Cloudflare CEO On The Coming Collapse Of Internet Infrastructure

0:00 In an internet where AI bots now outnumber humans, what happens to the web and can news information publishers and small businesses survive? That's coming up with Cloudflare CEO Matthew Prince. Matthew, it's great to see you again. Welcome back to the show. >> Thanks for having me on. >> All right. let's go with the headline here, which is that recently Cloudflare has found that AI bots outnumber humans browsing the web. I I can't tell whether that's surprising or not.

0:27 I think part of me would say, hey, you know, of of course that's happened and it's surprising that it just happened. there the sort of conventional wisdom is there's always been a lot of bot traffic on the web, but just talk a little bit about the escalation in AI traffic that you're seeing, whether that's, you know, crawling for new information or agents taking action and what it says about the rise in in these type of AI applications recently. Yeah, you know, I I'm I'm certainly surprised you know, I was asked at WebSummit in November of 2025 when when we thought that automated traffic would pass human traffic online. And we pulled all the data and looked at it and we and our best estimate based on trends was that it was going to be in the second half of 2027. at South by Southwest in March of 2026 was asked the same question again. We pulled the same data again and we're like, "Huh, that's interesting.

1:22 It's moved up. is now we think going to be in the first half of 2027. And so I was I was floored when our team came to us and said my team came to me and said in May of of of 2026, you won't believe it, but we now see that automated traffic has passed human traffic for the first time in the history of the internet. And and I think that that that really does just show how fast this is progressing and how fast the actually the derivative is is changing where the the slope of the line is getting is getting steeper. And if we if we just assume the current trends, what's what's blows my mind is that if you just fast forward 5 years from now, we estimate now that automated traffic online will be a thousand times what human traffic is. And that's not because human traffic is going to drop.

2:10 It's going to be because automated traffic is just going to continue to rise like crazy over over that period of time. And that's going to put, you know, enormous strain on the resources of the internet. It's going to put enormous strain on the infrastructure. And and it's a lot of what, you know, the team at Cloudflare is thinking about is how do we make sure that the internet is ready for what is going to be, you know, just a a complete and total onslaught of of of automated traffic, >> right? Right? And I'm sure you've heard of dead internet theory, but for those who haven't, it's this theory that one day the web is basically going to consist of of bots and automated content and there the bots will be manipulating that content for algorithmic curation.

2:50 So what you see in your social feed is effectively a dead internet that looks like it's real. this is obviously a little different than that. >> Yeah, I don't think actually the trends we're seeing are kind of the opposite of that. it's actually interesting for so you know Cloudflare started and launched in 2010 and and around 2015 the web stopped growing in fact it plateaued and then it was even declining slightly over over the last you know 10 10 years. and then right around kind of October of 2025 the the the web started growing again and it started growing at a rate which which most closely resembled kind of the mid 2000s and the content that's being created is not like I would not describe it as as automated sort of slop. There's some of that but there's always been you know there's always been some of that.

3:46 What I think has happened instead is that, you know, people who were not particularly technical, people who didn't have a lot of resources, but they had an interesting idea, they had a story to tell, they had something to share, all of a sudden were able to use, you know, incredible tools like, you know, lovable or riplet or, you know, what the team over at at, B 44, you know, at Wix is doing like and and then have those tools be able to build, you know, pretty sophisticated applications. and and that's like for the for the first time almost in Cloudflare's history, we're seeing an explosive growth in things which are getting created and instead of it making the internet worse, I actually think the internet's getting better right now.

4:31 it's just the interface that we're all going to consume it through is going to be very different. It's it's going to be much more mediated through some sort of an AI interface. but I I I think it's incredible that, you know, you can even if you don't have technical skills now, put together an application, put together a website, and bring that online and and and again, I think that that's, that that that's remarkable and I and I think it's actually making the the web is getting healthier, not less healthy right now.

5:00 And this is sort of picking up a little bit about what we spoke about last year is more and more that content is being crawled or accessed by AI, right? So you've seen this this massive increase in AI traffic to websites in the past year to the point where it's now overtaken humans. So do you have any visibility into whether that is AI bots accessing it to train new models primarily agents using those websites as a human would or you know maybe a a query that at test time. So as the bot is trying to find an answer it goes to a site and it and it ends up you know bringing that back into the chatbot interface without a person actually going to the website itself.

5:51 So what are the proportions if you have any sense of that? >> I mean it's it's all of the above and so I think all of those will drive drive things. There's there's there's definitely an uptick in training as more model providers are are coming online. I mean I don't think that you know like OpenAI is, you know, crawling a little bit more than they did, you know, a year ago. Anthropic is crawling a little bit more than they did a year ago. But but that it's more that we've just got more people building models.

6:14 And I think that's great. Like we should have like the world shouldn't be five, you know, frontier model providers. It should be 500,000. and how do we make it so that that's that's that's available. I think at the same time you've got sort of agents that are operating on their own. That's probably the smallest of the categories right now. I think that that will grow and become a pretty large category, but right now that sort of agent to agent kind of commerce and things is is is is less like that that's that's more of a future thing than a present thing. and then I think that you are getting more of the, you know, AI acting almost as on on behalf of humans going out and accessing things. But I think it's it's almost, you know, it's wrong to say that they're acting exactly like humans do. I mean, the scale is just dramatically different. If I'm shopping for a digital camera unassisted by an agent, I might visit, I mean, if I'm really interested, I might visit five websites, you know, and on every website maybe visit, I don't know, 10 pages or something like that. my agent is going to visit a thousand websites and visit every page on all of them. And so like the load that that then puts on this, especially if there's no if there's no financial disincentive to just crawl everything, right, is is pretty substantial. Now, I think you have got to balance that against you know, what's what's the benefit to the user and the fact that a agents are infinitely patient that they you know that they that they will look through everything. I think that that's that's that's really good for end users, but there's servers behind, you know, every one of those requests.

