AI时代内容价值的转移：Cloudflare CEO Matthew Prince

Today, we're joined by Matthew Prince, the co-founder and CEO of juggernaut Cloudflare, a company that quietly underpins a massive chunk of the internet. From stopping the largest cyberattacks to moving more compute to the edge, Cloudflare sits at the intersection of infrastructure, security, and the future of the web. We talked to Matthew about the economics of AI for publishers, how quickly chat is replacing search, the evolving architecture of the web, what's broken about how it's funded today, and what he wants to force Google to do. Matthew, thanks so much for being here.

Thanks for having me.

I want to get right into the juicy topics, but first, let's make sure our listeners understand the scale and current role of Cloudflare. Correct me if I'm wrong in any of this: Cloudflare is a $66 billion market cap company today, with about $1.8 billion in trailing revenue, and the biggest CDN by far, with a bunch of different products now, particularly in security. What else should our audience understand about the role Cloudflare plays?

1. Cloudflare Rebuilds The Network

Not to nitpick on one thing, but we've never really thought of ourselves as a CDN. We started out very much as a security company. The whole thesis was: Could you put a firewall in the cloud? We saw that servers were going into the cloud, and we saw that software was going into the cloud.

It seemed inevitable to us that the networking equipment would go to the cloud, and the big objection that everyone had was that you were going to slow things down. So we worked very hard to figure out how we could not slow anything down, and the goal was just to get back to parity. It turned out we were a little too good at our jobs, and everything got a lot faster. So, yes, we've ended up competing in the CDN space.

But really, Cloudflare is what the network should have been—what the internet should have been had we known in the 1960s, 1970s, and 1980s how important it was going to be. How can we make it faster, more reliable, more secure, more efficient, and more private? That fundamentally is what we're working on every day at Cloudflare.

How long has it been since you started the company?

We launched in September 2010, so we'll be coming up on our 15th year in September of this year.

Amazing.

I don't think there's a way to ask this question without somewhat trivializing the journey, but how did you become so dominant?

I don't know. I think we just focused on how to do the right thing for our customers and how to solve the problems that were there. At some level, the story of Cloudflare is that we have been customer zero along the entire journey.

Every step started with a question: Could we take a firewall and put it in the cloud? How would we get the data to populate that firewall? We had to have a free service. Once we had a free service, all of a sudden we had to figure out how to scale enormously across millions of customers in an efficient way.

That meant that we had a whole bunch of weird stuff using us. We got attacked from every direction. We had to build a public policy team in order to deal with those issues. We had to build our own security. Someone almost hacked or stole our domain at some point as a way of hacking into us, so the next thing you know, we built our own registrar.

To some extent, Cloudflare has been about starting with a relatively simple idea, making it as broadly available as possible, and then solving all the problems that become inherent once you've done that.

Now you're in the position that I believe you've described as the internet's traffic cop.

Mm-hmm.

A lot of people feel, with the sucking sound of attention toward AI assistants, that the shape of the internet is changing. Before we go to your point of view on what to do, what is your prediction for what's happening?

2. AI Replaces Search

No matter what, the dominant kind of value-creation model of the last 30 years of the web has been search. Search drove everything. It drove all of what you did online. Entire industries grew up around that.

The 3 ways that you could derive value on the web in the past were that you either sold a thing—whether that was a subscription, a product, or something else—sold ads against some content, or created content for the ego of knowing that other people were reading it. I didn't say the business model, but I said the value-creation model, because the third part is really important. The old adage is, "There are only 2 reasons why people create content: to get rich or to get famous." A lot of people are just doing it to get famous.

That's what drives Wikipedia. That's what drives a lot of content creation on the web. I think that the web is shifting now to a new interface. It's shifting away from search, and it's shifting to AI, and we can see that through the trends in terms of Google usage. We can see that in terms of our own usage, where more and more people are turning to these AI systems where they used to turn to Google.

Even Google itself is morphing into an AI company in its interface before our eyes. As we do that, the natural thing that's going to happen is that we're going to consume derivatives rather than consuming the original content itself.

A study that just came out from Pew Research Center says that if Google puts an AI Overview at the top of search, it's much less likely that people click on the links, and that seems like a "duh." The data that we have also substantiates that and shows that, compared with 10 years ago, it's become 10 times harder for the same piece of content to get a click from Google than it was before. That's because of the answer box, because of AI Overviews, and because the search interface has gone there.

