Cloudflare CEO Predicts AI Agents Will Outnumber Humans 1,000-to-1

0:00Matthew Prince:How are you doing? Matthew Prince, Chairman, Good to see you. I'm doing well. Great to see you. What's new in your world? You made some acquisitions. You made an acquisition. Take us through it. Yeah. So we bought a company called VoidZero, which makes Veet, which is one of the most popular developer platforms that's there. It's just an incredible team. Evan Yu, who's the founder, is just a first-class human being, someone who our team is super excited to work with. the team that he's assembled is just great. And I think that this is increasingly becoming the platform that is being used to power a lot of the agents that are running around the internet.

0:37And a lot of those agents are running on Cloudflare. And so we think it's just a really natural combination.

0:43Matthew Prince:How simple is the synergy? Is it you'll funnel those 130 million users who download Vite every month? I think it's 130 million weekly downloads. Weekly, wow, yeah, that's a lot. into preferring Cloudflare, defaulting to Cloudflare? Or is there more synergy under the hood around developer integration, company integration? How are you thinking about this playing out? Yeah, we plan to continue to leave it as an open source project and support it and invest in it that way. We want to integrate it closely with Cloudflare's developer platform and make sure that Cloudflare is the best place to run any sort of VEAT project that you have.

1:28But it'll work in any of the different platforms as well. And so we just really wanted to make sure that Evan and his team have the support to make sure they could continue to really invest in what's just an incredible platform. And we think that that is going to drive more developers to Cloudflare's workers' platform as well.

1:47Matthew Prince:What are the headaches for developers these days? I know everyone's concerned about token costs and token budget. Sometimes that doesn't show up for the developer. It's more like the CEO that's worried about it. But is uptime more difficult to maintain? We've been seeing screenshots of different uptime tracker status pages that have more and more red and yellow on them. Like what's at the top of the stack and how are you helping? Yeah, you know, I think that you need a different architecture than we've had to build sort of the last generation of applications for what's coming with agents. If you imagine, you know, there are about 100 million knowledge workers in the United States.

2:27if all of those knowledge workers had one agent, which was working on their behalf, and each of those agents was running in a container, a traditional container, like something you'd get at AWS or a Google Cloud. The amount of just CPU resources that would be needed to run just those agents, assuming they just had one agent per person, is about 50 % of the total CPU capacity that's generated by all the different CPU manufacturers that are out there today. And that's just the United States knowledge workers. If you take that to the world, then it's several times, 30 or 40 times the capacity of GPUs and CPUs that are existing today.

3:14And so what we really think is that as these agents are creating code, you need a different platform for it. And Cloudflare was built, Cloudflare Workers was built, not on containers as much, but on something called isolates, which is a much more efficient technology. And so what we're seeing is that as people are building these agents, as they're using them, it's just a much more natural place to be running them. And it's, I think, why you're seeing more and more of the big AI labs have Cloudflare as the preferred target for where their code gets run. OpenAI released a project for their enterprise users a little while ago that, again, is targeting us.

3:51And we want to make sure that with things like Vite as first-class citizens on Cloudflare, that we can help power that future. Because, again, it's not going to be kind of the same system that we built with the hyperscalers. It's going to be something different. And, again, I think that we have a really good shot of building that different thing.

4:08Matthew Prince:So can you help me understand more about the problems of CPU bottlenecks and then maybe some of the solutions? I'm just thinking about, I would think bandwidth would be an issue, obviously CPU load. But is there a world where we get to some sort of convention around, maybe it's the robots.txt or just the user agent. And when Googlebot shows up or any other AI agent shows up, you're just delivering a, it's almost an MCP server. but you're just delivering something that looks more like a JSON package, something that's a little lighter, a little less CPU intensive. Is there a path there to optimization?

4:46Matthew Prince:It feels like we're in the era of like squeeze every ounce of performance out of everything. But what's actually going to happen here? Are we just screwed or is there an opportunity? No, I think two different problems. So the first is, as you ask these different AI systems to perform tasks on your behalf, what has to happen behind the scenes, especially as those tasks get complicated, is that they need to be coordinated in some way. So if you say, plan a vacation for me or something like that, what goes on behind the scenes is that there's coordination there. And the best way to make that as efficient as possible, what agents are really good at and these various LLMs are really good at is actually writing code.

5:34And so that code needs somewhere to live. And the problem is that if that lives in a container, then you've got to bring in an operating system. You've got to bring in all the tooling and everything else. And that's actually extremely heavy weight. And so that's the first place that you've got both a CPU and a GPU bottleneck that's there. And so with something like Cloudflare workers and isolates, it's just a much lighter weight system. And so that means that you can have more agents running on the same CPU infrastructure and be able to provide that. And again, I think that's why a lot of these next generation tools are actually built using our platform.

