Akamai Technologies, Inc. (AKAM) Earnings Call Transcript & Summary

All right. I'm Sanjit Singh. I am doing my last presentation for TMT' 26, and we're going to end it strong with Akamai, we have Ed McGowan, Chief Financial Officer at Akamai. Ed, thank you for joining us at TMT 2026.

Well, thanks for having us. Great to be here.

There's a lot to talk about. The Akamai story has gotten really exciting. And so we're going to dive into all the things that's going on in the company. Before we get there, for important disclosures, please see the Morgan Stanley research disclosure website at www.morganstanley.com/researchdisclosures.

So I wanted to start the conversation with you, Ed, on the various acts at the company has been good going through, right? The company sort of pioneered the delivery networking market, you're sort of the market share leader there. Then you entered into an act 2, which saw you build a $2-plus billion security business over the last decade. And now in recent years, you're pursuing act 3, which has seen the company enter the public cloud market and more recently, the GPU as a service for Gen AI inferencing with Akamai Inference Cloud. So a lot going on. For those investors who haven't been close to the story for some time, can you lay out the vision here for Akamai's Act 3 and why the management team is excited about the prospect for growth heading into the AI era?

Yes. Great question. So most folks know us as for our first act, as you call it, with delivery. And we've always had a vision of some -- we started off actually edge computing way before it was even a thing, right? This was back -- I can remember when I first started in 2000, we are doing some form of edge compute. It was pretty limited in terms of what you could do with it. It was more a functions as a service. But the notion of leveraging our platform for compute was always something that we envision doing. Security was a very natural extension. So as we're delivering these websites, whether it's a commerce site or a banking site, there's so much malicious activity that happens, it was very natural to be able to protect those sites and do it in a way by leveraging our platform. And at the same time, I'm delivering you a shopping cart experience. I'm also using web application firewall rules to protect that particular interaction. So that business has been phenomenal for us and grown, adding $200 million to $250 million of new security revenue pretty much for the last decade, really. And it's a lot of internal innovation. M&A has been part of our strategy. We've been very successful in integrating companies into our portfolio and getting a larger and larger share of our security wallet, and we continue to expect to do that. And we actually think that we're the next generation of the web, if you will, of the agentic web, a lot of our security products will be applicable there. Obviously, we'll have to do some innovation and do some things to conform with the new protocols and things like that. Noname and API Security is a perfect example of that, where you've got massive amounts of data flows. Today, it's API, tomorrow, it will be talking to different models, going back and forth to different systems and all that will be exposed. And I think we have a great opportunity there to continue to grow our security business along with our compute business. So there's a natural adjacency with delivery compute and security. And most recently, over the last several years, what was happening is some of our bigger customers, more innovative customers are coming to us saying, we want to run this application and it requires us to have root access. We need to -- we can't just do functions as a service. I can't use your kind of edge workers platform to program in Java script or program in web assembly. I actually want to run my own proprietary code. I'll give you an example. We have customer who's using us with their own proprietary technology for live streaming. And they want to do things like synchronization to make sure that the video is synchronized across all different platforms, and they have some code that they use to get much better performance out of that particular stream. Now in order to do that, they needed to be able to run their code on our platform. And we're not designed to really do that. So we bought Linode several years ago, and we deployed our managed container service, where we're essentially slicing off a sliver of CPU from all of our servers around the world. And we have containers where they can actually run that code in that in all -- as many locations as they want to. And we also saw, at the same time, as customers are coming to us for these more traditional compute use cases where the public clouds weren't giving them the type of performance that they needed. One good example, we have a case study on our site about Apple, what they do with Private Relay, where you couldn't have, say, 4 or 5 big clouds backhauling all that information and actually run the service because users are everywhere. And at the same time, with our security business, we're spending a fortune on third-party climate, right? If you think of the massive amounts of data that we're processing. And we realize that there's a pretty good opportunity here for us to leverage what we do really well, running a massive distributed system and get into the compute business. Number one, it's going to save us a fortune. So we've started off as our first major customer, and we were able to offload a tremendous amount of workloads off of the public clouds, and that's continuing to grow today. So we really proved out the fact that we could run this. We needed to acquire some technology to help us get there, and we bought Linode, which was primarily focused on small, medium business, the hobbyist, if you will, and we turned it into an enterprise-grade system. And just so happens now with AI, the demand for compute being in a distributed fashion is even growing greater and greater and did some interesting work with NVIDIA back in the summer, early fall and just most recently deployed GPUs in what we call our Inference Cloud. This is sort of the latest innovation. So it's a very natural extension of what we're doing, well informed by customer demand, and it's been growing phenomenally well for us.

