All right. Good afternoon. Thanks so much to everyone for joining us at the Akamai session at the Goldman Sachs Communacopia + Technology Conference. My colleague, Max Gampl, and I are delighted to have on stage, Ed McGowan, CFO of Akamai. Thanks for being here, Ed. Thanks for having me. A great conference so far, so thank you for doing this as well. Oh, the pleasure is ours. Ed, you're in the middle of a very interesting inflection point for Akamai. Yep. Give us a little bit of a download of the last 6-12 months when you all as an executive team realized that this network that you've built over multiple decades can now essentially be upgraded along with the compute assets that you have into something that is a next generation AI compute network. Sure. Yeah, it has been probably the most exciting time since we first began. I have been with the company since 2000, and back then, demand was insane. Pretty much, the internet was just exploding. Yeah. Incredible demand. We are seeing that now with AI. We made a decision five years ago to get into the compute business, and we acquired Linode, and we always had this vision of, A, using it ourselves to start, because we were spending so much with hyperscalers, but to make that into an enterprise-grade network that would be comparable to using a hyperscaler. It may not have all the capacity and all the features, but it was an alternative, and we thought that, look, we can leverage a lot of what we have done over the years with our CDN network, leveraging a lot of the same teams that build some of the functionality, running the distributed platform, the operational teams, et cetera. I think the big turning point for us is when we sat down with Nvidia about a year ago or so and realized that there is this big opportunity to lean into not only GPUs, but compute in a much bigger way, and start building out bigger locations and offering an offering for compute from large scale CPU and GPU deployments, doing things like post-training or support systems that go along with a lot of these big AI companies. Leveraging our platform all the way down to our MCS, our managed container service, where we are taking the Linode technology and slicing off part of the CPU that we are using for CDN and for compute to run companies' code to do things like manipulation of video and synchronization of video, doing live, real-time ad decisioning, and things like that. It goes way beyond what we were doing before with functions as a service where you could do basic programming in WebAssembly or JavaScript. Now the customer has a lot more ability to do a lot more compute right out at the edge. So from core data centers all the way out to the edge. Tell us a little bit about how you've been able to merge the best of the Linode assets with the best of the Akamai assets to the point where you're signing multi-hundred million dollar, multi-billion dollar deals. Yeah. So it was really pretty simple. The first thing we did is we, over the years, built out a massive backbone. So we carry hundreds of terabits a second of traffic across our backbone. We stitch together our main CDN locations across the world. So we have this big asset that's connected up to all the locations that we acquired from Linode, and we built out another 20 or 30 or so sites that were initially 5-10 MW in size. So they could do a decent amount of revenue, a couple hundred million of revenue per site. All we're really doing is just now expanding those sites to be a lot larger. We're leveraging the same teams that build the CDN network, that negotiate for colo, working with a lot of the same colo providers we've worked with in the past, and building out the platform to be able to take on all sorts of different chipsets, both CPU and GPU. It's been a relatively light lift in terms of investment in people and technology. It's been really just doing a lot more of what we're doing at a bigger scale. Tell us a little bit more about how you go through a network upgrade cycle. What I'm getting at here is we sometimes hear the bear case that Akamai's network is built over multiple decades. It has a legacy feel to it, therefore it's not going to meet the performance and cost requirements of a frontier model today. Tell us how that statement is wrong. Yeah. First of all, most of the 4,000 pops we wouldn't use for a frontier model. Maybe if they were using, say, there was a CDN application that was going alongside that, sure, we could use the- Different use case. Exactly. Very different use case. In terms of what we're buying today for building out, let's say a customer comes to us and says they want 5,000 GPU of various shapes and sizes, or they want 10 million virtual CPUs or whatever, we're using the latest hardware, latest chipsets, et cetera. We have the visibility of getting a big order for multiple years, and it's just the most modern equipment that's necessary that performs very well. To the extent that it's at all connected to the network, if they're, say, running some code in the servers that we have for managed container service, that's