That's the bad news for content creators. In the case of someone like OpenAI, it's 750 times harder than it was with the Google of old. In the case of Anthropic, it's 30,000 times harder than it was with the Google of old.

What I worry about is that if the value-creation model of the web has been all about how to get traffic, the new interface of the web isn't going to send you traffic. If content creators can't get value from selling a thing or a subscription, selling ads, or just the ego that they get from knowing someone is reading their stuff, I worry that people aren't going to create content. That's going to not only starve the web, but it's actually going to starve even the AI companies that are using that content as effectively the fuel for their engines.

How do you think that evolves? If you look forward, the other thing that people are talking about a lot right now is agents and the fact that you're not only getting information through an AI, but it will actually go and take actions on your behalf. So the time you actually spend on the web is going to go down, or at least you're going to be dealing with one interface—this agent—that goes off and does things in the background for you.

Do you think that ultimately the AI companies will start paying for content? Do you think there will be other ways to monetize it? Do you think a completely different model starts to emerge in terms of how the web works?

3. The Web Needs Paid Content

There are going to be different solutions for different pieces of the equation. At some level, agentic commerce is probably going to be the easiest of these to solve. There are going to be certain companies that say, "Wait, listen, we'd love your agent to come and buy a widget from us."

There are going to be others that right now are aggregators of information that agents can actually disintermediate or disaggregate from that content, and they'll be quite threatened by that. But ultimately, I think commerce agents and AI are probably pretty good for—

So separate commerce from content, and content—

Right.

Content is what you're worried about.

Content is a different piece. The problem right now is that the default assumption has been that you get content for free. It's interesting that a lot of the content creators are looking to the law as the solution to this.

Generally, I'm a recovering law professor, so pardon me for going down this weird tangent, but I think it's really interesting. In copyright law, the more that you are a derivative as opposed to a direct copy, the safer you actually are, and the more likely you are to fall under fair use.

We've seen a number of court cases—2 that happened here in California within a week of each other—one of which basically said AI uses of content are fair use, while the other one said they weren't. There's going to be a whole bunch of things around that.

The probably more sensible rule is that the more you're creating derivative content, the less likely it is to be a copyright violation. But the opposite is also true: The more that you're creating derivative content, the more likely it is that someone isn't going to go back to that original source.

I worry that a lot of the content creators are focused very much on what the law says today and on copyright law, which may not come out in their favor because it is actually protecting those derivative uses.

I think what we have to figure out is probably a different business model where content creators get compensated. And I think the good news is, as you talk to the big AI companies—and 80% of the major AI companies are Cloudflare customers—we have good relationships with them, and we talk to them about that all the time. What they all say is, “You’re absolutely right. We should be paying for content.”

The devil is in the details, though, because what they are all desperately worried about is how to make sure that it’s a level playing field. They all believe their technology’s the best. They all believe that on a level playing field, they’re gonna win, but they’re really worried: “Well, if Google still gets content for free but we have to pay for it, that doesn’t seem fair,” or, “If I’m paying for it but somebody else gets it for free, that doesn’t seem fair.” So what we’ve been really working on is how we can create that really level playing field, and we think that if that’s the case, AI companies will actually be quite willing to pay for that content.

What are the approaches you’ve been taking to try and level the playing field here?

In order to have an economy, you have to have a market. In order to have a market, you have to have scarcity. No markets exist without some level of scarcity. The problem right now with content is that there is no scarcity. They’re giving it away for free.

We spent the last year working not only across Cloudflare’s existing customers, but also across the entire publisher ecosystem writ large—not just print publishers, but video, audio, music, film, across the entire spectrum—and saying, “We think that there’s a problem with AI, that it’s starting to actually take value and not give you anything back.” And across the board, for every publisher from the Associated Press to Ziff Davis and everything in between, we’ve seen incredible resonance with that message, where they’re all saying, “You’re absolutely right. Our business is getting astronomically harder over just the last 6 months, and we’re seeing less and less of our existing business model working, so we need to do something about that.”

And so what we did on July 1 was announce what we called Content Independence Day, where you could actually have independence from these AI companies. For free, across all of our customers, whether they paid us or didn’t pay us, we started blocking by default any training being done by any AI company. And it was really important that we focused on that because that meant that we could treat Google the same as everyone else. Now what we’re doing is working with the IETF and other standards organizations to say, “Let’s define how you have to announce what your crawler is doing as it behaves online.” And we’re really encouraged by the early work that’s there. As that happens, we think we’ll be able to set in place really fine-grained permissions for content creators or anyone else—

Mm-hmm.