But like from my perspective, I see stability in your delivery networking business, in your security business while you're seeing accelerating growth in your compute business and the CIS business. So let's dive into the compute and the cloud infrastructure services opportunity, starting with public cloud. Public cloud growth is accelerating across the industry, including for Akamai and your Linode business. What are the reasons why this part of the business has moved to a higher gear recently? And is this growth being fueled by a handful of customers? Or are you happy with the breadth of that contribution?

Yes. So I just actually came up with my last meeting with a group of investors where we were talking about, if you had asked me 3 or 4 years ago when we started this, we're now at a $400 million run rate in the CIS business, our cloud infrastructure service business. I would have thought we would have a much higher customer concentration. We would have gotten a few big customers to spend tens of millions with us, and then we would just start to get into a much deeper breadth across, say, finance and retail and travel and all that sort of stuff. And it's exactly the opposite, which is great. We actually -- that -- if I look at the composition of that, that's hundreds and hundreds of customers. It's anything from a customer doing a couple of hundred thousand dollars with us upwards to a couple of million dollars a month, but it's not driven by a small handful of customers. As a matter of fact, very little revenue today is coming from these mega deals, right? These are starting to ramp up, and they're going to become a lot more material to revenue over the next year or so. And we're also seeing our pipeline full of anything from customers working with our partners. So we have like a store, if you will, where you can go and say, if you're a media company, you want to do media workflow, we have media workflow partners where you can leverage their technology, use our storage and our compute. That's been going phenomenally well. Observability is one of the fastest-growing sort of use cases that we're seeing where customers are using an observability partner of ours and then ingesting information, storing it, using compute and creating all sorts of dashboards and things like that. And I wouldn't have thought that was how we would have scaled up so quickly. But it's been a great surprise and the demand is very strong across the installed base as well as new customers coming to us. We're starting now that we've -- started with the agentic web, we're actually getting a lot of interesting inbounds. It sort of reminds me of the early days of Akamai when you walk in when people used to have desk phones and you see the red light on and people will be calling and saying, "I need this stuff." We're starting to see that kind of activity again, which is kind of neat.

As I see investors sort of reengage with the Akamai story and maybe to take a break for a couple of years, but kind of looking at the trend lines of the business. The question I often get when it comes to the public cloud business is what gives Akamai or why do customers choose Linode versus the hyperscalers versus some of the alternative clouds out there. I think one of the important points that you guys make is that all three major hyperscalers are customers of Akamai but if you can just talk to the value proposition of Linode in the public cloud market.

Yes. So I think there's a misnomer out there that it's a zero-sum game. Most companies are doing multi-cloud, and there's applications even with some of the hyperscalers where we just get better performance, right? There's an application where you're just better served, whether it could be economics, it could be latency, we're using a more distributed cloud is a better answer. We've had companies trying to comply with standards in finance where, today we're complying with Apple Pay. And you can't get that out of a public cloud, but you can get that by using us. We have some customers that might have workloads that are in a third-party cloud or public cloud, that their data is accessed very frequently. So you store the data, but you're reaching in and grabbing that data is going all over the place. That's very expensive. You have a lot of hidden costs, egress fees, for example, it's something that we've had customers say, if I move to you, I'll use your computing storage and you're not going to charge them for the egress fees. And the reason we can do that is we have one of the largest backbones in the world and we deliver hundreds of terabits per second, so the cost for us is virtually nothing. And so we pass that on to our customers and we can differentiate from an economic perspective. So if you need economics, performance, sometimes it's just for diversity, where you want to have multiple cloud providers. It could be maybe a major AWS shop, but you have 2 or 3 workloads with us. So it's not necessarily a zero-sum game or one cloud takes everything.

