Guillermo Rauch: Why Software Development Will Never Be the Same

In many ways, I think MCP will be the new business development, but it's going to be at 100 times the speed. It's not going to be people meeting, and it's going to be agents meeting. We can call it AI transformation. Every company will have to rethink itself, what Jensen Huang calls a token factory. The fundamental thing is that companies will be producing intelligence at scale.

Welcome to the MAD Podcast. Today we have an epic conversation with Guillermo Rauch, the CEO of Vercel. A legendary engineer and one of the most prolific coders of this generation, Guillermo moved to the US from his native Argentina when he was 18 years old and sold his first startup to WordPress a few years later. He then started Vercel, a cloud platform to develop and ship fast web apps, which has grown to over 6 million users and powers websites like Stripe, Adobe, Midjourney, Nike, and OpenAI. On top of that, Vercel has scored a major hit with the launch of v0, their immensely popular text-to-code AI builder. This is a really fun conversation where we covered tons of ground, including the rise of vibe coding.

There's a way of building software where you don't pay attention to the code.

The importance of taste.

You need taste, you need to have a vision of what you want, and you need to be able to articulate it.

Vercel's brand new AI cloud that they just announced yesterday at their annual Vercelship conference.

An AI cloud should have agents that automatically produce solutions rather than give you problems.

The fundamental importance of the MCP protocol.

Think about MCP is like the HTTP for AIs, a protocol that allows AIs to talk to other AIs.

And what advice he would give himself if he were an 18-year-old developer today.

My advice would be focus a lot on getting to know everything that's available in the new world.

Please enjoy this fantastic conversation with Guillermo Rauch. Guillermo, welcome.

Thanks for having me.

So you're recording this on the big Vercel SHIP annual conference, and the big announcement of this week is the AI Cloud, which we're going to discuss at length in a minute. But I was hoping to start with v0, which is a big hit that you have on your hands, and it seems like it was an overnight sensation when you launched it.

Everything that works is an overnight success.

Yes. Yeah, so you launched it in the fall of 2024 and you've been on a tear since. Are there any metrics or any qualitative information you can share to help us get a sense for traction?

Yeah, it's pretty insane. We've passed over 100 million generations of applications. For those that don't know, v0 allows you to convert text to application, idea to application. It speaks to an audience that Vercel has typically not spoken to. It's basically everybody with a job or even everybody without a job, whereas Vercel Vercel required any new engineering skills. This requires that you have an idea. I think it's the embodiment of what people have been calling vibe coding. And to give you an idea of the difference between the two personas, every single second there is 7 app generations happening on v0. v0 has more than doubled the entire user base of Vercel, and Vercel has been around for almost 10 years. Uh, v0 has doubled our number of users in less than a year. And so what's happening really is that coding is being automated. More people can code with the ease that you could type into ChatGPT. And I think that's creating more software. Some people call it personal software or hyper-specialized applications. But the two things that are driving this is, one, I call it everybody can cook. You and I have an idea, we can start bring it to life as individuals, right? Like during the weekend, maybe with our kids. The other thing that's happening is that within companies, teams, and enterprises, people have the same fundamental need. They need to prototype their ideas. They need to communicate better with designers and marketers and engineers. Like, everybody needs to come together around something tangible and real. And that's what v0 is enabling.

And by the way, I read somewhere that Part of the reason why you were able to launch this quickly is that you had a separate SWAT team within the company of, what was it, 10 people? How did that work?

It's 10 people now. It's actually been— it was less. Yeah, I think I actually saw a photo of the original team the other day. It was 4 people, and one was me who was doing QA when we were launching it. So the idea was, because we're building the AI cloud, we need to be able to prove to ourselves and the world that this infrastructure has a good first patient zero application. We wanted to embody what we believe every company will go through. We can call it AI transformation, we can call it AI-fication. Every company will have to rethink itself as what Jensen Huang calls a token factory, or what I could call an AI-native interface, or you could call it an agent. It doesn't matter what you call it. The fundamental thing is that companies will be producing intelligence at scale. And Vercel, for many, many years, we've been in this business of creating what I believe to be really high-quality web experiences. And so we wanted to put that on autopilot. How can we get everybody in the world to be able to produce those high-quality web experiences. So we infused a lot of our knowledge, we connected it to the things that we are experts in. So obviously we created the Next.js framework. Next.js today is the largest web framework in the React ecosystem, 1.4 million monthly active developers. So imagine going from 1.4 million monthly active developers to 100 million to a billion developers. That's what v0 enables. So if you're cooking on Vercel, the thesis is you can do that with an extraordinarily small team. You can start with like 3 people, 4 people, because all the infrastructure stuff is taken care of. So v0 is full stack Vercel, frontend, backend, middleware, and some of the new primitives that we're introducing this week.