all modern and usable today. We don't need to go and upgrade the network. There's no notion of saying, "Hey, we have to throw out what we've done and rebuild something else." We're just adding locations with more servers in them. Makes sense. Let me ask you one more here, and then I'll turn it over to Max. When you sign these big initial commits with customers, you've given us a couple of nuggets on the unit economics, but I find what's interesting is this could be the beginning of a much longer-term relationship where you cross-sell into these initial customers that have signed up on the compute side. Yep. What does that customer journey look like for maybe take the two larger deals that you've announced today? Yeah. It's interesting, when you sign one of these big deals, it creates this sort of network effect where we get a lot of inbounds today. The pipeline is bigger than I've ever seen. I've never seen anything like this, where people realize, "Okay, you are an option at a scale that I didn't think you operated at." Once you get in, start doing work with any of these bigger companies, like the biggest companies that we're working with, foundation models as well as others, you uncover a lot of other opportunities, right? I think what people don't quite understand about us is that our business isn't just GPUs. That's not all that's growing. As a matter of fact, the pipeline is mostly CPU. A lot of our revenue today is CPU generated. It's a little bit different in terms of the unit economics, power usage, and things like that. What goes along with these, even the big investments in GPUs, there's always some drag along for CPU and storage. My CTO said to me today, think of the GPU as the brain, but the CPUs are all the arms, the legs, the tooling, et cetera. These systems have a lot of subsystems. I might be working with someone who is spending billions with one of the hyperscalers, but I may carve off a piece of a support system to run several hundred million on CPU. As you start working with these customers, you find more and more opportunities to grow with them. Over time, I think there's opportunity to cross-sell security. A lot of the customers we're working with today are already existing customers of security and delivery. The demand is coming from new logos as well as existing customers, and expansion orders with some of the customers that we've already done some compute business with already. You've now signed over $2.8 billion in multi-year cloud infrastructure commitments. What are the most important milestones, operational milestones between signing those deals and then seeing them show up in revenue? Yeah, great question. It's funny, the way we're managing the business now is pretty interesting. We start with the sort of three components, right? There's co-location or power, there's capital, and there's equipment. Each one of them has a little interesting twist and dynamic to it. We're very fortunate that capital is not an issue for us now. We have plenty of capital. We have the ability to raise capital if we need it. That's not a big problem for us. Co-location is probably the most challenging issue. Power and space is probably the area where that is the biggest blocker to near-term growth, and nobody has a lot of power and space just laying around. We have a very different model, where we work with hundreds of colo providers around the globe, and we're actually a very interesting buyer for them, because one, we're investment-grade credit, and two, we buy a fair amount of colo. If someone's like some of the major colo providers are building five or six sites across the country, we may take 20 or 30 MW from them in five or six different locations or whatever. We can make long-term commitments of good financial backing, and we're able to cobble together in multiple locations, stitching that all together with our backbone in areas that might be cheaper than some of the major cities and that sort of thing. We're not building these gigawatt facilities or anything like that. Getting power is step number one. I've got a lot of demand right now. What I do is I sit down with my operations team and say, "Let's map out how much power we can light up between now and, say, the end of 2027, and then give me a look at what 2028 looks like." We'll start to map that out by month and by quarter. Then I'll go back to the customers who will come to me and say, "Ed, I want to do 5,000 GPU. I need it to perform like this. I want to use this chipset. I want to do this or that, and I'd like to have that up as fast as possible." I say, "Okay, I can slot you into these locations over this time period." Usually, you get an agreement. It might take you 90 days or so to hammer out an agreement. At the same time you lock in your colo, you're now locking in your supply chain. I'll be ordering the chips, be putting together the equipment. We use contract manufacturers to put all that stuff together. Sometimes you're buying, say, like a total NVIDIA stack. Other times, you're just using CPU and putting the pieces