CEOInterviews.AI
Start App
Get an email when Matthew Prince gives a new interview
One email per new interview. Unsubscribe in one click.
Matthew Prince
Co-Founder, Co-Chairman & CEO, Cloudflare

Cloudflare CEO Matthew Prince and Nous CEO Dillon Rollnick on who owns the internet in the age of AI

📅 Oct 22, 2026 USV 43 MIN 56 VIEWS 146 SEGMENTS · 3 SPEAKERS
The internet as we know it will die. But what emerges in its place is still very much up for grabs. To get some clarity, we asked two of the most important figures in The Rebel Alliance - Cloudflare CEO Matthew Prince and Nous Research (creators of Hermes, the largest open source AI harness) CEO Dillon Rollnick -to talk about who gets to build, control, and benefit from an internet run by AI? The conversation was part of a series we launched at the @usv CEO summit where we pair founders for conversations with each other. No moderator. Just two people building at the frontier, geeking out to...

What Matthew Prince said

Written from the verified transcript and checked against it. Every figure links to the moment it was said.

Cloudflare CEO Matthew Prince said he would not trade places with Sam Altman or Dario Amodei because the frontier labs are spending hundreds of billions, if not trillions, of dollars to build what is effectively a commodity. He argued that calls to pause research are untenable and that rivals, including Chinese companies using Cloudflare for security, will keep advancing. Prince highlighted that agentic traffic has already surpassed human traffic on the internet, reaching 60/40, and predicted it will be a thousand times greater within five years, presenting a serious infrastructure challenge. He emphasized the need for a new internet business model beyond advertising, because agents don't click ads, and the need for more efficient GPU utilization. Prince also discussed Cloudflare's work with Nvidia on Open Shell sandboxing for agent safety, and the importance of open models and data access, noting Google's advantages.

Key takeaways

  1. Agentic traffic surpassed human traffic by May 2026, already at 60/40 and accelerating faster than Cloudflare's end-of-2027 projection.
  2. If every knowledge worker had one agent, the internet would need 40 times current annual CPU production.
  3. GPU utilization across hyperscalers is typically only 5-8%, presenting a major efficiency opportunity for virtualization.
  4. Prince would not join Anthropic's board because Cloudflare must remain neutral among rival AI labs.
  5. Cloudflare is co-designing Open Shell with Nvidia to give agents guardrails beyond simple prompting.

Numbers and commitments

FigureWhat it refers toTypeAt
60/40 Ratio of agentic to human internet traffic metric 1:04
40 times Required increase in annual CPU production for one agent per knowledge worker metric 22:37
5 to 8% Typical GPU utilization at hyperscalers metric 28:04
92% Share of world's data centers located in the US metric 28:04

Chapters

  1. 0:00Agentic traffic surpasses human traffic
  2. 6:25Open models and competition
  3. 11:16AI safety and scale
  4. 13:55Agent security and Open Shell
  5. 16:12Future of AI differentiation
  6. 20:36Dead internet theory debunked
  7. 22:37Internet business model challenge
  8. 28:04Data center capacity and GPU efficiency
  9. 30:57Inference location and policy
  10. 33:23Chinese models and open source