—to say, “Here, humans can get my content for free, but robots have to pay for it,” and then figure that out. That first step of creating scarcity is what you have to do in order to figure out what the market is, and then after that, I think figuring out the market is gonna be what takes some time. I think we’re still experimenting with different variants.

No, super interesting, because if you look in the search precedent, we had a robots.txt file, and that’s where you’d specify whether a search engine could come and crawl the content. It sounds like you’re really extending some of those concepts through to the AI layer.

Yeah, that’s right, and I think robots.txt was a relatively simplistic and blunt tool, where today it basically says you can either allow something or disallow something.

Mm-hmm.

You can basically do it on a directory on your site or on the entire site, but there’s not that sort of fine-grained control. And so we think robots.txt is like the street signs that are on the road. A lot of people don’t necessarily follow the speed limit, though.

Mm-hmm.

And we actually see plenty of examples—in fact, some really prominent companies—that do some very, very, very shady things—

Mm-hmm.

—where they basically say, “Absolutely, we follow the rules of the road,” but when push comes to shove, if it turns out they’re blocked, then all of a sudden they’re doing a bunch of things that look not dissimilar to what we see Russian hackers or Iranian hackers do in order to try to get around those blocks. At Cloudflare, we’re really good at identifying that and stopping it, and we’re also really good at embarrassing those companies that do that. I have a feeling that some prominent AI companies that are misbehaving are gonna get called out pretty soon.

Should the idea of a marketplace for contributing to training work out, what do you think that does to the landscape of the types of content companies that win or lose? I can’t imagine it’s gonna look like it does today because there’s some notion of incrementality.

This is gonna take us down a little bit of a tangent, but I think a lot of the things that are wrong with the world today are ultimately Google’s fault. They’re not the worst actor, but they—

I’m glad we started with “we’re all friends here”—

We’re all friends here.

—and then it’s all Google’s fault.

Yeah, it’s all Google’s fault.

Yeah.

I think Google has been a net force for good in the world. I think that they actually believe in ecosystems. I think they’re trying to do the right thing, but they taught everyone, if they’re content creators, to worship a sort of deity, which is traffic.

Mm-hmm.

Right? And that was the proxy for value. It was: How do you generate the most traffic? That led to Facebook as the next iteration; it led to TikTok. It led to folks like the Huffington Post, which would literally write a piece of content and then A/B-test headlines, trying to figure out which one generated the largest cortisol response to get the most clicks.

Or if you guys remember Demand Media.

Demand Media. BuzzFeed. There’s a whole bunch of folks that were just trying to figure out: How do we actually stimulate rage and get people stirred up so they’ll click on the thing, so that I can either sell them a subscription or sell ads against a piece of content?

And, again, I think that led to a lot of me-tooism. That led to a lot of people writing the same story with a slightly different bent. I don’t think it led to a lot of us actually figuring out how to advance human knowledge.

And so what I think is interesting is, if you think about the AI companies en masse, they’re a relatively good approximation for the sum of human knowledge. They’re not perfect, but probably the best we’ve ever had, right? They’re where it all comes together. And the reality is that, in aggregate, they’re like a giant block of Swiss cheese, where there’s a whole bunch of cheese there, but there are holes in the cheese as well.

Their algorithms, as they come across a piece of content, prune off that content that’s already part of the meaty part of the cheese. Whereas the parts that are in the holes are actually super valuable to them. And so I actually think that if we could create a market where you’re rewarding content creators not for who stimulates the most cortisol—

Mm-hmm.

—but for who fills in the holes in the cheese, and you actually pay people for that, that is a better outcome, and that’s actually advancing human knowledge. And that’s really amazing if we can do that.

Isn’t that arguably what companies like Mercor, Surge AI, or Scale AI are doing, as they do data labeling and hire human experts to basically fill out content areas for AI companies?

Yeah.

So do you basically view this as a distributed model of that, or sort of a web-based—

I’ve spent the last year talking to a lot of people. One of the more interesting conversations that I had was with Daniel Ek. I flew up to Stockholm and saw Daniel. I think there’s really nobody who has compensated content creators at scale like Daniel has. And it’s amazing.

The day before iTunes launched, the music industry was about an $8 to $9 billion industry. Spotify, on its own today—

Mm-hmm.