together. We'll make sure that if there's any delay or any sort of a time difference where we're, say, risking memory prices going up, we factor that into our agreement. It gives us an enormous visibility to where we can start to slot in some of these bigger opportunities. I'm making my colo buying decision based on what my pipeline looks like. I want to secure. Let's say I'm going to sign 100 MW, say, in round figures. I'd like to get, say, 70%-80% or greater of that secured with deals behind it before I place that order. I'll leave some amount, 10 or 20 MW or whatever, for stuff that's not done or maybe there's options that I have to take additional colo. I try to keep that as tight as possible because that's the one area where that can do the most damage to your P&L near term. If you think about signing the deal, ordering the equipment, getting the colo, probably takes six-nine months between signing the deal and getting to revenue. You start to take the expense for colo the minute you get the keys to the facility. Depreciation can be pretty tightly linked to revenue, so there's a bit of a delay. If I take on 100 MW of power, I'm going to have a pretty big expense. Even if it's for 90 days, it can put a point or two of pressure on my margin, so I try to keep that as tight as possible. It's really about as soon as I light up power and I know when the power's coming in, I can start to slot deals into that. Operationally, it's a fairly easy thing to run, and power is definitely the one area that is sort of the gating factor, if you will. Is there any color you can provide us on the CIS pipeline beyond the $2.8 billion committed amount right now? Is this the limiting factor, mainly colo and power, or is there anything else that would hold you back from onboarding more customers? Yeah. I would say we have line of sight to add a significant amount of colo between now and the end of 2027, and even more in 2028. I would say that's not necessarily a big factor as far as prosecuting the pipeline. I've got a very robust pipeline. You need to go and qualify it. Some stuff that's in there doesn't meet our margin requirements, might not be the right fit. You don't want to take too much of a bet on startup companies, for example, that might have a different financial profile, especially if I've got enterprises in there that are better credit quality and that sort of thing. The near-term problem I have here is that I don't have a ton of colo just sitting around. There's the growth. We've already talked about having growth accelerating from the 6%, 7% now to the teens for next year, and that's with what we have signed today. I still have the opportunity to sign more, but now I'm starting to build stuff that will come on late in 2027 and into 2028. I see really no issue as far as having this massive pipeline. It's just a question of how quickly you can get the colo lit up, and then there's backlog for computer equipment. It takes some time to get certain equipment. Let's say, for example, I wanted to sign Rubin c hips, I can start taking order now, but I won't get the chips for at least six months to a year. They're just not physically available. We're in that period of time now where we are signing up big deals. Some of them are already inked, and we've already announced them. They're starting to produce revenue in Q4 and will ramp up throughout 2027. A lot of that's in the first half of the year, and we're starting to sign up additional deals to put more revenue into 2027 and a lot of revenue into 2028. That makes sense. You talked about a meaningful acceleration in CIS revenue starting in Q4 and then into 2027. Are there any directional clues you can give us to help us model this ramp from the outside in? Yeah. We've given a little bit of a heads up in terms of some of the bigger deals we've announced and how much revenue they'll produce. We're expecting $15 million from one of them and somewhere around $20 million for the other. So that will significantly increase the growth rate for Q4. There's some risk that if stuff comes in a week or two late, maybe it pushes a couple of weeks, but generally speaking, you're going to start to see a big ramp here, and then it continues to ramp into 2027 as we get more equipment and we get more power lit up. But it's all structured that we should start to see a significant acceleration in Q1 and then into Q2 and a little bit into Q3. Understood. The upside from AI compute is clearly large, but investors are also focused on protecting the downside. So how are these contracts structured to protect if any of these deals continue to slip? Or you mentioned that memory costs, you have hedged against that, but what about other delays? How are you protecting that side? Yeah. Generally speaking, what you do both on the colo side and on the customer contract side, if you're signing, let's say, a four-year deal or a seven-year deal, you will build in some type of an escalator to cover costs. Labor costs go up, co-location costs go up, and with our colo deals, we'll do the same thing. So I might have a 