6:08You mentioned something else, which is, you know, as these agents go out and interact with the rest of the web, you want to make sure that that is done in the most efficient way. And so if they're pulling down, you know, all of the HTML from a web page and those web pages are designed kind of to be consumed by humans, there's just a lot of cruft on that that isn't necessarily as important. And so some of the things that we're doing are, you know, for those customers of ours that want to make sure their content is consumed by agents, want to make sure it's as accessible as possible, We are automatically converting things into Markdown, which is a much simpler system.

6:48That saves you a ton of tokens. It saves you a ton of processing. It means that your context window can fit more useful information into a context window. So I think there's a ton of optimizations. And we're helping both on the developer side as well as on the content side, making sure that we can have these agents be as powerful as possible and get as much done as possible.

7:11Matthew Prince:John Gruber, total victory. You know he invented Markdown? Grubinator. No way. Yeah. Isn't that amazing? Yeah. I mean, it's one of these things that just, you know, it turned out, it was ahead of its time. But it's such a key for making sure that we can take information and make it as compact as possible. John Gruber created the format for God. It's a funny, funny world we live in. Talk about, it was last week you guys announced that the threshold had been passed around agent versus human traffic. Talk about that moment. Did it happen sooner or later than you expected? Much sooner. Much sooner.

7:51I mean, it was, I, at the end of last year, so the end of 2025, I said that I thought that we would pass, he bought traffic. and that's, you know, across the board. So that's like Google's crawler, but also, you know, the new agents which are coming out, I thought would pass human traffic by the end of 2027. About three months ago, I revised that based on the traffic that we were seeing to say that it would actually be in the first half of 2027. And so the fact that it actually happened in the first half of 2026 has just been extraordinary. And it just shows how quickly this is growing. And the real key here is that I think that if you or I as humans were researching to go buy a digital camera or something, we might visit five websites and do a little bit of research and some price comparison.

8:44You just watch as you use these agents. They have boundless attention to be able to just go to maybe 5 ,000 websites to find exactly what you're looking for, the best price, the best delivery, the best service, and everything that's there. And so that's just driving an enormous amount of consumption of the Internet. At the same time, the other thing that's happening is that for a long time, since about 2015, the web has kind of plateaued. There were more websites that were being shut down than were being created during that time. In the last 18 months, though, we're back to the web growing at a rate which is exponential.

9:27And it looks sort of similar to what was happening back in the early 2000s in terms of growth of the web. And so I think you're seeing both more sort of consumption of what's online, but you're also seeing more things online as it goes forward. And so in both of those directions, you know, that's going to continue to just drive, you know, more and more use of the Internet. And I wouldn't be surprised if going forward, say, five years, that bot traffic will be a thousand times human traffic online. And we've got to make sure that we make the internet work for that new future.

10:04Matthew Prince:Yeah. Like every year, we're just going to add another 99.999 % of internet traffic as bots. Do you know what the baseline was? Like pre-AI, were we at like 1 % bot traffic? I mean, I've said it was more than that. It was about 20 % for a long time. So yes, Google was the largest. And then obviously there's a lot of malicious things that run around online, but that was around what it was for. And it was pretty stable over the history of Cloudflare, at least, where we could measure it. So since 2010, it was always kind of around that 20 % range. And then it's grown. In the last few years, it started growing steadily, and then it really accelerated in the first half of 2026.

10:55Matthew Prince:I want to talk about the inference stack, I guess. We're seeing two things play out, sort of a bifurcation. At WWDC, Apple's launching on-device inference that's not going to be frontier, but it's going to be usable for sure on your phone, on a computer. They can obviously go to their private cloud as well. And then you have the new NVL72 models. There's stuff in between on a Cerebris chip, fast but expensive. Then there's everything from, oh, it runs on commodity hardware, but it's still pretty big. You need a real desktop for it. And I'm wondering about how you see Cloudflare fitting into that.

11:36Matthew Prince:It would be a logical extension to Cloudflare workers to have inference on the edge, inference in different places? Like, where do you see yourself offering inference, if at all? Yeah, well, it's actually kind of funny. Back in 2022, we issued a press release that there was a graphics chips company that we were going to partner with in order to put GPUs at the edge of our network. And that graphics chip company turned out to be Invidia. And what was amazing was at the time, it was just crickets. Nobody cared. But we sort of did a little of it, but we hadn't really rolled it out broadly. And what was funny was, and then you fast forward two years later, and all of a sudden everyone cared.