And that's a theme that people have been thinking about a lot in the last, whatever, year or two, this idea that you can build very large companies with very few people. So you've you're the living embodiment of it. You've created that startup within the startup.

That's right. And right now, this is probably breaking news of this podcast, but we've actually, as of last week, fully adopted what I call the GM model. There's a general manager for v0. He's like the CEO of v0, and he's a customer of the Vercel platform.

This is not unlike people as a service kind of microarchitecture for people.

This is not unlike the microservice architecture with Amazon and AWS. Paradigm, right? Like, you can think of v0 as amazon.com. We all use it. Anybody can shop, anybody can cook on v0. But as far as the Vercel infrastructure, we actually host a lot of prolific competitors to v0 as well. I believe that there's going to be billions of agents. Literally right before this meeting, I got another pitch for a code review coding agent that had very interesting properties and niches that it was satisfying. So as the market realizes that every company, every task, every role will need these complementary agents, we want to be the platform on top of which they're built. So v0 actually benefits a lot from that kind of infrastructure, but also the agility that this GM model gives us.

You mentioned competitors. I think it would be really helpful for people to hear you compare and contrast the various companies and products that fall under the general AI coding umbrella, whether that's Cursor, Windsurf, or Lovable, or Replit. Does everybody do the same thing or different things in different ways?

Great question. So as of last week, an article came out that I loved by the Zed team, the high-performance text editor with AI-native capabilities, or code editor. They called what their focus will be agentic engineering to contrast it with vibe coding. They kind of, I think maybe provocatively, some people took it as like, ah, vibe coding is old, agent engineering is new. My framework is that there is vibe coding and there is agent engineering in the sense of, look, if you're working on Vercel's mission-critical infrastructure, I literally just announced on X that we hit a trillion function invocations on our new Fluid compute platform. A trillion function invocations. The people that work on that infrastructure platform, you know, sure, they should vibe code maybe to like test things out and whatever, but I want them to— even agentic engineering is like, hey, like, you have to be on, you have to put a lot of attention into everything that the agent generates, etc. This is hardcore deep engineering. Vibe coding is the coding that will be available to everybody in the world. I think both have a place in this world. And our bet with v0 is actually that there is connective tissue between the two worlds. You can start on v0 vibe coding, and with our Git integration, for example, you can begin collaborating with people that are doing more agentic engineering. You can eject to professional coders, you can get help from professional coders, etc. So there is a binding between these two worlds. But v0 is sort of the top of funnel personal software and everybody in the organization participating. Whereas I think agentic engineering, I think for the foreseeable future, until we really reach SSI or safe superintelligence or whatever you want to call it, or AGI, it's going to be a more reduced number of software professionals. And the way that I anticipate it going is, look, mission-critical infrastructure, and really low-level stuff, more on the agentic engineering side. And then if you're creating apps, interfaces, demos, prototypes, specifications, slide decks, Vizier is a very versatile piece of software. You're going to be on the sort of vibe coding world.

And again, in an effort to make this interesting to a broad group of people, could you actually define vibe coding? I think for people that live on X minute by minute, it's obvious, but there's a large group of people people that are just starting to hear about it.