—pays out over $10 billion a year to the music industry. And so, done right, these can be very much pie-expanding. There’s plenty of cheese to go around—

Yeah.

—if we do this correctly. And I remember he was telling me a story, which I thought was sort of in the same vein—

Mm-hmm.

—which was that Spotify actually looks at queries that people have run that they don’t have good answers for, where somebody searches for—I don’t know—“I want a disco song about how fun it is to dance with your dog,” right? I don’t know.

And if somebody searches for that and it doesn’t get it, they actually publish that list back to content creators and musicians.

Mm-hmm.

And there are several musicians making literally tens of millions of dollars a year just writing songs for what people care about listening to—

Unmet demand.

—that is unmet demand.

Yeah.

And so I think that it’s not exactly the same as data labeling. I think it’s actually saying that, for the first time in human history, we can very accurately identify where there are holes—

Yeah.

—because of the very nature of the pruning algorithms that these LLM models are trained on, and we could then, if we basically resurface that and say, “Hey, we don’t have enough articles about the wing-toed ferret,”

Right.

—that people will actually go out—

Mm-hmm.

—and do that. And if we can then compensate people for that, that actually is much better than—

Mm-hmm.

—yet another article about what’s happening in Washington, D.C., yet another article about—

Mm-hmm.

—you know, how much San Francisco is in decline or on the rise.

Sure.

I mean, again, that’s not actually adding to human knowledge. That’s just rage bait, effectively.

Yeah. How do you think that plays out? If you look at some of these labeling companies that also hire experts to provide at least some of the expert content that you mentioned, and if you look at models like Med-PaLM 2 from Google, which is a couple of years old now, it outperformed the average human physician in terms of output. So if you rated its output against people, at what point do you think we’ve run out of good content from people? In other words, there is some limit.

I don’t think that’s true. I do think that there’s always going to be people running new experiments and new tests, finding new things and new discoveries. Maybe we can imagine some distant future where it’s all robots doing this in the labs, but that’s a long way off.

Sure.

And so in the meantime, I think we can do that. My Black Mirror kind of version of the future, though, is actually one where we’re not going to get rid of journalists, we’re not going to get rid of scientists, and we’re not going to get rid of researchers. You’re still going to need that work.

What I worry about is if we don’t figure out how to compensate independent content creators broadly, we actually go back to almost a time in the meta sheets where the web had historically been this incredible distributor of value creation and knowledge creation. You could imagine a world in which, all of a sudden, you have 5 big AI companies. You have the conservative one, and you have the liberal one—

Mm-hmm.

—you have the European one, and the Chinese one. And they all actually hire and run their own teams of journalists, researchers, academics—the experts that fill in the holes in their cheese. And again, it’s not too hard to imagine that, in some not-so-distant future, that becomes a thing.

What I hope is that we figure out a way to compensate independent content creators—

Mm-hmm.

—and share that knowledge across all of them, as opposed to creating these silos of knowledge behind each variation of an LLM.

Do you feel like the large labs agree with you on how much can be paid out to creators to fill those holes? Because you look at the scale of ad revenue—I mean, even ignoring things like commerce and whatever from the open web. But in aggregate, what’s been paid to labeling companies—$10 billion or less? We’re really far off if people are starting with a very large free base today.

Sure.

But from that first-principles view, I see it.

Yeah, and so you—

Yeah.

—and so you’ve got— So there is some value there. I think the mistake that a lot of content creators made was that they did deals that don’t scale as the business models of the AI companies scale with them.

Yeah.

So if you do a deal that’s like $20 million and you get all my content, I mean, that’s an incredibly naive deal, right? It might seem like a great deal to the content providers on day 1, but it’s exactly the opposite of what you want to do.

What you really want to do is say, okay, if you imagine that there were a way to look at all of the content that’s available and say, “Here’s how much value this is creating,” that’s going to be some percentage of whatever the subscription fee is for your AI model. Or, if you’re an ad-supported AI in the future, it’s going to be some percentage of that.

And then as the AI companies grow—and that will inherently then mean the ad revenue shrinks—you share in that upside as your downside gets diluted. That, I think, is the right model. There’s still going to be advertising out there. There are still going to be subscriptions. There’s still going to be tentpole content that people just have to consume, even if they’re AIs.

But what you also want to do is allow that content to get into the AI systems, and the content creators should get compensated for that. And again, if there’s scarcity, a market will determine how valuable that actually is.