2% or 3% escalator each year. The way the accounting rules work, you have to straight line that. You also have to straight line your revenue. So it just comes out as flat across the period of time. You don't have that bump in terms of revenue and costs and that sort of thing. But that's generally locked in, so there's really no risk of that blowing up on me. If something comes in, say, a few weeks late, let's say there's a delay in construction on a facility for colo, and we expect to start on October 1st and it starts on November 1st. All that does is just start the revenue clock later. There's no penalties. The deal doesn't get shortened or anything like that. It's just literally, you're working very collaboratively with your customer saying, "This is the best estimates we have from the construction site to the manufacturer in terms of when we get the equipment." As soon as the equipment is lit up and it's available for the customer to use, these deals generally are take or pay, right? As soon as the capacity is available, you start billing for the amount of capacity that's available. Sometimes it might be a ramp of, say, three or six months to get to full ramp. Other times, you can do it in a much shorter period of time. If I'm building out, say, a big deployment, and I say, once I sign the contract, I'm signing a contract with the manufacturers and with the colo providers. Colo is easy. Usually, you can get that locked in, and I can actually say in the contract, "Here are the three or four or five or six locations we're going to put you in, and if there's any kind of an issue, maybe there's a backup location of where I might put you in." With the equipment, let's say I can get 80% of it contracted with all the memory and everything like that locked in the day I sign the contract, and now there's this 20% or so that could potentially flux. Memory being the biggest driver of that. You work something in the agreement where you would say, "Look, if the prices go up by a certain percentage, we just pass it straight on. If it goes up by, say, 50% or more, we'll have a conversation and maybe we put a pause and wait a few months, or maybe we just continue to build and we just pass the price on." The concept is that we've sort of agreed on what the "margin" looks like at a high level. In other words, how much pretty transparent with what our costs are, that if the component parts go up in price, we are going to capture that. Once I have that locked in, now I've got the hardware, I'm just appreciating that the cost is already set. The colo, we've taken that into consideration with building in slight revenue ramps as well. So we have definitely modeled in the "risk or downside" as we're signing these big deals. To the extent that I'm building ahead of demand, I have the ability to raise my price. So let's say if component parts go up or if colo costs go up and I have to secure at a higher price, I'm just going to raise my price. We've done that. We've actually raised prices in some cases. Got it. Historically, you talked about CIS resulting from $1 in CapEx Turning into $1 of revenue over time. Yes. Is that still the right way to look at this? Yeah. What is the timeline to get there? Yeah, good question. When we first bought Linode, that certainly was the case. We would see that over time, you spend $1 in CapEx, and you get about $1 of recurring revenue. I think initially when we started, there was some confusion because we were our first biggest customer. If you looked at our CapEx and you looked at our revenue, you would say, "Well, wait a minute, Ed. I am not seeing that manifest itself in revenue." Keep in mind, we are spending well over $100 million on third-party cloud, and we brought a lot of it in-house. I guess if I translated that into revenue, if I were a paying customer, it would probably work out to be pretty close to that. With some of these bigger deals, we are not necessarily getting a dollar for dollar. It is generally between, say, $0.50 and $1. There may be an occasion where I might take something a little bit less, but I am doing that keeping in mind that there is sort of two main drivers of cost. There is power and there is equipment. If I am willing to take a slightly lower yield, that means I am doing much better on the power throughput, right? I am getting significantly better dollars per megawatt of power in a situation like that where I have slightly higher depreciation. But I gave you on the last call some margin guidelines. As we look at these big deals, I am getting somewhere in the, call it mid-20s on the low end on the operating margin to mid-30s on the high end. That is for the bigger stuff. For the regular way business, it is even better than that, right? Because customers do not have the purchasing power. They are not locking in for four to seven years. You also have to assign some value to the relationship as well. There are strategic deals where you might say, "I am willing to maybe go a little bit lower on that margin scale in the 20s because I get a relationship with customer