Vercel CEO: 70% of Our Traffic Is Now AI Agents "Nobody Was Prepared" | Anthropic, OpenClaw, OpenAI

70% of traffic to Vercel Docs is now coming from coding agents and only 30% of pages are coming from humans.

Yeah, it's because it's that same intuition. Vercel's growth was capped by how many humans exist and how many times they deploy by submitting PRs. If now coding agents start taking that job of writing the code and deploying it and testing it and submitting PRs, all of a sudden our infrastructure has like 100 times the demand. Seems like agents are the new computer.

Excellent. It's good to be back in SF from Japan.

Yeah, we 5x'd the guess. That's good.

Yeah, we're only bringing you back if you raise another round. Done.

Yeah, I think agents have taken everybody by surprise. I've been sharing with people that I think even Anthropic has been surprised cuz I remember when one of their researchers was in a podcast, Sholto, talking about how in order to get an agent to be a very good product, you needed to have this chain of events, all of which needed to have a high probability of success happening in a row, and agents sounded like something was far away. But then Claude code happened, which is a coding agent. And I think one of the key features of a coding agent is that it does a lot of problem solving, just like a human engineer does. You'd never write code that's perfect in the first try, unless you're Dave Moran. But typically you write some code, you get a syntax error, and then you have to call this tool, I forgot to install this package, and you do this, and you do that. Turns out that a coding agent was a necessary DNA for full personal agents and general purpose agents to grow from. And that's been the, I think, I'll credit Pete, Open Claw, with the discovery basically of like, well, what if I just toss a coding agent into a container of sorts and I give it access to messaging platforms, and I give it a bunch of tools that are useful like a CLI that can read your email. It's fundamentally a heart of coding agents. And so, I think there's been an emergence that no one expected. You know, I've spoken with a bunch of people about this at every other lab, and they themselves have been surprised that a coding agent is the most promising path to AGI. They were like, oh yeah, it wasn't in our training set to do so much tool calling or getting so good at calling CLIs and shells and things like that. And so, that's been surprising. The other thing that's been surprising to us at Vercel is just, by the way, it's intuitive in a sense, but it's still surprising is that the amount of compute in the world necessary for responding to human intent is kind of predictable. You know, you're awake a certain number of hours in a day, and then you prompt, and we need to do stuff based on that prompt. Agents can just do stuff in the background, and they can quote unquote, we talked about this in Japan with a crazy conference about artificial life, they can reproduce and call each other and fork themselves and claw, as Pete has said, claw into other places. And so, the demands of compute that an agent will put on the world is just so much greater. And maybe a data point from Vercel that I haven't shared broadly that is super interesting is we're actually, an agent told me about an insight in our data that is nearly 70% of traffic to Vercel Docs is now coming from coding agents. 70% and only 30% of pages are coming from humans.

The timeline is that I think last year it was like 90% humans. I can't remember that. But like, yeah, like that's just how insanely fast it's changed.

Yeah, it's because it's that same intuition, right? Vercel's growth was capped by how many humans exist and how many times they deploy by submitting PRs. If now coding agents start taking that job of writing the code and deploying it and testing it and submitting PRs, you know, GitHub also published their data and it's insane. Now all of a sudden our infrastructure has like 100 times the demand. Our sign-ups are growing by 50% month over month. My high school friend who's never written a line of code in their life came to our office yesterday and he's like, by the way, I've been following you obviously and what you've been doing for years, and in Argentina there's always news and whatever, but I never understood what Vercel was, but now Claude introduced me to Vercel. And so, there's this new agentic driven interaction, and I fully expect agents to be signing up for services directly, zero human intervention soon, making payments on behalf of people imminently, you know, getting credit cards and budgets and solving entire problems for people where the service provider becomes a vendor of the agent, not a vendor of the human.

Actually, it's probably closer to like it's all electricity but on steroids in the sense that one of the things that Vercel had embarked on that was very timely for this AI revolution is the switch to consumption-based business model, right? I was actually sharing the story to a group of founders today that I think people underestimate just how complex the problem of building per token or building per CPU cycle is at the scale of providers like us and just how many teams we had to basically devote to the problem of just counting and billing tokens. And now we're trying to basically share that back with the world and provide the infrastructure so that you can build agents that can charge by token and participate in this emergent economy, but the barrier to entry is the fact that agents will basically be very much pay-per-use. The SaaS business model definitely gets disrupted because it was predicated on selling seats.

By the way, one prediction that I'll make is that it's going to continue to get more and more complicated for them to sustain the subscription thing because it's almost an impossible problem. We confronted this at Vercel early on because what customers really want is predictability. And they want to know exactly how much they're paying. But I don't know exactly how much I'm paying AWS and the world that we're going in is so incredibly dynamic. So I would try to explain to e-commerce companies, look, on Black Friday, you're going to use 10 times as much Vercel. So I have to meter you on each unit of consumption of what you're producing and consuming because it's going to be highly fluctuating. And I think this is the problem with selling you $200 Claude Code Max is that they're trying to say like, look, most people hopefully will not use it. But the thing here in the unit of intelligence is so valuable that why wouldn't you use it? Why wouldn't you use whatever high ceiling there is, which may be like $5,000 worth of inference or whatever? You're going to want to use the 5,000. And that's what we realized is that it becomes this insane game of cat and mouse because once people realize the value that is behind the illusion of predictability, they will try to maximize its utilization.