One of the things that Cloudflare is known for, to your point, is really speeding up web pages and the internet. And as we shift from serving pages to models being run, you’re shifting from a world of caching and serving pages to inference. How do you think about that in the context of Cloudflare or some of the directions that you all are going?

4. Cloudflare Moves Inference To The Edge

Well, I think we leaned in heavily. Nobody remembers this, but back in 2020, we partnered with this graphics chip company to put GPUs at the edge of our network in order to allow people to do inference. And, by the way, it was crickets. We launched this product, and no one responded. There wasn’t a single sales inquiry to use it.

And so we apologized to the partner, which happened to be NVIDIA, and went on our way. Now, 4 years later, the market was ready for it. We basically reissued the same press release, and it’s taken off like gangbusters.

Mm-hmm.

I think we’ve leaned in heavily. We believe that a lot of inference is going to happen on your end device.

Mm-hmm.

But there will always be some model that is too big or too resource-intensive, and in that case, the next best place to run it is going to be inside the network, at the edge, and that’s what we’re delivering.

More importantly, I think that if you look at whether it’s MCP or whatever the next protocol is that connects agents to services and allows these things to connect, inherently, because of how much of the internet we sit in front of, they have to pass through us. And so we’re investing heavily behind those protocols, making sure that they have all of the security, the underlying rails and payments infrastructure, and everything else that you need.

And my hunch is that what we solve in the content space and the rails that we create for the payments there very naturally then become one of the models to do sort of agent-to-agent over MCP, or whatever the final protocol becomes—payment infrastructure to be able to handle that as well.

So Cloudflare fundamentally is a network. And I remember when cryptocurrency and blockchains and everything were getting big, people were like, “Aren’t you worried about this?” And I’m like, “They still need a network.”

Yeah.

As AI gets big, it still needs a network. And so I think we sit in the center of this, and as you especially have more agent-to-agent communication, I think the network actually becomes more and more important.

Is there a bet you’re making on what changes in terms of models or compound systems that drive more model traffic to the architecture you describe, where it’s in-network versus in a large data center today?

Well, I wish I could say we were that strategic. I think we go wherever the market demands that we go. And so today—

Including building a neocloud?

I don't even know what a neo-cloud is, but sure.

It's very important.

But I think we're fundamentally always just trying to say, “How do we respond to whatever either our own team needs as customer zero or what our customers need?” The fact that, again, 80% of the AI companies are using us means they are constantly pushing us: “Can you do this? Can you do that?” I think our team has been uniquely good at executing and innovating and staying at least up with whatever the trends are.

Again, I’m proud of the fact that we have ended up in a lot of these conversations, and that so much of the internet does flow through us. One way or another, I think we end up being in the center of a lot of these transactions.

Does that imply any particular belief around open or closed models as people continue to develop capabilities?

We have closed models that run on us. We have a lot more open models that run on us. We have historically been a company that believes very much in open source, and most of the things that we build internally, as long as we can, we try to open-source all of that technology. I tend to be pro-open models.

We work very closely with the Meta team and Llama and everything that they’re doing. But again, I think there are going to be different flavors of this, and we’re happy to have customers at either end of that spectrum.

I’m actually quite skeptical of the “if we allow open-source models, the world is going to end” arguments. That seems histrionic to me. There are things we should worry about, like synthetic pathogens and other things that can be created, but it seems to me like the place to regulate and control that is in the machine that can actually print the pathogens, not in the AI model that can come up with them. That seems like a pretty flimsy argument for why we shouldn’t have open source.

What needs to happen for your view—sorry, I’m still going back to the shape of the web.

Yeah.

What needs to happen for your view of a marketplace for content to emerge? What are the next—

Yeah.

—signs that this is actually happening?

5. Google Must Play Fair

I think the very tactical next step is that we’ve got to get Google to not be a special snowflake. Because Google has had such a dominant position in search, they almost believe that it is their right to have access to content without having to pay for it.

The conversation that we have with them is, “We get it. The deal that you made with content creators in the past was that they give you their content, and you send them a certain amount of traffic. Over time, you have taken just as much of their content, but you’ve sent them one-tenth of the traffic. If we just plot those trends out going forward, it’s going to become a smaller and smaller part. At some point, the content creators will say, ‘We’re just going to block Google.’”

That was unthinkable 10 years ago. It seemed radical 6 months ago. It is what people are talking about today. Why Google is so important is that, when you talk to all of the other AI companies, Google is the one company that they’re most afraid of. The reason they’re most afraid of Google is that they think Google has privileged access to content in a way that is much more difficult for them to replicate.

