, say, crashing into something or doing something destructive, that needs to be very instantaneous. We actually had one of my sales guys talk to me about a, it was a crazy idea, but someone in the agricultural space that was developing a sensor on a tractor that would be able to identify a type of weed and zap it with a laser based on what it saw. Now obviously, the ROI is pretty interesting for the farmer, for the equipment guys, et c.
But you are talking about being in the middle of the country, middle of nowhere, latency sensitive obviously matters. Cost matters too. The fact that we are more distributed, we can get, say, West Coast performance without paying San Jose colo costs and things like that. There are examples of, think about the next generation of your web interaction with folks, like a travel agent, like I talked about. Well, that matters, right? You do not want to abandon a shopping cart. Similar thing, but now it is a lot more difficult than just serving up an image. Now it is making sure that interaction, when you put your credit card in or you are looking at an image of yourself in different clothing or whatever it is, needs to be responsive and perform well, and also be secure.
Yeah.
I do think this is playing into our strength. We are leveraging our customer base who is doing these physical AI things and developing sort of the next generation of interaction from a commerce perspective or, say, from financial services or whatever it may be. They are the ones that are out there innovating, using AI to make that experience better, and it needs to perform well and needs to be secure.
Yep. So maybe let's build on that. You've just launched Akamai Inference and Agent Sandbox.
Yep.
Can you talk about kind of the edge inferencing opportunity, which obviously expands on what you're doing with the central cloud, and what does the product roadmap from here need to deliver to unlock that opportunity at scale?
Yeah. So, it's just more of development of things that we've been doing.
Yeah.
I think it's more just waiting for the demand to show up as folks start to leverage the technology that's available to them and start to generate these new use cases, and we're seeing these come up all the time. But it's right in our wheelhouse and something that is core to what we do, so we're leveraging a lot of the same engineering team that built the distributed platform. That's really the hardest part, which is, how do you do this in a distributed way where the customer will come to you and say, "I need to run an application.
I don't care about, I don't want to think about whether it's CPU or GPU. I just need it to perform well everywhere." That's where we come in and we take that, just like we did with the CDN and with security, we are able to make that sort of seamless for the customer. So think of it as like a serverless type â
Yeah.
âenvironment.
Yeah. Okay. That makes a lot of sense. Then, I want to think about, we have talked about the backlog. Now, one of your large deals has a big expansionâ
Yeah.
âopportunity. Can you talk about what needs to happen for that to be exercised, and what do you need to do in terms of execution? Then, we will bring it back toâ
Sure.
âthe whole financing and balance sheet.
Yeah.
Maybe walk us through those.
One thing, I did get some folks thinking, they were concerned it is a put option, meaning they can put this on us and say, "You have to deliver it." It is not like that. It is literally a collaboration. If there is a project that we want to bid on and it makes sense and we both agree, then great, we do it. The relationship with that customer is fantastic. We got a lot of great technical relationships, a lot of great collaborative thinking, that they will come to us with a project and say, "Here is what I am looking to do. We want to see if you can do this for us."
It may be something in terms of like, "We want to run it in this type of environment. We want this type of architecture," whatever it may be, and if it makes sense for us, we do it.
Yeah.
They come to us all the time with all sorts of different projects, and if we are interested in bidding on it, we do. Obviously, it comes down to my earlier point, which is, if there is capacity for it, we will let them know, like, "Yeah, we got capacity here in this particular location or at this time." Really, just again, it is a both sides have to agree on something, but there is no shortage of ideas coming from them.
Yep. Okay. No, that is helpful. Maybe let us turn, we have obviously spent a lot of time talking about CIS, but maybe let us talk about security.
Sure.
We have seen high single-digit type constant currency growth on the cybersecurity side. At the same time, we are seeing AI-driven bot traffic up, I think you said 10x year -over -year. How should we be thinking about the kind of base case for the cybersecurity, especially just given, we just had a huge enterprise AI panel, and clearly cybersecurity is one of the top areas with AI today.
Yeah.
Can you maybe walk us through, and then what needs to happen to see that accelerate from here?
Yeah. Clearly, AI has driven a lot more demand for security. It also helps us in our products, and we have been using it for years.
Yeah.
In terms of making your products work better, et c. Right now, our customers are grappling with, "What do I do with all these bots that show up?" There is not a revenue opportunity from a traffic perspective because it is very small traffic, so it is not like the CDN is going to grow as a result. From bot management, which is a security product, it also gives customers the ability to say, "This is a type of interaction. It is not a human, it is a machine. What kind of machine is it?" You let them know what it is. "Do I want to block it? Do I want to allow it to happen? Do I want to do something else?" We are seeing demand for that obviously increase.
Yep.
There is not necessarily a direct correlation to volume to revenue per se, but there is obviously a lot more demand for that type of a product, for sure. Also, just in general with how quickly vulnerabilities are discovered, the idea of a customer, say, that might be using us for web app firewall, might be protecting, say, 50% of their sites that they consider to be their core sites, are now talking about, "How do I expand that to cover everything?" We are seeing demand for that, for sure.