X, and I have always wanted. I think I can either make up that margin with other products down the road, or it is just an opportunity to get significantly more business. Ed, I'm curious on this point. We have so many data points that are a little bit apples to oranges on ARR per megawatt, revenue per megawatt, GPU payback period across the hyperscalers, across the Neoclouds. I'm curious when you do your internal benchmarking how you think about your price point relative to all of these other data points Yeah floating out there on the ether. Yeah. Obviously, with the bigger deals, it's a lot easier to do the math, right? Yeah. You can say, generally speaking, you are getting $15 million to $20 million per megawatt. There are cases where with some CPU deals, you might get $25 million or better. Right. You could even see sometimes in the 30s with some of the smaller customers and things like that. We will not take something that is significantly lower than, say, 15. Maybe there might be some strategic reason you might take something slightly lower than that, but generally speaking, you are somewhere in that range. Now, initially, there can be some build-out where you would say, "I need to build out for spares and some excess capacity." So if you look at some of the metrics, you might say, "Hey, you are putting an extra 5% into this deal to build out some headroom. Sure. Now, as you get bigger and you have more deployments and Yeah locations, that comes down over time, so I would expect your revenue yield to be a little bit better. Also, not every deal behaves the same, right? So I could have two different customers that are using two different chipsets. Let us say they are using Compute. I might get a better yield on my revenue per megawatt from a particular customer just because of how efficient the chipsets are and the servers are and that sort of thing. But generally speaking, we try to land somewhere in that range. If you think about cost, the cost of a megawatt of power is, say, $3 million- $4 million a year. It is sort of on the higher end, I would say, in the U.S. I am sort of building in a little cushion there for inflation and that sort of thing. Your depreciation can run $3 million-$5 million, something in that range. As you work through it, you can see where you get to those margin targets. Right Now, there's some other ancillary costs in there. There's a little bit of people cost. There's maybe a little bit of, say, maintenance cost for your switches and your equipment and that sort of stuff. But generally speaking, the bulk of the cost is colocation, and by far, the biggest cost is depreciation. One of the debates we've been having actively in the last few weeks is we're in a period of time where supply and demand is just so tight. What happens to that $15-$20 when the industry potentially goes into excess supply? How does Akamai think about Yeah their role in perhaps de-risking some of these contracts in Yeah an eventual excess supply environment? Yeah. It is funny. If you look at the history, you just do some research and just say, ask one of the big AI models to do it for you, and say, what is the hyperscalers typically got? What do the other folks get out there? You see that that range is pretty consistent, right? Yeah. It is sort of held up historically. Could you get into a situation? Maybe we decide there is business we are not going to take. Also, we have a value proposition as you think about the next generation of where this is going, where there is a lot of infrastructure going out for building models. You have these open source models. You have a lot of post-training going on. You have now just starting to see some of these agentic applications that are being built that will use a lot of compute, but not in massive amounts, right? You might have someone who used to have a website now has this agentic agent that is being like a travel agent or a personal shopper or something like that. Well, that is going to require a lot of compute, and there, latency really does matter. I think you're going to get a higher yield for providing that type of a service, right? Maybe the model switches, and we're not going after some of this if Right the yields don't make sense. Here's the interesting thing. Right now, I can see demand into 2028, right? I'm able to park business out into 2028 today, which is very unusual, right? That's sort of the way the industry is. Nobody's sitting around saying, "Hey, I've got X amount of power available. I can deliver it to you on Friday. I can get you all the equipment." It's not sitting around. Now, will that work itself out over time? Probably. But I think it's several years out. I don't see this being like a 2027 or a 2028 phenomenon. One more on CIS before we move on to some other topics. One of the advantages for Akamai is clearly the distributed network, but how do you think about the balance between centralized GPU deployments and more distributed inference at the edge? What customer use cases are actually ready for edge inference today versus still more of a longer-term