What I think we have to be able to do, first of all, is say to Google, “Listen, you can still do search indexing, but if your bot is taking content and then transforming it in some way—making it into the answer box, making it into AI Overviews, turning it into Gemini—that’s a different action. That’s a different deal, and you have to be in the same bucket as everyone else.”

Google is going to resist that. I think the good news is that they really do believe in the ecosystem. I think they are trying to do the right thing. Maybe this isn’t good news for Google, but it’s good news, I think, for the web: They have a ton of regulatory, legal, and legislative pressure coming down on them. One way or another, I think we will flatten that out.

Once that happens, I think that’s when we can actually start to say, “We’re going to shut off access to content unless you pay for it.” In the beginning, most of the deals that are done—the actual money being exchanged—will be between large content producers and large AI companies. That’s happening right now, where Condé Nast, Dotdash Meredith, The New York Times, or Reddit is doing a direct deal with a large model company. That will happen a lot.

Where I think we can play a role is when you have either a large content provider trying to make a deal with all of the AI startups that are out there, which they really do want to do and want to do in a way that scales, but they can’t do one-off deals in those cases, or you have the long tail of content with all of the different AI companies. In both of those cases, I think Cloudflare can play a role in helping set basic rates.

How that model looks, I’m not sure. It might be that we negotiate, basically on behalf of a number of the content providers, with all of the different AI companies—a pool of capital, much like how Spotify does—and then distribute that out. It might be micropayments every time you access a piece of content. It might be that training is actually a different payment rate than search. That’s something that we’ll have to figure out.

But step 1 is that we’ve got to get Google to play by the same rules that every startup and every other company is playing by. The minute we do that, I think the rest of the marketplace will actually happen a lot faster than you think.

For any content company or individual provider—since that used to be a big part of the web—there’s no business model for them today to make money off content going into these AI experiences. They can’t easily predict what is incremental to models. What advice would you have for them?

I think the first thing is that you’ve got to get back to controlling your content. You have to create scarcity from the beginning. How do you make sure that you’re not just giving your content away for free? Again, we’ve made that easy. There are other companies that are working to try and make that easy as well.

One way or another, create scarcity, and then start to have conversations. You can see which AI companies are the most likely to deal with it. Just today, there was news that Google is starting a pilot project to pay news providers, something they swore they would never do. But again, I think they can see that this has to happen. They do believe in the ecosystem.

If the incentives for creating content go away—if you can no longer sell something, if you can no longer sell ads against something, if you can no longer even get the ego hit because, if people aren’t going to the original source, you don’t even know—then we’ve got to figure something out around that. If you write some incredibly influential piece that ends up in millions of AI responses, you don’t actually ever even know that happened. You’re yelling into the void. We’ve got to figure something out around that.

The first step for content creators is to recognize that the business model of the web is changing. Second, recognize that there is something you can do about it. You can actually create this scarcity. And then, third, participate. Start to go out and say, based on the data, “Hey, you keep trying to crawl my stuff. Let’s figure out a way that we can have some fair exchange of value for that.”

One other thing that you mentioned as a side note when we were talking earlier was that you felt a lot of the models would actually be running on-device and locally. Obviously, there would be things on the edge or in the cloud that would be the bigger models, perhaps doing more complex tasks. When do you think that’ll happen? Do you think that’s based on when the device is advanced in certain ways? Is it model size? Is it something else?

6. AI Moves Onto Devices

I think a lot of it’s happening today. On your phone, there’s a lot that your phone is doing locally without having to go out.

Sure.

There are certain places and certain applications where it has to be local. If you have a driverless car and there’s a red ball bouncing through a yard with a little girl running after it—

Sure.

—whether to hit the brakes or not can’t be dependent on network conditions, right?

Mm-hmm.

So that has to run locally. I think the big place where there’s going to be exciting innovation that doesn’t feel like it will be today is in how you take, especially on the inference side, and make it significantly more power-efficient.

That ends up being the biggest limiting factor. Apple has shown that it’s possible, and that you can actually have relatively power-efficient GPUs and TPUs that are out there. When we talk with the folks at NVIDIA, it feels a little bit like talking to Intel back in our case in 2010—

Sure.

Or Apple’s case in 2005, where they were like, “You’re doing it wrong if you care about power efficiency.” I remember sitting in Intel’s research lab outside of Portland in 2011. We were a tiny little startup, but we were doing interesting, innovative things, and we were using their chips. So they brought me in, and I was like, “The only thing we care about is cores per watt.”