Just a couple of follow-ups there. In terms of those lower egress fees, can give us a range on how much cheaper you are versus your -- versus the hyperscalers. And the other question was around -- let me just stick with that first.

Sure. Yes. So a lot of times, even in the CDM business, we started to see this, and we actually started to build some products to help with our customers, where they might have their web infrastructure at a hyperscaler. And you have the notion of your origin, right? So we're all the -- say, if you're a video customer or even a retailer, you have lots of images and that sort of stuff. The idea of having a CDN, obviously, is you want to get great offload meaning every time a request comes in, you want it to hit an edge server and not have to go back to the origin right? Better performance, lower cost. We have customers where sometimes we'll have 99% offload or, 99.5% offload and their bill for egress could be equal to what they're spending in CDN. But just that little bit of a cash miss. Because it's so expensive. And I don't have all the numbers handy, but you can look at some of the public costs for what it would cost for egress. And it's -- compared to CDM pricing, more expensive.

With the managed Kubernetes surface that you guys have launched, is that expanding the types of workloads that customers can run? Is that becoming a criteria for why customers gravitate to [ the low end ]?

Yes. I mean that obviously like if you think going back to the beginning of the discussion, where our first origins were in this notion of Functions as a Service, programming your cash, programming some basic content awareness where you could do, tailor personalized content based on an IP address. You can program the CDM to do that. That's very basic stuff. The -- having a managed container service, now I can run all sorts of different code and workloads in that container that I couldn't do in the earlier days on programmable CDM, if you will.

And in terms of the hyperscalers as we've mentioned, that their customers of Akamai and their public cloud service, what are the hyperscalers using you guys for?

Yes. So we've talked to sort of at a high level, some are using us for video delivery, some are using us for API management, some are using us for advertising decisioning and things like that.

Awesome. So let's move the conversation to the inference opportunity. Akamai Inference Cloud was announced at the end of October, beginning of November. The edge computing opportunity has been discussed about as a potentially big market for several years kind of across the industry. I would say it's generally been slow to ramp, and maybe it's having its moment now. But what -- with the opportunity around the inference, what gives you guys the confidence that a significant portion of AI inference will happen at the edge? And what will be the drivers for this opportunity to materialize?

Yes. Good question. If you think about the next iteration of applications. Today, you might go to a retailer or to like a Home Depot or a Lowe's where you interact with it with a search, right? So you do a search, and you get a bunch of stuff and it's helpful. But imagine now you work with a virtual travel agent or a virtual contractor. The amount of data that needs to be processed and the amount of compute that happens is much more robust, right? You might be -- have a small language model that you're using in real time. And you also want to make sure that, that is trained up and informed, et cetera. But the amount of compute for that is pretty significant. And you couldn't do that using, say, like our edge workers or program that in JavaScript and that sort of stuff. But latency is a big issue, right, where just like in the old days, when you had a shopping cart, if your shopping cart took forever to load, you see the circle going around, you're going to abandon that. You don't leave the site or refresh it. And the vendor, the partner of ours, ours customers will lose a sale. Same thing will happen in this world. It's just now compute is going to be a much bigger component, right? Because in order to deliver that experience, maybe that virtual travel agent is booking your dinner reservations for you. It's putting markers in your calendar for travel might be getting you a car service and recommending all sorts of different things. That's a much better experience, but it requires a lot more compute, Latency is a big issue, right? Workloads like robotics, you need to be close to where the factory is. Autonomous driving is going to require extremely low latency. So those are just a few ideas of what we're seeing from customers. And what's interesting is we started working with NVIDIA and doing the trial to see what can we get out of these RTX 6000 Pros? What is sort of that token to megawatt mathematics look like? And we were blown away with how -- what we can do with those in even relatively small deployment, say, maybe a megawatt of power gets you about 1,000 GPU, you can do some pretty powerful workloads. And we have customers coming to us, anything from, say, AI start-ups that want to do some rentals and just get access to GPUs. And we also have others that want to do clusters where they say, I've got a -- maybe it's a research project I'm doing. Maybe it's a post training. Maybe it's robotics. Maybe it's a really robust advertising decisioning model that wants to work on the fly and requires an enormous amount of data. So there's -- we're starting to see some really interesting use cases now and a pretty nice pipeline that's building up. And latency is a factor. As a matter of fact, with this very large customer deployment, the latency was the requirement here was such that typically, within the U.S., you get pretty good latency, like you're in the middle of the country and you're on one of the coast. It's usually tolerable. In this particular case, it wasn't. We had to be very, very close to where the application was being run.