vision? Right now, I would say anything that has a latency requirement like robotics or autonomous driving, anything that has to do with interacting with the consumer where you don't want to introduce latency into the experience would be anything that could be effectively a use case for that more distributed type platform. There's also an element of performance too that comes into play where customers will have an expectation of investing a dollar and getting a return out of that equipment, whether it's a number of actions that they can produce per dollar of spend, et cetera. So that's another element of performance, and in some cases, we can outperform competitors just based on the economics that they can yield for however they're looking at how their platform runs on a particular set of machines and equipment and that sort of thing. Also, some cases for inference, you do not necessarily need GPUs either. You can run it with CPUs, right? We are perfectly positioned for that, and I would say we are just starting to see more enterprises heading in that direction. I think so far the dollars have been primarily in building the big training models and all of the subsystems that go along with it. So that need for more of a centralized, but distributed, meaning in a handful of locations, not in tens or hundreds. Do you have a latest update on how many locations you have deployed GPUs thus far? You also have RTX, like lower-end GPUs deployed. Are these also included in these? Yes. I forget the number we last quoted. It might have been 20 something, might be the number. That will just go up based on customer demand, right? If somebody wants to go more, there is nothing preventing us technologically. As I talked about with colo, we can get colo in many locations, so it would be fairly easy for us to increase that. It would just be based on customer demand. Got it. Moving on to capital allocation. You raised convertible debt earlier this year. You still have a very strong balance sheet. As the pipeline grows, how do you think about the right funding mix between cash, future debt capacity, and then internal cash generation across the business? Yeah. I want to fund as much of it as I can out of my cash flow. It is interesting, if you model this business out and you say, just do a fun exercise and double the size of the company with just CIS revenue, and use the metrics we gave you, and then say, at the end of that, just go to say the company goes to 10% growth or something like that, and watch what happens to the margins. It is fascinating. EBITDA margin expands dramatically. Free cash flow expands dramatically. If you want to grow at 10% and say you are twice the size of the company, you might need 20 something percent of CapEx to grow at that level. So EBITDA at a 50% yield, 50 cents on the dollar for CIS. The multiples will expand, the free cash flow will expand, EBITDA margins will expand. So higher EBITDA margins give us a lot more ability to raise debt if we needed to do that. We are also going to produce a lot more cash. These models are great. Once you spend the money, you get the payback relatively quickly on your CapEx. Sometimes it is two years, sometimes it is three years, and they are very high EBITDA margin, very high free cash flow margin after you spend the initial capital. So we will be able to fund a lot of this growth in the future from our cash. So far, the convertible market has been very good to us. We do want to maintain an investment-grade credit rating to the extent that we can, and the only reason for that is it is an advantage when you are dealing with colo providers. If we lose it is not the end of the world. I think if we lose it would be very temporary because of the EBITDA margins will expand very quickly thereafter. Sometimes you see the rating agencies, if you do an acquisition and you have a thesis where you drive synergy, they will give you credit, and they might put you on negative watch, but you get back to investment-grade or stay at investment-grade. So that will be our goal. If not, I have to put maybe some deposits down when I do some of these colo deals or get letters of credit, so it is a little bit more costly, but not the end of the world. But our goal would be to try to keep that. So far, debt has been the right instrument for us in terms of raising capital. Moving on to security. AI seems to be helping some of Akamai's more mature security products, including web application firewalls, DDoS protection. Do you view this as a durable re-acceleration in that more mature security product category, or is this more so a one-time upgrade cycle that you are seeing? Yeah. It's interesting. I talk about it as like demand at the top of the funnel has increased dramatically ever since Mythos has been released, that CISOs are bringing all their vendors in and saying, "Okay, what am I doing today? Am I utilizing everything like I should?" For us, an example would be, hey, I'm using WAF for all my public-facing applications, but not my, say, supply chain or my, say, private network