Questions asked in this interview

12
  1. 0:39Is it okay if we talk a little bit about OpenAI?
  2. 0:54What do you think about that?
  3. 5:31Like how did you even start down that path?
  4. 8:22Open AI you did a deal with them right on the security?
  5. 13:12Are you using any of them?
  6. 20:20Which is why distillation is so scary, right?
  7. 26:03Do you have any ideas or answers?
  8. 28:00See the answer mostly on the create more hardware, create more efficient software, or do you think that there's a business model solution in this as well?
  9. 36:12We talked earlier, like what does it take to make a good model?
  10. 36:31And the Chinese companies have no issue taking data from wherever the hell they get it. You know what I mean?
  11. 36:58There's big American companies that could participate in this, but ultimately the burden of being open invites the scrutiny of where did this data come from?
  12. 38:11If you were to co-author a blog from Cloudflare and put it on your website, and then I eventually made its way into my model, did I distill whatever that source model was?
Matthew Prince 0:00 ↗
I would not trade places with Sam or Dario right now for anything. Because I think they have a real challenge that it might be that they're spending hundreds of billions, if not trillions, of dollars and what they're effectively making is a commodity.
Interviewer 0:14 ↗
Marker.
Wait, wait, wait, wait.
Oh, that goes away. All right.
Now you can get started.
Now we get started.
Tell me a little bit about what Neus is.
Dylan Patel 0:24 ↗
We started in 2023. We actually formed out of a Discord. We thought like, can we fight for a corner of this that stays open so that the tinkerers aren't shut out, which made us famous for the first time of like it got embedded into every major model in the world.
Interviewer 0:39 ↗
Is it okay if we talk a little bit about OpenAI?
Matthew Prince 0:43 ↗
Almost all the big labs are our customers. We got this massive concern where the big model companies are coming out saying, you know, maybe we should even pause research because the security concerns are so...
Interviewer 0:54 ↗
What do you think about that?
Matthew Prince 0:55 ↗
I mean, that's it's pretty untenable. I think that the most interesting question is what's the future business model of the internet?
You looked at the data starting at the end of 2025 to project when would automated traffic surpass human traffic. And our best guess that it was going to be the end of 2027. So I was just stunned when my team came to me in May, just a couple months later, and said, "You won't believe it, but agentic traffic is now already crossed human traffic." And now we're at 60/40 in terms of that and it's accelerating. It shows that we'll be at a thousand times more agentic traffic online than we have human traffic in just 5 years. And if I had to take the over/under on that, I'd probably take the under. I think it's going to be faster than that.
Interviewer 1:45 ↗
So, um, Dylan, it's great to meet you. We've not met before. Neus, Air Mass.
I was bad at French. Like, are you just a Francophile?
Dylan Patel 1:54 ↗
The most common question we get is it Neus and Hermes or News and Hermes? We named it News and it is Hermes. The roots are Greek.
Interviewer 2:01 ↗
What is News in...
Dylan Patel 2:03 ↗
The Greek word for intellect, but it's funny we actually do I'm wearing the shirt. We do a party every year the last 3 years called NewsCon, but it turns out NewsCon means we idiot in French.
Interviewer 2:12 ↗
I can't bear it.
So tell me a little bit about what you're building kind of at, you know, 50,000 foot level. What what's sort of the mission of the company?
Dylan Patel 2:20 ↗
We started in 2023. We actually formed out of a Discord.
Interviewer 2:24 ↗
Okay.
Dylan Patel 2:25 ↗
And we're one of the I think maybe...
Interviewer 2:26 ↗
Like Discord the the chat the chat group?
Dylan Patel 2:28 ↗
Yeah, exactly. We were all kind of watching what is sort of an inexorable trend in technology towards centralization and seeing it especially play out in AI. And we had always been tinkering with these open-source libraries and tools and all that, but we thought like can we fight for a corner of this that stays open so that the tinkerers and the people that want to have access to these things as they create it aren't shut out. And when we started the company, we thought like the chat version of what you do with AI, though it's magical, could not be the final form factor. We thought it would be things out there making economically productive work like, you know, being agentic. And actually the first product we tried to create was an agent and it was sort of this like graph-based agent architecture where you're like cobbling together all these different processes to get something from start to finish that's like acting on its own. But the models were so shitty that the product was shitty and it was duct-taped together. And so we kind of like spent the next few years just finding different points in the stack where a team of 10 anons on the internet could like actually have an impact.
Interviewer 3:32 ↗
That's amazing. So So I So I asked for 50,000 feet and you're down a couple inches off the ground a little bit there, but let's let's let's let's zoom let's zoom out and then start like you said we, like who is we?
Dylan Patel 3:43 ↗
It's a bunch of anons on the internet. So...
Interviewer 3:44 ↗
Yeah. I mean have you met at this point?
Dylan Patel 3:46 ↗
Yes, of course.
Interviewer 3:46 ↗