Yeah, vibe coding came up in an X thread of Andrej Karpathy, who was the lead of machine learning at Tesla, worked on self-driving cars, later joined OpenAI and later left. And he started building software and he realized, look, there's a way of building software where you don't pay attention to the code. It's almost like the post-code way of building software because the LLMs and the agents have gotten so good. You give them your product requirements, you give them what you want to build, and they just start building. And they— and you can see what's being built and you can give feedback. So it's not fully autonomous. You're still, you know, driving, but you're not caring so much about what's being generated. You're caring about the product that's being like, what does it look like? What does it feel like? And this was excellent in the context of v0 because v0 has always had that focus on like, the most important thing is does it fulfill the vision of the end user? Like, do you want it to look a certain way? We always started, you know, arguably we started all of our engineering kind of like interface first. We care a lot about how things look and feel at Vercel. So vibe coding is now become— it started out as a way of describing how Karpathy was using Cursor. He was also using a tool called SuperWhisper so that he could talk to the machine. So imagine just being in a sci-fi movie where you sit down like this and you just start talking and then interfaces show up and you're like, all right, ship that. That's kind of what vibecoding is. The really cool thing that I think is new, and this is the space of v0 and Lovable and Replit, is that this is continuously getting more and more democratized. Everybody can now use these applications and it's very exciting. I think, you know, when I think about the TAM, the total addressable market of each tool, when I think about the coding IDEs, you still have to be a coder, meaning you know how to download an IDE, you have to, you know, how to set up, you know, the dev server, the tooling, the SDKs, the libraries. Whereas a product like v0, the TAM is everyone with a computer and they can just open a website as opposed to having to download something. You can just go to v0.dev and you start cooking. So I think it's exciting because again, there's a connectivity issue. So Cursor and Cloud Code and Windsurf, they're really good at what they do. One of the other announcements that we've made recently is we took the v0 model, the sort of underlying code generation engine, and we've made it available to code editors as well. So there's a lot of cross-pollination that's happening in this, in this ecosystem, right? So you can take a tool like Cursor, VS Code, OpenCode, and you can say, please use the v0 model. Now the question is, why would you want to do such a thing? Why would you use that over the stock Anthropic model, etc.? Well, we're continuously infusing our expert knowledge on building Next.js and React applications. We have up-to-date knowledge on the libraries that are really important, like the AI SDK. We infuse design patterns and taste. We have a bunch of best practices about how to build applications, which is super exciting to me because the story of my life is that I've just been a door-to-door evangelist of best practices for web applications.

I still do. Your new Twitter bio right there.

And, you know, Vercel actually started because I'd gone to Brazil.js, a massive conference with thousands of engineers, and they gave a presentation called the 7 Rich Principles of Web Applications. And I did two things there. I presented this essay that I'd written, and I also made a joke about Brazil losing 7-1 with Germany. And I'm from Argentina, which is like Not a super smart thing to do. But then when I came home, I was living in San Francisco at this point, I realized, well, I can't scale myself by just going to every conference in the world, making bad jokes about soccer and teaching people. It's just too much. I said, could I create a framework that embeds these best practices? And that's how Next.js was born. It was literally an essay to web framework pipeline. Now web frameworks get you far, but there's—

and this makes sense. Again, in an effort to like for everyone to understand, Next.js being the open source library that you created on top of React.

So yeah, for context, Next.js powers a lot of your favorite web applications. You don't know it. It's sort of the plumbing of the pixels behind. It's like for a pixel to appear on your screen, there has to be a technology that renders that pixel, uh, for web applications. So Next.js is one of the most popular in the world. For example, we were just talking about Midjourney's new model. Midjourney.com is built with Next.js, but so is walmart.com and so is nike.com. And that's the technology that I mentioned. It has 1.4 million active developers, but it's not everybody in the world. But also, and I've talked to Nike's team about this, when I go to nike.com, me, Rauchy G, I can spot a few things that are like, eh, I would have done this differently. So how can I scale my knowledge and my team's knowledge? My CTO was at Google for 11 years. My CPO was at Meta and created React. He was there for 13 years. Would— could there be a technology that allows us to download our brain and sort of disseminate those best practices and ideas that I would go to conference to talk about? And that's a model. So the v0 model is essentially my brain into tokens and my team's brain and the community's brain, and we're constantly making it better. And I think that is the crux of every company in the world. I believe it will have to go through a similar process. How do you go from, you know, ad hoc or maybe SaaS software, etc., to tokens?

v0 as a result is very front-end centric. Is that fair as of now? And then if so, is part of the roadmap to add more backend to it?