Yes. That makes a lot of sense. You mentioned that large customer, it's a $200 million 4-year deal with a major tech customer at the forefront of AI. Why did that customer choose Akamai? And what needs to get done between contract signing and the timing of revenue, which you sort of pointed to at the second half of the year, maybe Q4 in terms of what we're seeing? Just the time line between why do they choose you guys and then the time line between to get revenue to start to materialize?

Yes. So this was an existing customer, not one of our top 1% customers. It's been a customer for a while. And when we announced the Inference Cloud back in the fall, they had called us and said, "Hey, we've got this unique circumstance where we want to run a cluster of GPUs. And here is the type of performance we're looking for. We want to trial it you, right? And we actually have some -- like NVIDIA does refer some business our way too where customers will come to them and say, "Hey, I've got this use case, I want to do X, Y and Z. I need this type of performance. How would I go about doing it?" and in some cases, working with us is the best chance you have to meet certain requirements. So it was just an idea at first, and then it was -- let's run a trial, so you do a proof of concept, and you run that on a small scale in one of our locations and see what type of performance you get. Once the customer was comfortable with that, they placed a very large order. In this case, it's a fairly good-sized cluster. We had to light up some more data center space in particular for this. So in terms of timing, signing a contract of that size depending on -- it's an existing customer, so that helps. So you usually have an MSA in place and that sort of thing. But it does take time, approvals internally and that sort of stuff. So it might be 3 to 6 months to do a contract from proof of concept when it actually gets signed with that scale. Maybe it's 2 months on the short end. And then there's a lag between -- in this particular case, since we had to get some new space before we get revenue. But the pipeline is informed from existing customers, new customers, referrals from our channel partners, also from some of the partners we have on the technology side like NVIDIA.

That's great context. I want to have a discussion on what it takes for the build-out of the AI inference opportunity. So we have about 20 locations today supporting Akamai Inference Cloud. What is the build-out of colocation facilities look like over the next 12 months? And then maybe looking beyond that?

Yes. So in some of the -- like right now, the 20 locations we have were all existing facilities, and we just can roll out racks of GPUs. So all of the sites we were in had all the necessary cooling requirements and things like that. So that wasn't a problem. With certain clusters, if someone who says to, "Hey, I want 1,000 GPU." we may not have space in the particular areas where we need to add on to an existing site or maybe deploy into a new site. So there'll be a mix of rolling out into existing sites where we have capacity. And a lot of our -- the way we're building our CPU or our compute centers is essentially you look for our bigger sites, the 41 sites or so are generally 5 megawatts typically the initial size. I may only be using 1 or 2 megawatts and I'll roll up to 5, but I'll usually have an option to go to 5, 10 or 15 depending on the type of relationship. A lot of the folks we work with will be doing a multiyear build-out. So there might be campus one is available today, and then we have an option on campus to that might be available next year. So there will be continuous build-out of capacity. Now in terms of like if you want to think about unit economics, I've given us-- a little bit of this on the call. So if you think about just GPUs in general for 1,000 GPUs, we'll stay in base 10, it makes it easy. It requires about a megawatt of power. It's about 600 kilowatts just for the GPU, but with all the other stuff, it's about, call it, a megawatt power, okay? So a megawatt of power in the U.S., you can get it for anywhere from $2 million to $4 million a year, typically. California is a little bit more expensive, maybe $6 million. But generally speaking, somewhere in that range. For the CapEx, it's funny. I did this the other day. I did -- I'd like to find public references. So I went to ChatGPT and said, "How much would a 1,000 GPUs cost with all the servers and switches and all this other stuff." And it gives you an interesting answer. It comes up with anywhere from, say, $12 million to $16 million. So let's round that up, we'll call it, $20 million on the high side. The rental market today is around -- I think we launched on Saturday $250 an hour. Okay. $250 an hour for 1,000 GPUs for this year would be $22 million roughly. $2 is around 17.5, okay? Let's take that number divided in half year, okay? So if you think about the unit economics, I've got high end, $20 million of CapEx. I appreciate that over 6 years, it's roughly $3 million, $3.5 million of depreciation, a couple $2 million to $3 million of co-location costs. And even if I get half of that price, it's still very good unit economics for me. The operating margin is in the 60% -- 40% range or better. $2 is phenomenal.