Were there some of them that you met and you were just like actually don't really want to start a business with you?
Dylan Patel 3:50 ↗
No.
Interviewer 3:50 ↗
Okay, that's good. That's a good sign.
Dylan Patel 3:51 ↗
One of our co-founders is pseudonymous. Like we all call him by his screen name. And we call each other by our screen names, but we all know each other. We're all in this thing together. And the 10,000-ft view is we want to bring AI to as many people as possible as fast as possible. And we want to do that in an open way and a incredibly sort of neutral way. And what we're building with Hermes, I think is just the interaction layer between man and machine.
Interviewer 4:11 ↗
So So are there Were there literally 10 of you that were the the initial co-founders? Or how big was the original team?
Dylan Patel 4:16 ↗
Four, but quickly we filled in. Like we had this community, like I said, it was a Discord. Current company's Discord is actually still the same Discord. It wasn't a company. It was just a loose affiliate of people. So it was really easy to draw a talented folks in.
Interviewer 4:29 ↗
Yeah, and how did you early on decide like who worked on what?
Dylan Patel 4:32 ↗
It was a process of learning. Not your typical founding story. Like we obviously wanted to create something commercial so that we could continue to do the work we cared about, but not a very clear division of like, "Hey, this is you. This is..." It evolved sort of naturally.
Interviewer 4:46 ↗
And just And is that Is it Is this company still work that way? Is it Is it kind of consensus-based and you come together as a you know, Discord chat room? Or how How does How do you Yeah.
Dylan Patel 4:55 ↗
No, no, no. We're We've, you know, totally buttoned-up. Looks a lot more like a normal company. We're still primarily on the internet. We do have an office. We We We function and we make money. We do things like a normal company.
Interviewer 5:05 ↗
Yeah. And And what was your background before before this? I mean, other than, you know, hanging out on on Discord.
Dylan Patel 5:11 ↗
I studied computer science and I also studied some humanities. I double-majored in history and computer science.
Interviewer 5:17 ↗
Yeah.
Dylan Patel 5:18 ↗
That seems like a theme with a lot of the folks that are in AI. It's like It's almost like the the sort of double major and that, you know, sort of the revenge of the of the word cells.
Interviewer 5:27 ↗
Well, I was seeing you as well. Similar I mean, went to law school and all that, right?
When you when you guys started out, I mean, building your own model sounds like a, you know, enormous challenge. Like how did you even start down that path?
Dylan Patel 5:40 ↗
Back then it wasn't. We We found post-training. Post-training was the resource relatively resource-light part of the equation where you could take like this amorphous hunk of clay and sculpt it into something useful. So, we kind of pioneered synthetic data at a time when people didn't think that that was necessarily like additive. It was There was a discussion going on like is that necessarily a watered-down version of the original source data? But we thought definitely not. Like you can push these things in different ways and create the data that you want to create. So, we were able to do a lot with with very little and just get performance that kind of rivaled closed labs. I think I heard you say previously like you would love a world where there's thousands of labs doing that, right? I think post-training is the way to do that. Pre-training a little bit harder, but I'm curious if you have any thoughts on what that might look like.
Matthew Prince 6:25 ↗
We've just got to make sure that it's in none of our interests for there to be, you know, just one superintelligence that's out there. We want as many different ones as possible. Like and and I think you want actually really just very specialized different versions of it. I wouldn't hire my friend who would perform the best on Jeopardy, right? That like you don't want just some general intelligence. You actually want sort of specific um different things. And so, the ability to actually shape the weights, shape how the models look, and be very open I mean, it seems inherently like that's got to be a big part of what the what the future's going to look like. For us, it's so much about how do we just make sure people have choice, right? And so, we're we try to be that neutral player in this space. Like that, you know, the folks at Anthropic talked to me at one point about, you know, potentially joining their board. I was like, I can't. Like I can't pick a side in this. Especially with all the rivalries that are in this space. I mean, it's it's really intense. Like the, you know, OpenAI exists because, you know, Elon and Sam got so upset with Larry over a dinner, supposedly. I mean, over what was effectively like a religious dispute. That they went in separate directions. And then obviously Dario broke off from OpenAI. I was in the room in India at Modi's Modi's um conference where they refused to hold hands. And no matter how awkward it seemed like in the pictures and the video it was way more awkward in the room and you're just like guys and you can so the fact that you know you have all these just incredibly intense rivalries and yet all of these companies use Cloudflare and we're almost sort of like the electrical grid for them like we try and stay incredibly neutral and I'm really proud of the fact that even these intensely kind of rivalrous companies at the end of the day say hey we need the best network we need the best security and therefore you know we're going to turn turn to Cloudflare and so it's fascinating to watch how just massively traffic for entrants swings based on you know whatever model is right at a particular given time.