Yeah. So people have asked us for a long time to create entire web applications. When v0 came out, the models couldn't really do it. They couldn't generate enough thinking to sustain data flows and state. State is like a really complicated word that we deal with in development a lot. So we started out with more of just, you could call it simple stateless applications. Over time, we've developed more of an ecosystem of integrations. So it's now possible to get data from other platforms that bring full stack to v0. So Supabase is a good example of, we have an integration that allows you to basically have a Postgres database so you can read and write data. Another really exciting integration, actually, and we haven't really publicized this broadly, but it's the beginning. We built an integration to Salesforce Commerce Cloud. So we showed how you can vibe code e-commerce, and you could imagine vibe coding interfaces for sales processes. In so many ways, v0 feels a lot like the what the spreadsheet and the word processor allowed people to do is like you can, you can turn simple tools into very, you know, flexible and potentially sophisticated business processes. We're hoping to do that same thing through this ecosystem of integrations. The other huge integration obviously is the AI models. People want to use AI to create AI apps or to prototype AI apps. One of our largest customers, it's a Fortune 10. And the reason that they use v0 is because they're trying to think about what does our future with AI look like? We have to very quickly prototype. We have to, you know, see things and experience them at the speed of thought, not at the speed of like multi-quarter planning. Think about a quarter in AI could be 20 different models. 3 new architectures, et cetera, et cetera. And it doesn't seem like it's slowing down anytime soon. So v0 is being useful for the individuals that want to create full-stack apps. It's also being very useful for the enterprises that want to prototype and come up with their next big app.

Let's spend a second on the implication of all of this, and then we'll go into the backend, how it works behind the scenes, the AI cloud and and all of those things. But so vibe coding, curious about your thoughts about the sort of general implications of this, including how does one become a good vibe coder?

Yeah, I believe that the Rick Rubin thing that went viral quite a few times now, which is you need taste, You need to have a vision of what you want, and you need to be able to articulate it. Those are like the fundamental, like, literacy of a person. A successful individual will be able to discern what they want and what they don't want. They'll be able to give feedback. You know, so much of being good at or successful in Silicon Valley is receiving and giving feedback, right? Or even paying attention to the feedback that your customers are giving you. In so many ways, I think you and the AI are like you and a coworker that are going back and forth. Not unlike if I had an idea and I hired someone off of a jobs marketplace to do it, right? So now you have AI, which never gets tired, which is being kept up to date with everything that's happening in the world. And also has companies like Vercel that are infusing new knowledge and expert knowledge in it. So you're basically having this like worker that can make your ideas come true. But obviously you still need to come here and input the idea. There are things that we're doing to help you. So a good example is the v0 community. It's massive now. So we help connect people that have ideas with all of these examples that exist in the world of, you know, might be a dashboard, it might be a crypto app, it might be and cool animation. And so there's a, there's a lot of things that you can one-click clone, which is a very successful sort of like go-to-market motion for a lot of really popular apps like Notion, for example. The cool thing about Notion is I could go to Notion, they were like, all right, it's really cool with an empty document, but that requires like a certain amount of creativity and intention.

Yeah.

Or you can just get something from the community. So that's a way.

Yeah, which is a big problem for AI in general, right? Like we in front of ChatGPT or whatever.

All of these UIs look the same. People make fun of this sometimes. Like they take screenshots of like 4, 5, 8 apps and they all say something similar. Like v0 asks you, what do you want to ship? If you don't know what you want to ship, you can start with something from the community. Another way to start is from a screenshot, so, or even a diagram. Sometimes people just don't know how powerful these things are. You could literally take a photo of your napkin idea, like you can draw it, take a photo of it, upload it, and it'll give you something pretty good. There's a couple other things. We have a little button called, uh, embellish prompt or enhance prompt. So if you wrote a very simple prompt, it'll enhance it. By Enhance It, it's like someone that is giving you ideas, like that friend that's like, oh, that's cool, but why don't you do this other thing? So you have to be aware that when you press Enhance Prompt, somebody else is giving you ideas that are not necessarily yours, right? So I do think it's— it's the— there's still a ton of value for the would-be Vibe coder to get a lot of exposure to the world, to get even just looking at products. Looking at what other people are doing gives you more of that context so that when you prompt next time, you come up with a better thing that you want. And I think the eloquence to express that is also very important.