I think it is just a healthy demand environment. With a recurring revenue business at this scale, it will not show up in any one quarter. This is something that I look at and say, "This business has been great for over a decade, producing $200 million-$300 million of new revenue every year, high margin, high free cash flow business, that I feel very good about the fact that this demand will continue into the future.
Yeah.
Is there an inflection point? There might be. I would not call one right now to say that, "Hey, this is going to re-rate the growth rate per se," but obviously, the environment is there that it could possibly happen.
To build on security, I want to talk a little bit about API security.
Yeah.
I think, especially with the acquisitions in the space.
Yeah.
But you've talked publicly about how there's just a massive opportunity.
Yeah
Because API and MCP with all these.
Yeah.
Agentic traffic presents so many more vulnerability surfaces.
Yeah.
Can you walk us through that, and how confident do you feel about your current portfolio within API security?
Yeah. I think we obviously bought the market leader in API security, and we've obviously developed from there and added significant enhancements and capabilities. If you think about the market, you can use web app firewall as a proxy and say, "Okay, that's a billion-plus dollar product." Obviously, every website has lots of API calls to it, so that is obviously a great upsell opportunity, and we've been doing that very well. That product is growing very, very fast. But also, non-web properties, and it's not properties, but applications have APIs. If you think about that's obviously, that market opportunity is bigger from that perspective. As you think about the next generation of application, and let's say, I'll stick with this virtual travel agent, for example.
Think about how much more interactions that application's going to have, model-to-model traffic, model-to-application. It's going to be able to do a lot more things for you. Think of what Muse does today, where it might be booking something on your calendar, maybe making a dinner reservation for you, or something like that. There's just going to be an enormous amount of more, whether it's APIs or MCP, whatever the protocol is, doesn't really matter. That notion of having this exposure needs to be protected, that fits right into our sweet spot. So I think it's going to be a great opportunity for us over the long run.
Yeah. Okay. Maybe I want to get into M&A as a topic. This has obviously been very critical for your growth levers, starting from the Prolexic acquisitionâ
Yeah.
âto get into security back in the day.
Linode.
Linode, massive to get you into compute.
Yep.
Look what you've been able to do with it. How are you thinking about M&A as further platform expansion opportunities for you? We always ask that build versus buy versus partner debate.
Yeah. I mean, I think, if you look at our acquisitions, the transformative ones were the two you mentioned. Prolexic really sort of cemented ourself in security at the time. We were always doing some level of security just in providing our basic service, but this sort of put us on the map there, and we've obviously added enormous amounts of functionality, whether it's developed ourselves or through acquisition. That will continue. I don't know that there's necessarily a transformative thing we need to do there. Then obviously, Linode was very transformative. So those are rare. Rarely what you see from us is we either do like a tech tuck-in where you're buying talent, and certainly in AI, that's going to be more prevalent, where you might be able to pick up a team that can do some development for you and get you to a spot much faster. Certainly, as you go into some of these different areas of technology, there's expertise in certain scaling things.
Yeah.
It's amazing that there's these small companies that have some experts that can do things around storage or something like that can really help you scale. We also have done a nice job of buying a company like Noname that had spent four or five years developing the product. They had it in market, and they were at that point where you either take on a big round to go spend an enormous amount of money to build out a big enterprise sales force, or you exit. We dump that into our force and sell it from there. I think we'll continue to look for those opportunities and be very opportunistic. I don't see a significant change in our posture towards M&A. Obviously, from a capital perspective, you always look at what are your alternatives to that capital. Right now, CIS is growing so fast, and it's got a great return that it's got to be a pretty good challenger to that in terms of spending capital versus putting it into the business to get a return on CIS.
Yeah. Okay. So maybe now to close it out with the 90 seconds we have left. We've been talking about your opportunity in central cloud. We've talked about your opportunity in edge cloud, which obviously gets even amplified with physical AI. As I can see, you're the only platform that provides both of these from a single vendor, kind of integrated. Is there an opportunity to have kind of a full-stack integration that you're running AI workloads throughout the full life cycle, from that central to edge, right, for these sort of use cases? What needs to happen from a technological roadmap perspective to get there?
Yeah. I mean, I think, we're doing that today. We're obviously not doing the big training per se. We are doing post-training.
Right.
So, you are seeing some pretty big, intense compute being done on the platform. I don't see us changing our posture there and say, building out gigawatt facilities and stuff. That's not where we're going to head. But we are seeing it today. In terms of innovation, it's more about getting that serverless experience, right? Which is, how does somebody come to us with an application, just dump it on us? Today, typically someone might say, "Hey, I want to build out my own model. I need X number of B300s or RTX," or whatever it is, to, "Here's an application, just make it work." Right? There's just more development that goes into that, but that's already underway, and we have some applications that we're doing that today. Our managed container service is a great example of what the early innovations are there, where you have folks running massive applications for millions of users, running their own code in hundreds of locations.
Yeah.
We just make it work.
Yeah. Awesome. I think that's a great place to jump off. Ed, thank you so much, and thank you everyone in the room for being here.