Mm-hmm.

“And we just need as many cores per watt as you can possibly deliver.” And they just kept saying, “You’re doing it wrong. You should be water-cooling your systems—

Yeah.

“—doing these things.” And we kept trying to explain, “We don’t have that luxury.”

Mm-hmm.

We have to go into what are oftentimes the oldest, most legacy data centers in the world, where there is a relatively limited power envelope, and we’ve got to fit within that. I think the same thing is going to happen in the AI space, and I’m very hopeful—

Sure.

NVIDIA’s been a terrific partner to us, but it’s been sometimes frustrating to see how more GPU capacity comes along with having to stand up your own—

Yeah.

—mothballed nuclear power facility. That can’t be the solution.

Yeah.

And there’s no physics reason why it needs to be. So I think that we’re actively looking around the ecosystem, trying to figure out who can deliver the most tensor units per watt, or whatever the sort of GPU equivalent is. And that, I think, is going to be the big unlock that allows you to have more running on your device, whether that’s your phone or your driverless car, or frankly, at the edge of the network. Because again, we also have to live within a power envelope, which is not the same as if we were standing up a 100-megawatt data center.

Yeah, I was kind of thinking of it, I guess, from 2 or 3 perspectives. I mean, to your point, there are the actual chips, and in the context of mobile, obviously there was Arm and then Qualcomm as other approaches to basically get to some of the things that you’re mentioning for devices. Separate from that, there’s actual model size and inference time, and a few other things that are overlapping but different.

And so, to some extent, it’s how large of a model that’s high-performing can you actually load on a device, when does that happen, and how well can it run? I was just a little bit curious how you thought about all those different pieces, because there’s also the model component or the other side of that that seems to matter quite a bit. And it’s all coming. It’s all inevitable. I’m just sort of wondering about inference and—

Yeah. I think most of the AI companies are still relatively inefficient in terms of their utilization, from everything that we can see. It is in our interest, based on just how our business model works and everything else, to make inference—to make anything we do—as efficient as possible, because we only charge customers—

Mm-hmm.

—based on the actual work that we do. We’re different than hyperscalers. The hyperscalers, you go out and rent a—

A compute unit.

—a GPU. They don’t care if you use it—

Yeah.

—or you don’t use it. There’s actually no incentive for them. In fact, if tomorrow someone announced that they had made inference 100 times more efficient, that’s great news for us. It’s terrible news for the hyperscalers because our business models are very different.

I think the great lesson of DeepSeek—I just wish that it had been a group of students out of Hungary that had designed it, not out of China—was that there were some really significant, innovative steps that they took to make training and inference significantly more efficient.

Mm-hmm.

We all got distracted in the US by the fact that it was China, and whether it was real or not real or anything else. It was real. There was really great science done there to be more efficient.

Mm-hmm.

I think we’ve just barely scratched the surface on what we can do around model compression, what we can do around much more efficient pruning. There’s a ton in these models that are branches of the tree that, literally, the probabilities of going down them are so incredibly low that you can prune those branches off fairly efficiently and still get incredible performance.

So I think we have an enormous amount of the kind of hardcore computer science—which is different than the hardcore AI science—to do, to just say, “How can we now take these things and make them massively more efficient?” And my hunch is that it is not particularly long before you’re running something that is the equivalent of the current generation of ChatGPT on your iPhone or your Android device.

Does that advancement happen at the model providers, in open source, or at an infrastructure provider?

It’s hard to predict where it is. We’re not investing in how we build a frontier model. That’s not our job. We are investing in how we take any models that we’re running and run them significantly more efficiently. So that’s how we think about it.

I think the question is: Who’s going to be the VMware of AI? And who’s going to create that ability to just slice it? Because we haven’t even done that basic work.

Mm-hmm.

Today, for the most part, if you want to spin up a GPU, you’re taking an entire VM. There’s not even a container because you’ve got to get that low level, or you’re taking an entire machine that is running, and that’s extremely expensive. In most cases, with most of the hyperscalers, to get anything close to attractive pricing, you have to commit—

Mm-hmm.

—to a year of that, and now it’s up to you as a customer to figure out all that efficiency. Someone will come along and figure out a way to say, “Here’s how we can slice these things up. Here’s how we can make them more efficient.”