12:29And so we basically just reissued the exact same press release. And now that's become a pretty important part of our business. And so today you can run inference at the edge of Cloudflare. And because of the fact that we're in over 350 cities worldwide, we're within milliseconds of the vast majority of the world's population, we become a very natural place for inference to happen. My working assumption has always been that about 50 % of the inference that happens will be on device, whether that's your phone or your laptop or whatever. but that there needs to be some standard protocol where your phone or your laptop or whatever that local thing is can hand those either long-running tasks or larger tasks off next to the network.

13:14And so that would be to us. And then if for some reason you need something that's more than that, that you could hand it back to some centralized data center with more capacity than we may have. And I think that that's what we're increasingly seeing. I think my assumption was a little bit probably off. I actually think less is going to happen on device today because I think more and more of the tasks are going to be these long-running tasks where it's not going to be just, you know, what's the temperature in New York today? Instead, it's going to be something like help me plan a vacation, go take into account all of these different things, shop for different hotels in different places, plan all the travel between the different locations.

14:02Here are the criteria that I have. And that might be something that takes maybe – it's not going to be seconds. It might be minutes or hours or days in some cases to run in that case. I feel like you're going to have long-running agents just for Park City snow forecasting and reporting, tracking like individual runs, you know, maybe a network of drones that are identifying what's tracked out, what's not. Satellite photos, Planet Labs. Optimizing your routes on the mountain.

14:32Matthew Prince:A lot of stuff. That's exactly right. Hopefully it snows this year, unlike last year. So what are the decisions when you're doing inference on the edge? Because it's sort of a hard pitch to say I'm going to shave off 300 milliseconds, 600 milliseconds when you're talking about a 20 minute workflow or even a 10 second workflow. You're getting me a 3 % boost. Is that going to make the difference? But are you optimistic that there will be sort of like a in between state where there will be inference that happens and it needs to happen within the request loop under a second and then the latency matters?

15:14Yeah, well I think there are sort of three different buckets, why people are choosing to run inference tasks on Cloudflare. I think the first bucket is exactly what you said, it's just it's latency. And so especially for those things that have human and computer interaction where any delay feels like it hurts you, that people are trying to make sure that they can get the best performance possible.

15:39Matthew Prince:For voice models, I Voice models have got to be local. Yeah, I love that. Yeah, and so, you know, again, there's a case for that. I think the second case is that a lot of times for either privacy or regulatory reasons, people want to keep things as close to where they physically are as possible. And so, you know, I think that a lot of, especially in Europe and otherwise, of places think that, you know, they made a mistake with the Internet kind of 1.0 and 2.0 of having everything go back to Ashburn, Virginia, everything go back to the United States. And so I think that there's a real kind of sovereign desire to keep things local, and that's important.

16:17And then I think the third thing is we actually can be oftentimes much more cost effective because of the fact that we are a network, and because of how we're interconnected with networks around the world, bandwidth for us is effectively free. And so it's very easy for us to get things onto our network. And because of where we deploy systems, we often are in places where we don't have to pay for the space or the power, which allows us to then pass those savings on to customers. So it can be significantly more cost effective to use inference with us than it can be in some other places. places. That's somewhat counterintuitive, but it's sort of the nature of what we've built and kind of the secret to what's allowed Cloudflare to deliver as much as we have over time.

17:01Matthew Prince:There's a ton of debate over data centers, how they get built, how big they should be, what we need, the pushback, regulation, etc. Is any of this affecting your business? Are you thinking about how you position your footprint in the real world? How are you processing the evolving discussion around data center construction? Yeah. So first of all, we're going to need more data centers. I think a lot of the discussion is somewhat silliness. The water consumption, these are closed loop systems. A golf course uses more water than probably all the data centers in United States over the course of a year.

17:45So I think there's a lot of kind of silly concerns that are there, but there are real concerns as well. We've got to make sure that there are efficient ways to get power to these things, that it's not taking power from other systems that the grid can support that. And so I think those are all good considerations. For Cloudflare ourselves, You know, we're a little bit different where in the case of, you know, if you're an AWS or a Google or, you know, a SpaceX, you're building one big facility or a handful of very big facilities and putting a ton of machines in that one facility. Cloudfire is different.

18:24We have a ton of machines, but we spread them massively all around the world. And oftentimes we want to go into the places where networks come together. And so those can be some of the oldest data centers in the world. And in any of those facilities, we may have only hundreds of machines. But collectively around the world, we've got what is effectively much larger than any individual data center, which is out there. So I think we're less impacted from the new builds of data centers. I think we're much more impacted by how do we find our way into the existing data centers? And then how do we make sure that the equipment that we're deploying is as power efficient as possible?