Well, it's actually in my, I haven't looked into it too deeply yet, but my, it seems like a continuation of the same strategy of can we verticalize as much as possible?

And there's two ways that they can subsidize. One is investor money and the other one is growing a formidable enterprise business such that you focus on the subscription for individuals and engineers that are working on their own and whatnot. But there's also another third one, which is they train on the subsidy, right? So if you agree to getting cheaper or more high rate limits or whatever, you become the product, you become the data. Funny enough, I was just joking when I was tweeting before we kicked off. We announced today on Vercel AI Gateway zero data retention guarantees so that you can turn it on and say, Vercel has a negotiated contract with all of the AI labs on zero data retention. We will not allow a single token through these pipes that can put you at risk of being trained on. Because that's going to be the other encroachment is that, yeah, here's a bunch of AI playing the casino money, but I'll learn from everything you do. I'll upload your code base into my training data set, right? And on managed agents, it goes back to like the way that you justify a subsidy is can you lock the customer in to your ecosystem? For example, I don't know if they've adapted yet, but Anthropic for a long time was not playing ball with agents.md. You could only call their instructions Claude.md. To their credit, they actually pioneered a number of open spec solutions. I actually think they've been a really good actor on the open ecosystem side. There's a case to be made that their business incentive is to verticalize. Only do things in the Claude ecosystem. Store your data there. Store your file attachments, artifacts, PDFs. Store your vector DB there. Run your sandbox compute there. And Vercel's answer is the exact opposite. It's like, no, no, no, no, like get the tokens there. Reserve the optionality of using Codex if you want. And here's all the services to build a managed agent that doesn't lock you into a particular vertical model app stack.

Yes. So the agentic internet requires agentic infrastructure. So that's the focus of the company right now. Being the AWS of agentic infrastructure.

I was going to say there is an important but, which is that the existing business was also at the right time and the right place because when you deploy all these apps and when we were in Japan the previous trip, we were talking about the dream of personal software. And then we were joking it only took a year and now we have real personal software. It's kind of nuts. I remember actually when I was in the plane to Okinawa, I sent tweets. And I felt like every good banger, I felt a little bit of like, hmm, I don't know if this is just me being high on my own supply. And to put this in context, that tweet was what? Less than 18 months ago.

Yeah, it was this idea that generating software was in competition now with Googling for software. But now generating software is not just in competition, it's going to win. It means that everyone can produce their own software for personal use or for enterprise use. Because why would you go through the hard work of procuring a new vendor if you've already procured Vercel, you plug it into Codex or Claude Code, and now you generated the software that you thought you were going to procure? Like this is literally happening as we speak. What makes Vercel really good for this hosting solution for generative software? One of the things that we bet on was serverless. Serverless has this property of scale to zero. A lot of software will be so spontaneously created and used that it may just get intense use for 5 minutes, for 2 weeks, maybe a year, maybe it's just for me, maybe it's for a medium-size group, maybe it takes off on X. Like literally every day there is a new generative personal app that takes over the world on X. There's one called the Monitor, which is Vibe Coded Palantir. It literally became one of the top 100 host names on Vercel for a week because everyone in the world was talking about it.

It actually surprisingly is still very high. But some of these probably like blips in time. They rise, they fall. This speaks to the durability of some of the Vibe Coded software. Another example is J Mail. J Mail was a group of guys saying like, oh, like what if we produce a better UI to the Epstein files?

Lo and behold, you have a massive, it's become the Wikipedia of that whole case that everyone is talking about and whatever. And so what do we call our compute infrastructure? It's fluid compute. Fluid implies again, it has the fluidity that CPU cycles might go down to zero or CPU cycles might go so high that you're running something that's bigger than New York Times. Maybe for a week, maybe for a year, who knows. And so we made the right investments to be very well positioned for this era. But now the agentic era also requires innovations. The sandbox is the most prescient one, which is that very fascinating thing about agents is the more power an agent has, the more intelligently it performs. And so if an agent has an entire computer, like a Mac Mini, you feel like you have AGI in your hands.

And this is the problem that people I think run into is if your agent only has access to your rack database and one tool that one engineer wrote, it's like you're wasting this incredible IQ capacity that you just purchased.

Yeah. And so I think what I'm helping enterprises do now is how can we get them to achieve that power but not exfiltrate all their precious data and IP and the keys to their kingdom. And so that's the fine balance that, and again, requires innovations like we invented a new kind of firewall for Vercel sandbox that literally did not exist in the world and things like that.

I'm very bullish on whoever focuses on coding because as I mentioned, I think the foundation of a successful agent for personal use or enterprise use is the ability to code. In fact, one of the things that we learned with the agent that helps us run Vercel internally is that you still need to give feedback to the agent. So Dave and I were talking about soul.md.

No, no, you can't tell everybody about this. We're going to sell these. It's coming. It's merch. We're making soul.md hats. But now you know, I did this feature this, well, dreams, Dave. You have to do dreams.md. I worked on this feature on Open Claw this week called dreams.md, which is pretty fun. But what's another amazing emergence from agents and from Open Claw? And by the way, this is also the sharp edges, by the way. I literally saw that tweet from Gary Tan today that said, I told my agent to improve itself. I literally knocked itself offline and I had to SSH into the machine and reboot it.