7:41 There's bandwidth that has to, you know, get paid for. There's content that had to get created. And so, there's got to be some exchange of value. And if the AI companies are just taking without returning some sort of value, then that then that then that just doesn't make sense. And so, that's that's again what I think we've got to we've got to figure out, >> right? So I think that what you know when you talk about this in your view is that we're we're eventually going to go to I think you said it's a thousand times more bot traffic than traffic right in 5 years which is crazy. Okay, first of all, just quick some sense of how like how intense this is. the internet, European internet came this close to completely collapsing in COVID because in COVID over a period of two weeks, internet traffic doubled and especially doubled for for video streaming which was a which was a very high bandwidth intensive thing. I remember being on a phone call with senior leaders at Netflix, Disney, AAI, YouTube, you know, all the major streaming platforms. a and it was basically like this Mexican standoff where where where where nobody wanted to reduce their the quality of their video stream because you know they were like people pay for us because we have you know 4K video and we're not going to reduce it. But what was happening was everyone was their kid like in Europe, everyone was putting their kids in front of the TV, you know, streaming first thing in the morning and that was unfortunately lining up with Asia where everyone was getting, you know, was was turn like after dinner was turning on, you know, something to watch and it was just causing so much traffic that we were we were running out of capacity online. And eventually people had to we had to cut their video streaming resolution down in order to make sure the European internet stayed online. And by the way, if the European had fallen over, then Asia would have been next. The US would have been okay for a little bit longer. But at some level, like there's a finite capacity to all this stuff. So you've got to be able to support it. And that was at 2x. At a,000x, it's going to be completely different. And so again, we're spending a bunch of time trying to figure out how do we make sure that the technologies are ready? So like WordPress most popular platform for hosting content that's out there is run a programming language called PHP not the most efficient programming language and so we're like listen we want to support the WordPress community we want to make sure WordPress is successful but if it costs I was meeting with one of the largest WordPress hosts a little bit ago and they said you know it is it cost us about $3.17 to host a website on WordPress and then we charge like $12.99 to it and that's that's sort of what their business is. I was like, if traffic went up a,000x, what would happen? And they're like, well, then it would cost us over $3,000 to host that website. And again, the average WordPress user isn't going to be willing to spend over $3,000 to host their WordPress website. So, we've got to change some of these fundamental technologies. And again, I think in the best case, it's still the user interface and the ecosystem, everything that people have come to love about WordPress, but with the underlying technologies made much much much more efficient. so that you can actually, you know, scale to what's coming. And that's that that that volume of what's coming is just is just again, it's it's almost inconceivable to to to think about.

11:06 >> All right, we have to spend some time talking about this because this is very interesting. So, first of all, to anticipate a 1,000x increase in web traffic coming from bots in 5 years. I mean, yes, like you know, we we've seen a or you've seen quite an increase recently, but in a thousandx is is a crazy number. So, what is going to make up that thousandx? Like what what makes you believe that that's going to be the case?

11:32 >> I mean, we just look at the trends right now. I mean in in May we saw that that you know the automated traffic crossed and we kind of extrapolate out what that rate of change is and then we assume that again the the the exponent is going to is going to increase on on on that cuz cuz it's you know AI is still way underpenetrated with what's there but we've got to invent new technologies in order to support it.

11:57 Just to give you some some some sense, let's imagine that just every knowledge worker on earth had an agent that was working for them 24/7, right? Helping them do their job. And I and again, like that's that seems reasonable. Like every knowledge worker is going to have some agent that's that's working for them. Very few do today really, but but over time like that that becomes becomes reasonable. just with that and and putting GPUs aside, just looking at CPUs, if every one of those agents ran in a traditional container, like you know, which is which is not even a VM, not even its own server, but just a container, right? you would need more than 40 times the current CPU production in an in a calendar year in order to support that level. And that's with one agent. I would guess that most knowledge workers are going to have many agents that are going to be working on their behalf. So we just don't we using the current technology we don't have the ability to support what's coming. And so again what we're spending a lot of time at Cloudflare is thinking about like okay like we went from servers to VMs to containers in order to get more efficient. But even containers you've got to pull in an entire operating system. You've got to build all the tool chain. All this stuff it's pretty heavy weight. So what's even lighter weight in order to support what's what's coming and a lot of what we've built with our developer platform which we call workers and and leaning heavily on technologies like isolates that allow you to you know have much much much lighter weight services. We think that those are the sorts of things that the internet's going to have to have if it's going to be able to continue to scale to support the the the level of agent traffic that just is inherent to to to what what I think the future holds.

13:45 >> So just to put a fine point on it, the increase is going to be agent driven, right? That 10,000 increase is agent >> or h humans interacting with their agents, right? So, but just going to be automated. And again the way to just think about it is again we've if if you've used chat GPT if you've used claude if you've used Grock like and you play with them and you just ask them a question just count how many websites like you ask a lot of sites >> simple question and you go across everything like now imagine that not in the hands of you know probably about 1% of global knowledge workers like in terms of using it in a regular basis just just scale that up to 100% of knowledge workers and instead of it being something that you use maybe once or twice a day. It's something you're using for every task. And it's really easy to just see how a thousandx is not in five years is is not only inconceivable, but it might actually be a low estimate.