I had a whole section unit economics, but you got in front of me and that's super interesting. In terms of -- on the theme of the build-out of the AI inference opportunity, how do we think about the existing network and growing to that? So whether it's the roughly 50 Linode reasons that you have the broader 4,000-plus points of presence, is that going to be leveraged in terms of this build-out? Or are you building out like net new or renting out net new cohorts?

Yes. So there'll be a combination, I'd say, the hierarchical compute capabilities, functions as a service available in all of our servers. That's sort of serverless request based edge Java web assembly type programming for basic things. the container service, we actually have available in all of our locations as well. Limited amount of capacity, obviously. And some of our locations are pretty robust, where you might have -- we call them our multilink e-cores is the term we use internally, which effectively is a big server farm where you have most of your big telcos all connecting at once. And so you've got maybe several hundred to maybe 1,000 servers or something like that in those locations. And there, you've got more power, more space where you could run your GPUs in those locations. So we could leverage those and roll in some racks of servers and get maybe 100 locations or something like that. But there will be purpose-built locations for customers that have -- that when the math makes sense, where they have, say, a cluster that they need in a particular area, we'll purpose build for that. So it'll be a mix of leveraging some existing sites. But some of the sites are like in telco hotels. We don't have the power and the cooling necessary to run even a compute server, the 1U, 2U rackmount servers you use for CDM, pretty cheap. They can -- they don't require a lot of power, they don't require a lot of cooling. So you can run those in pretty small locations. But for the -- even a compute server is much more robust for CPU that you wouldn't run them there. So -- and also, I don't know that the economics would necessarily make sense for somebody to run GPUs in 4,000 locations. The math probably wouldn't work at that point. But it's not saying that we couldn't go a lot more distributed than say, 100, but practically speaking, probably 100 locations for GPUs at some point makes sense. We'll probably grow from 20 today to somewhere 20 to 40, something like that. But a lot of this can be informed by customer demand will be the main driver of why we pick additional locations.

And just as a follow-up there, I think the team mentioned that the initial capacity for the 20 has already been contracted out on the 20 existing location. How do we think about where do we stand in terms of the utilization potential and potentially going denser in those 20.

Yes. So that's not -- we didn't spend a heck of a lot of money there. There wasn't a huge revenue opportunity.

Just sort of proving out the concept out?

Yes, it was proving the concept out and we do have a sliver of that we use for proof of concept. So like I said, the typical sales motion is someone comes to and says, "I want to run a proof of concept." And the proof of concept might be running in a couple of locations, using a few GPUs or whatever, that we do on a regular basis. So we do have some of that in terms of being sold out, that is set aside for proof of concepts. But it wasn't a huge revenue opportunity. The bigger revenue opportunity is this initial buy of $250 million of CapEx. Only a portion of that goes to this big customer. There's still thousands of GPUs to sell there. So there's an enormous amount of revenue to come with that as well.

It's also the basis of my next follow-up is the timing of these purchases on the GPU, CPU side, does that happen once a contract has been signed? Or are you purchasing ahead of that, anticipating future demand?

I mean I would love to be in the situation I was just in where I had a big customer willing to -- if I didn't have the chip inflation, that probably would have covered practically all my CapEx, right? That's an awesome position to be in. It will be informed by the pipeline, growth of existing customers and demand. We will invest ahead of time like we do, but it will be informed based on what we're seeing in the pipeline and what customers are doing.