Interviewer 8:22 ↗
Open AI you did a deal with them right on the security?
Matthew Prince 8:25 ↗
You know with Open AI um you know we said you guys have amazing models one of the concerns is how do you make sure those models are being used by the good guys versus you know being used by the bad guys and like they're very very busy with the models and doing all of those things and and then they came to us and said we could be a good clearinghouse to make sure that models were not only being used um for security in the in the right way but we're actually being used by the right people um to do that because I think you know we've got this massive concern seen you know even in in the last little bit where the big model companies are kind of saying you know maybe we should even pause research because the security concerns are are so are so so vast.
Interviewer 9:04 ↗
You think about that?
Matthew Prince 9:05 ↗
I mean that's I like I think that's pretty untenable at some level I mean I think that even if you have a couple of companies that pause research I doubt you're going to pause pause what you're what you're working on and I certainly don't think that...
Interviewer 9:17 ↗
Dog but yeah.
Matthew Prince 9:18 ↗
But but but I I mean don't don't don't sell yourself short I mean all these things start start somewhere and so I think that and then and then rivals you know I mean that what's what's happening in China and again a lot of the Chinese companies are are using us to to be secure as well and and it's amazing to see some of the work that that they're doing I I everyone's going to keep moving this forward so I sort of more of let's keep building let's keep building this stuff up and I it's it's I mean it's it's a little freaking me out. I mean I'd be curious like that so many of my smart friends are like we might be at the edge of like the the you know the sort of Doom... I'm sort of like it feels like we're at the edge of something pretty amazing and yeah it's going to be disruptive and some things are going to going to change but it feels like we can stay stay in front of things. Are there things that you're seeing that you're seeing that sort of spook you?
Dylan Patel 10:04 ↗
It's a tough conversation to have because like there's so much nuance around it but the most like bombastic things are the things that are repeatable and easy to keep going on like the safety thing. Like my sister she's a pediatrician. She text me it was the first time she'd heard about the paperclip analogy like what if what if it synthesizes all humans into making paperclips. She's like holy like do I have to be worried? I was like no you should be worried about you know saving kids right? The thing that...
Interviewer 10:30 ↗
Is a pretty relatively safe job for for quite some time I would I would think.
Dylan Patel 10:35 ↗
Yeah I think so. She's good at it. I know she'll be good. So it's I mean like yes there's so much to be hopeful for and I think the issue with the fear thing is it's very easy to press on and it's very memetic and it's hard to make good decisions out of a place of hysteria. And so like if on the one hand you have like the doom of all of humanity and on the other hand you have like oh I'm getting better search results for my you know restaurant or whatever. It's hard to have that discussion but obviously that's not what we you really have on both hands. You have like potentially solving diseases and curing cancers and like...
Interviewer 11:09 ↗
And and some risk. I mean it would be pretty scary if we all of a sudden give anyone the ability to create the world's kind of most evil virus.
Matthew Prince 11:16 ↗
I mean like that's kind of I I take the same view it's sort of like we we've done this so many times like that you could buy a book that tells you how to make chemical bombs like it's...
Interviewer 11:23 ↗
I have the Anarchist Cookbook.
Matthew Prince 11:24 ↗
Yeah yeah. There you go.
I wish that there was more less like hysteria around it so there was more ability to be nuanced about yeah I mean there are risks to like the security hacks those are real they're not like completely fabricated you know depending on who you ask but...
Interviewer 11:40 ↗
Is there any like who are the people that you're looking to and listening to when you're cuz you're you're at the cutting edge of a lot of this stuff and and thinking about it who do you think has a kind of clear view of it?
Dylan Patel 11:49 ↗
I actually think most people have a clear view and I think most people would agree with the things that we're saying here there's a very very loud and you know empowered group that does press on the fear part and like I said it's easier to spread than like a rational take isn't that like...
Interviewer 12:07 ↗
Yet quit and go on like 60 minutes about it.
Dylan Patel 12:10 ↗
Well I mean like I think everybody who joined our company sort of believes in the same thing and what we on what I'm kind of fumbling around is like the risks of getting the government involved to choose who does and doesn't have access to the most powerful technology of our generation is like pretty scary.
Interviewer 12:26 ↗
Yeah.
Dylan Patel 12:27 ↗