client portals and things like that. Well, they're just as susceptible to these types of attacks. I need to put WAF on every application. We're seeing a big jump in bot management, demand for bot management. What's happening with these AI models is it's creating a lot of machine traffic. So understanding what is that machine, and then what action do you want to take? Do you want to block it? Do you want to send it different information? Do you want to keep it away from the crown jewels, maybe keep it away from the paywall, but let it get other free information? So there's a lot of opportunity for that business to grow, and I think that will continue. Then for newer products like LayerX, there was that announcement today from, was it Meta around that whole, what's it called? Muse, I think, where they're creating basically the ability for individuals to have their personal agents do things. Did I say that? Somebody told me about that in one of our meetings. Anyway, what that does now is it creates a big threat vector for all the endpoints. So LayerX actually can help defend endpoints from, say, data exfiltration or data loss and that sort of thing. So I think there's an opportunity for that as agentic applications grow. Then for Noname Security, with our API security, with these AI models, there's a lot of back and forth. It might not be API, but maybe MCP is the protocol. That also creates a big threat vector, and I think that we're perfectly positioned to try to leverage that to solve that problem. On that point, across the security portfolio, where do you think the biggest tailwind is from AI? Is it the products tied to CDN, Zero Trust, or API security? Yeah, it's a little of everything, really. For example, Guardicore. You might be deploying Guardicore and you say, "Okay, the benefit of Guardicore is once something gets in, you limit the damage. You can set up rules in your network to segment things that if a virus were to get in, it will block certain bad things from happening." And you might say, "Hey, I'm going to roll that out to a piece of my network today and then slowly roll it out over time." This puts more pressure on the CISOs to have that, if that's the way they believe in the last sort of layer of protection, to roll that out faster. I think Guardicore gets a leg up. As we talked about, LayerX has got a big opportunity initially with just protecting the browser and the user from taking the company information and sharing it out with either an open source model or one of the foundation models. You can use LayerX to prevent that from happening. Then we're working on with LayerX and with Noname and some other technologies, creating the ability to protect the cloud deployments for agentic applications running in the cloud, not just at the endpoint. Ed, you made a really good comment earlier in this conversation about Akamai's been through a cycle like this before in 2000, 2001. And you're one of the few management teams that actually have that tenure under your belt. What do you think is different this time? Well, obviously it's a lot easier than the equation we talked about in terms of having capital, being in a great position to be able to start a business like that is more capital-intensive. Right. Back then, we had a metric called QTL, quarters to live. Because you were burning cash, right? Here we are in a position where we are a profitable company making an investment. I think we are being very smart about the customers we are working with, the amount of capacity we are adding, trying to keep that supply chain as tight as possible, right? And knowing that, look, these things, they do not live forever. There will be some. At some point, things will slow down. We do not want to overextend ourselves. We do want to invest in growth. I think that definitely helps you. But also we are leveraging, we are monetizing a part of our business, which was our ability to scale and build a network. That now is a very valuable skill. The relationship we have with our colo providers has provided us this sort of niche where we are really the only ones out there working in this model of cobbling together, across many colo providers, an interesting fabric where you can do billions of dollars of CIS revenue, right, without having to go and build your own data centers. I think it is sort of leveraging all that experience and then the 20-plus years of enterprise relationships with customers. That is a huge advantage. You hear about some of the Neoclouds or even DigitalOcean saying they want to get to the enterprise customer. That is hard to do. We have built up 20 years of trusted relationships, and so we are already having conversations with customers that say, "Hey, I want to build my own open source model and run it," or, "I have some agentic application that I want to build. How do I think about performance? What can you do for me?" So it gives us a significant advantage across many different vectors. Ed, well said. Congratulations. Thank you. on your milestone. Please join me in thanking Ed for his time. Thank you.