Awesome. And I want to talk -- spend the next couple of minutes just talking about some of the debates coming out of Q4, which -- I mean the results in Q4 were really strong. You have a CAS business that's accelerating, Global Cloud business accelerating $200 million customer Inference service. One of the things, though, is that margins did come down to 26% to 28% from about 29% in 2025. As CapEx steps up -- I think to the range you provided was 23% to 26% versus 19% of revenue in 2025. So with the full understanding that growth doesn't come free, and we need to fund this investment. How should we think about the margin trajectory from here as the Inference Cloud business scales?

Yes. So if you look at -- we gave this disclosure on the call. The -- where we're seeing the pressure points on margin are in cost of goods sold primarily around colocation and then depreciation. So if you look at that, almost the entire amount is those two items. And with colocation, there's two factors. One is we are buying ahead of time so as I deploy these thousands of GPUs with this latest purchase, well, I haven't sold them yet. So I don't have any revenue against it. Even with this big customer, I'm going to have that facility up and running for several months before it ramps fully up with revenue. So there's the timing aspect. And that goes for my entire compute platform, right? So I've got capacity. Any time I have capacity available I have some inefficiency in my math on [ colo ], right, which you'll always have that. The other thing is, as I talked about, how we buy some of the bigger locations, we'll work with our vendors for 7-year deals, 10-year deals, in some cases, and there, you're making a fairly good-sized financial commitment saying, okay, I know I'm going to be using 5 megawatts over the 10 years, and I want expandability, but I want unit economics today at that purchasing power. So I might make a, say, a $50 million commitment over a 10-year period and maybe I'm only spending $1 million in cash. But with the lovely lease accounting standard, you need to put an asset and a liability up. And whenever you have a lease that has either free rent or have escalations or commitments, you need to straight-line that. So there is some element of noncash colocation costs in my cost of goods sold. So there's some inefficiency there. So that's part of what's driving it. But that said, there's a trade-off between growth and margin. So if I want to grow faster, let's say a customer came to me and said, I want to spend $500 million with you, whatever, pick a number. Well, my margins are going to go down a touch before because my payback on my servers might be two years or whatever. I have to get colocation because a deal like that, I don't just have that sitting around necessarily, unless they want to distribute it everywhere. And perfect locations and fine everywhere that I have capacity that's not reality. It's probably going to be in several locations might be a U.S. opportunity or Europe opportunity or whatever. So there's going to be some inefficiency associated with that. Where we get good operating leverage is on the go-to-market side, on the engineering side and on the operations side. And we have some overlay sales for compute. It's a relatively small group compared to the size of our sales force. The people that are going out and getting the colocation deals and managing the operations associated with that, doing the deployments, buying testing the hardware and the different routers and switches and chips and all that stuff. The same group that does that for CDN and CDN basically runs our security on the same platform. So I get an enormous amount of leverage there. We've moved about 1,000 engineers out of our CDN business into our compute business because there's a lot of similar things that you do in terms of [ your ] engineering and running the distributed platform. So we will get a lot more leverage on the operations side over time.

Maybe to look at the unit economics question from the lens of a customer so sort of lifetime value of a customer, let's say, it's maybe not this particular $200 million customer, but you signed similar sized deals with commitments for 4- to 5-year ranges, what do you expect the contribution margin to like over the lifetime of that customer commitment? And I was sort of referring to, I think, a couple quarters ago, you mentioned that you think that maybe -- the Inference Cloud business could get back to kind of where your public cloud gross margins are.