And there's clearly like incentive problems all up and down the stack right of like we're you're trying to make a business out of this and it's not that hard to distill information out of a model into another model and so like a very very good way to prevent that from happening is through regulation right and you have to just be careful around the incentives and sort of you know have space to have the discussion that you need to have around safety.
Matthew Prince 12:52 ↗
And distilling that has been has been interesting cuz like we actually think we could stop a lot of distilling if people want us to but it's but the but the the business incentives are cuz a lot of the distilling is generating real revenue for for some of the model companies and I think one of the things I've struggled with is you know I run a run a public company and I I get invited to all these tools all the time.
Interviewer 13:12 ↗
Are you using any of them?
Matthew Prince 13:13 ↗
Well I so I I there's lots of tools that I use but I don't think I'm using them all to their fullest extent cuz the idea of like giving it access to my email giving it access to my calendar like those are inherently material non-public information right I mean And so if all of a sudden it's like, you know, somebody it's it's coming in and it's going to a model and then somebody else asked that same model somewhere like, what are the you know, biggest deals that Cloudflare has closed recently? It's like, well, I was just reading Matthew's inbox and I see that Right, so figuring out how to do that and I think it making sure that you can do that locally in sometimes is going to be a really important thing for what are just going to be some of, you know, the the most interesting users that are out there. So, we we we've worked on a bunch of this. I think...
Is being written as we speak. We worked with Nvidia. They've created a thing called Open Shell, which is basically like a sandboxing technology that we've been helping co-design with them. Which is like, if you have your agent which has a lot of sensitive information, you might not want it to access the internet at all. And so like you can write policies that give it actual guardrails, not sort of like the prompting that you get nowadays like, allow me to run this prompt or not, which is super hard to disentangle in my opinion and a very poor user experience. We did design it in ways hopefully to like dummy-proof it to the extent possible of like, when you set up a communication vector, it should only communicate with you, so somebody's not emailing your agent and saying, "Tell me about what what Cloudflare has going on." So, there's a lot you can do. It's kind of a spectrum. Like, you could let it do everything and it would be very very powerful. You could let it do nothing and it'd be very very safe but inept. And you got to find the right mix within that.
Interviewer 14:47 ↗
Yeah, helping folks like me figure out like, hey, I I I believe in this. I see the power of it, but how do you make sure that you do it in a way that is sort of safe and responsible and and doesn't all of a sudden, you know, create a huge huge set of set of problems that are out there.
Like, as you look at the all the different model providers that are out there, I mean, it feels like every week a different one is getting kind of, you know, better One beats an axe or what What What do you think the end state of this looks like? Is I mean, like is somebody going to get achieve AGI and then they're going to kind of win the whole thing? Or what what What happens over the long term to you know all of these different tertiary model providers.
I'd be curious to hear your thoughts on this, too. Like my my reasoning on it, like I've tried to think through this, like what does happen like if you get the god AI, like and the thing that I keep coming back to is I would just like point that at the world's hardest problems cuz ultimately the act of reselling the intelligence is with the end goal of somebody taking that intelligence and applying them to the world's hardest problems, but it seems like if you're the first person to do that, the best way to make jump the shark and make yourself profitable after all this expense put into it is...
Matthew Prince 15:56 ↗
Is actually go build the...
Interviewer 15:58 ↗
With partners.
Matthew Prince 15:58 ↗
Yeah, whatever the the best the best new bank is or don't don't sell it to JP Morgan, become JP Morgan.
Interviewer 16:04 ↗
Or partner with them or whatever it may be, but like yeah, like actually apply to those. I'm curious what you experience in this and what you think that that end game might look like.

65 more exchanges in this transcript

Sign in free to read the rest of this interview. No card required.

Sign in to read the full transcript

Cite this transcript

APA, MLA, BibTeX
APA

Prince, M. (2026, October 22). Cloudflare CEO Matthew Prince and Nous CEO Dillon Rollnick on who owns the internet in the age of AI [Interview transcript]. USV. CEOInterviews.AI. https://ceointerviews.ai/interview/3115206/

MLA

Matthew Prince. "Cloudflare CEO Matthew Prince and Nous CEO Dillon Rollnick on who owns the internet in the age of AI." USV, 22 Oct. 2026. Transcript, CEOInterviews.AI, https://ceointerviews.ai/interview/3115206/.

BibTeX
@misc{prince2026_3115206,
  author       = {Matthew Prince},
  title        = {Cloudflare CEO Matthew Prince and Nous CEO Dillon Rollnick on who owns the internet in the age of AI},
  howpublished = {Interview transcript, USV. CEOInterviews.AI},
  year         = {2026},
  month        = {oct},
  url          = {https://ceointerviews.ai/interview/3115206/},
  note         = {Speaker-attributed transcript with timestamps}
}