And we’re just going to speed-run, essentially, the CPU efficiency gains, including a lot of the security things, like Spectre and other speculative attacks that you had in CPUs, which will come to GPUs. All that stuff is coming. We’re going to speed-run the last 30 years of CPU efficiency gains in the next 5 to 10 in GPUs.

I think fundamentally there are a lot of really interesting next-gen models around physics and materials and—

Sure. Yeah.

—different things that may actually be interesting. Obviously, a few really odd, future-looking things on the infrastructure side, I think, will be important. You’ve probably been following a lot of the agentic-related infrastructure that’s actually necessary for multi-step agents, so I think there’ll be a few big companies there. And then there are all the vertical app things.

We will probably compete in that agentic infrastructure space. It will probably not be one provider. It will be something where you’re going to have to have a whole bunch that actually work together in some way, and so figuring out the standards behind that, I think, is going to be important.

That’s super interesting, yeah.

I think that whether it’s MCP—

Mm-hmm.

—or Google did its own flavor of it—

Yeah, yeah, yeah.

—which felt very Microsoft-y—kind of embrace and extend. But all of that space is going to be—

Sort of like Temporal and Graph. All those things are early indicators of new agentic infrastructure that’s coming.

Yeah. That’s right.

Yeah.

I think there are entire, very large domains where a lot of the architectures probably apply.

Yeah.

And the data collection—efficient data collection—is the question, because we’re not going to get robots from Common Crawl.

Yeah.

But we’re probably going to get them. We just have to figure out how to pay for the data.

Yeah.

And so I think figuring out if there are interesting models to get to generalization in these other domains is just something I’m looking at.

Yeah. The other thing that’s going to be really interesting is that I was actually pretty much a skeptic around blockchain and cryptocurrency.

It may be that this shift is the thing, because now we're looking at this and we're thinking, “Okay, let's say we got to a place where it was actually micropayments for every—

Yeah.

...page view.” We do something like 15 trillion requests every day.

It's an obvious use case, yeah.

But how do you then scale these things to be able to work that way? You can't do that with Bitcoin, right? There's—

Yeah.

...and you can't even do that with Solana or other things. I think it's going to be interesting how all of these things that have developed over the last 10 years, how they kind of come back together—

Yeah.

...in interesting ways to invent whatever that next future is going to look like.

Yeah, super interesting.

So.

I think a lot of people also talked about agentic permissioning—

Yeah.

...and identity—

That's right.

...as part of that, too.

Yeah.

How do you actually embed identity on the blockchain so we can do these advanced—

Well, and also, the whole question of identity is going to be really, really interesting because there are times when I might want you, the human, to be there—

Exactly.

...and be okay with that. There might be times when I'm willing to let your agent connected with a human be there, and then there might be times I'm willing to let some sort of agent that is self-directed be there. Setting up the differences in those—

Mm-hmm.

...permissions is going to be really interesting. Browsers are another place. Everybody is building a browser right now. And that's an interesting question for us: How much do we lean into supporting that versus how much do we say, “Okay, that's actually just a way of leaking data back out”?

And so I think getting to some sort of way of saying, “Here's who I am—

Mm-hmm.

...as an agent, a bot, a browser. Here is what I have agreed I will do with whatever it is that I'm taking from you.” And cryptographically signed: “I have—

Yeah.

...I agree to these things.” I think that'll probably be the same fundamental infrastructure that regulates how bots access the web that ends up regulating how browsers access the web. And a browser that takes data and then immediately feeds it back to an LLM might have more restricted access than one that doesn't. And that's going to be an interesting market to tap there. So—

That's super interesting, yeah. And I guess, on the identity side, ZK is a very natural way to actually do a lot of really interesting—

Yep.

...prove you have the right to a credential but not show the credential, do other—

Yeah.

...things that are complicated, and so the blockchain's perfect for that.

It feels like all the building blocks for this have been coming for quite some time, and it wasn't until fairly recently that it felt like, oh, that starts to be the shape of how these blocks come together.

Mm-hmm.

But it feels like both from the need for a change in the business model and from the fact that the technologies have matured to the point that they're starting to be able to handle these volumes and have the broad adoption to actually take off. Again, all the work of the last 10 years—if you'd asked me six months ago, “How is this all going to come together?” In the last few months, it's felt like, oh—

Mm-hmm.

...now you can start to see the shape of what this future might look like.

Mm-hmm.

I think that's what we have time for. Thanks, Matthew.

Thank you so much for joining us.

Thanks for having me.