Yes. I mean just if you go through the numbers we just talked about, let's say you get per 1,000, let's say you're getting $12 million to $15 million of revenue against cost of, say, $3 million to $4 million for colocation to $3 million of depreciation. Your operating margin is north of 30%. Your gross margins are in the 70-ish percent. So you can get there certainly and that's assuming a discount off of what the published price for rentals are. Now the other misnomer I think that's out there is people believe, well, 6 years depreciation, that's probably not the right way to think about it because what about the next generation. Well, sure, there'll be a next generation in that 6 years, probably many next generations. That doesn't necessarily mean that everyone is going to move to that next generation because you've done -- like one of the things I'm finding is customers will come to us and say, my particular application runs best on this chipset, maybe it's AMD, maybe it's NVIDIA, maybe it's Intel. And I also have a particular workload in a set of economics that I might not be able to afford the next generation of chips. So that may not make sense. And I look to our Linode business. Before we bought Linode, they offer GPU as a service and it's been a nice little business for them, and we're selling generations that were put in place before we even bought Linode. We have grown Linode now for 4 years. So there's -- I anticipate -- I mean, I could be wrong on this, but I don't think I am, there will be a market. Now maybe the unit price has come down a little bit. So how you get good cost recovery of say, a year to two on my service. And the same is true with CPU. You -- kind of the dollar CapEx dollar revenue gets you above a year or so. Even if it's $0.85, you still get somewhere between a year or two of -- maybe it's 18 months of recovery on your CapEx. So it's a great business. I'm not too concerned about -- most customers would probably renew unless their application goes away, but generally speaking, it's tied to a business that has some notion of scale with it as well that you might say, "Hey, I want to start off with x amount because I'm doing an application that is tied to something else I'm doing that's revenue generating. And as long as that's growing, I'm going to continue to need more and more. And we do see that. We see very nice, call it, same-store sales growth of folks that are using us. So that's very true with the public clouds.

Maybe last question [indiscernible] time about security in the delivery business. So I'll kind of bundle the question. Your confidence in sustaining the growth that you have in [indiscernible] security, was 9% constant currency. I think you guys guided to high single digits, your confidence in stating that level of growth in security? And then the delivery side of the equation, it does seem like the pricing environment is a little better. Fewer competitors out there, we got Olympics, we got the World Cup. And how should we think about the growth trajectory and delivery over the next year?

Sure. Delivery has gotten a bit better. One of the things that's a little difficult to digest this year as we had acquired the assets of NGL. So throughout the year, we had modeled it that some of these customers, some of the bigger customers were multi-CDN and we bought the other CDN. So they have to rebalance. So let's say we have somebody who's doing 50% with us, 50% with Edgio. We combine it, we probably end up with 75% share as a general statement. So 25% goes to fight somewhere else. That happens throughout the year. It's not all like I closed on December 20 and December 21, everything rebalances. It happens gradually throughout the year. So you did see revenue declining. And part of that was like modeled and planned, we knew that was going to happen. Someone else was going to pick up some share where they didn't have share before. Now we are seeing the pricing environment get better. And we're taking a much tougher stance with pricing now saying, look, we just talked about on our fourth quarter call, $200 million of additional CapEx just because memory prices went up, right, colocation is not getting cheaper. In some markets, it's solid, but other markets more expensive. Labor doesn't get cheaper. We can no longer eat all of this. We have to start passing along some of that price. And certainly, not accepting a 15% or 20% price decline in the delivery business. So we are taking a much tougher stance. We are trying to raise price on renewal. We'll be successful in some areas and others we may not be. But that's a much different stance from us. We've never raised price before. We're doing it for the first time in our history. On security, I gave out some really interesting statistics on the fast-growing parts of the business. API security exit over a $100 million run rate, growing at over 100%. The penetration rate in the installed base is sub-10%. Very applicable for our WAF customers to use API security. So that's going to look -- enormous runway. And I actually think in the agentic web, as we talked about earlier, with all the data flows that go with that more robust application that's going to be a massive attack vector. So we think there could be even a second generation of growth out of API security, protecting those agentic applications in the future. In terms of Guardicore, again, sub-10% penetration of the base, the majority of that business is coming from new customers, a very robust channel there. Very great dynamics with existing customers. The typical motion is to add more agents to protect a greater slew of your network. So you start off running maybe 1/4 of your network and then you expand our license business, which isn't a significant business, but has 90x percent renewal rates. Usually, it's an upsell. They're term licenses. So very healthy dynamics in that part of the security business that was growing at over 30% last year.

Well, we're out of time, Ed. Thank you so much for giving the update on the Akamai story. A lot of great things happening in the business and all the best for 2026.

All right. Thank you, Sanjit. Appreciate it.