19:06Because we're often given sort of here's the envelope that you have to fit your equipment in. And we're often guests of whoever the local ISP is or whoever partner is that we're working with in the thousands of buildings that we're in all around the world. What's new with Italy? Yeah, Italy and Spain.

19:28Matthew Prince:Like it's it's it's it's you're so you're fighting Italy your summer plans your summer plans are limited duty. No, no, no Ibiza for me or multi coast. No, it's it's, you know, I think it's interesting. It's been actually kind of puzzling for us. So all of this stems back. So for those who don't know in both Italy and in Spain, there's been pretty aggressive. tactics to come after either me personally or Cloudflare. And it's all been driven largely by the football leagues, the soccer league in those places. And what they're concerned about is piracy. The thing that's been ironic about it is, we don't like piracy on our network either.

20:16We don't make any money off of it. It takes bandwidth, it steals resources. And so we have a whole team that's working all the time to shut the pirates down. And yet, the pirates are clever. They find ways to use our system. And because, you know, a huge percentage of the Internet sits behind Cloudflare, you know, there's going to be stuff from time to time that we don't we don't capture. It takes us some time to catch with leagues everywhere else in the world. So with the NBA and NFL and MLB and the Premier League in the UK and others, we work really closely with them to if they identify a pirate stream for us to pull that down.

20:52Because, again, it costs us money. We don't like it. But for whatever reason, La Liga in Spain and the league in Italy have decided that instead of partnering with us to take this down, they would sue us instead. And so, again, yes, it does cramp some of my summer plans.

21:13Matthew Prince:I want to talk about IPOs. Life is a public company. There's a few entrepreneurs that are trying to go public this year. You had a very successful IPO. The stock is up more than 100 ,000 basis points since you went public. Fantastic run. We're putting everything in basis points now because it sounds even better. The stock's up 1 ,000%. It's fantastic. But 100 ,000 basis points sounds even better. 100 ,000 basis points is a lot. 110 ,000 basis points. Who's counting? But anyway, what was your process like? We were talking about the level of float lockups. What were the hard decisions? Were there things that we're talking about now with these big IPOs that weren't even on the top 10 of your to-do list?

22:03Matthew Prince:What is important to a successful IPO and then running a public company? Yeah, I think there are a handful of things that are interesting lessons from it. First of all, I love being a public company. It's, you know, I think that it's interesting, you know, that in jurisdictions where you don't have no fault divorce, spousal homicides are much higher, which is, I think, kind of an analogy to the difference between private market investors and public market investors. it's really hard to fire your VCs. It's hard for your VCs to fire you. And so as a result, like actually it's kind of dysfunctional.

22:48Whereas public market, like I love our investors because if I do something stupid, they call me up and they say, that was stupid. I sold your stock. And then we, and we have a conversation and sometimes they say, oh, actually that makes sense now. And I bought it back. Right. But I think you can have actually a much more honest, much more real conversation. And, um, and, and that's, and that that's felt actually just a lot healthier than what you see in the private market. Yeah, it's so true. I mean, everybody that's angel invested in any number of meaningful companies is going to have portfolio companies where you've fully given up on the company and you just try to be nice and, yeah, help if you can.

23:27But at some point, you're just totally disassociated with the investment and you're just like, and that's the reality. But, yeah, so it's very healthy to be like, yep, I'm disassociated, like moving on. I'll put my energy elsewhere. And in both directions, right? I mean, it's just a better relationship. And so I loved the process of going public. I think the thing that we did, like it was an opportunity for us to really kind of retell our story. Like you don't get to reinvent yourself very often, but the process of writing the S1 was incredibly just an opportunity to sit down and really do it.

24:03And we really dug in and told the story in a way that to this day, we still refer to parts of that document to sort of explain what is Cloudflare and how did we work. I think that the most clever thing that we did, and this was advice that I got from Ryan Smith from Qualtrics, was he said, you know, he's like, what are you doing about friends and family? And I'm like, we're not going to do it because you can take 5 % of an IPO and allocate it to friends and family. And I didn't want to get my aunt like over if the stock went down or something, I didn't want to have to explain that over Thanksgiving dinner.

24:40And he said, no, no, no, you're thinking about it wrong. Think about the people who if they, you know, if they owed you a favor, that they could make a meaningful difference in the future of Cloudflare and then offer them to the ability to invest in the IPO. And I was like, is that like, is that legal? He's like, check with your lawyers, check with the bankers. But the answer is yes. And I was like, some people are going to have like conflicts. He said, it doesn't matter. Like, even just the fact you offered it, even if they can't do it, will mean that, you know, that they'll always remember that super fondly.