It's fascinating. We were having on this podcast a conversation with Benedict Evans. I don't know if you follow his work, but we were talking about the open box on ChatGPT versus a GUI and the beauty of GUI being that it already embeds a finite universe of things you can do by clicking on buttons. So it sounds like a lot of what you're talking about is sort of one step towards the GUI for AI.

Yeah, I think the problem with GUI is that, to your point, it's trying to guide the user, but it's trying to show the entire kitchen sink at once. If I think about the GUIs of products in this category, like the— if you even go back to things like Dreamweaver, FrontPage, but then you look at other website builders, etc., you know, I never know which one to press for the thing that I need to do. If you really decoded the neural pathway of how a human thinks, a lot of it is a train of thought that you're following and then it gets parsed into, okay, I parse that train of thought. All right, now I need to use this tool. But it starts with the intention. The intention is please create a form, please move this thing to the left, please do a different style for this thing, please connect it to this API, et cetera, et cetera, et cetera. So you're starting with the intention and then it arrives to the outcome or the tool, whereas GUI is front-loading all of the tools. In fact, you know, something that's fascinating is What an agent fundamentally is, is there's a context, there's a prompt, and there's a set of tools. But it's hard to visualize it because it's all APIs. It's behind the scenes. The agent has been given these tools. So I'll give you a very simple example. When you give a URL to v0, v0 can visit that URL. To obtain information or take a screenshot because you might say, hey, you know, for this page, I really wanted to be inspired on, I don't know, NewYorkTimes.com. So the agent has that tool to be able to use a web browser and get inspiration from it or do research or whatever. So we're entering a world where it used to be that the tools get drawn in front of your face. And knew how to wield them. In this new world, the tools are given to the agent, and the agent wields them and decides when to use which. You could also kind of influence it. You can, of course, tell it, you can guide the agent. But for the most part, it's just really, really powerful to give the agents more and more tools, a little bit of context, or perhaps a lot of context, depends on what stage of the development of the process you are, and then the magic happens.

Two things I want to talk about agents, and I want to talk about taste. For agents, do you consider v0 to be an agent or a copilot in a world where some people in the AI coding world have made the claim that they're full-on agents?

Yeah, it started out as a copilot, like you mentioned, kind of for more like design and frontend. But as you get more ambitious with the things that you wanted to do, like create more interactive code, data fetching code, et cetera, you always end up with an agentic architecture because you realize that LLMs are not bulletproof. LLMs, just like us, right? Like, if you watched a developer code, one thing that you would notice is that almost constantly they're running into errors. I would say, you know, developers get paid really well, but their job is actually kind of— it's like not that pleasant. It's just like constantly being berated. Compile red. It's all red. You never see red in UIs. Why? Because we reserve that for like negative feedback, right? So other than, I don't know, I can think of like CNN.com, like there's very few staples. There's very few major brands that even go with red these days. It's all blue, grays, greens, et cetera, or black, of course. Nice.

We coordinated nicely. Coordinated today.

Yes. And, but a developer is just constantly getting— so if— forget about AI. If you watched a bunch of recordings of what your software developers do at your company, you would have seen, oh, they get this pop-up if they're working with tools like Next.js, or they're getting this pop-up all the time that says error, type check failure, undefined, whatever. Or if they're in a terminal, They're getting all these red letters, or the developer forgets something or doesn't quite remember an API. So they, they Command+Tab, which is a shortcut for going into another window, and they go into Google. This is before AI, right? And they would search. Sometimes they would copy paste the error, put it in there, open a website, tells them something about the error, come back to the text editor. What's happened in the last 2 years essentially is that we've taken that workflow and we've given it to an agent instead. Mm-hmm. So the agent reads the error. Sometimes in v0, we don't even render the error because we know with a high probability that the agent is going to fix it. So that's why it feels magical. You don't even see this stuff is happening behind the scenes. And that's why I mentioned that the magic of agents is that they took away all the graphical user interface because they're taking all those tools and they're using it themselves and they're just coming to you with the, the result. And increasingly that result gets higher and higher and higher quality, more bulletproof. And so obviously it's happening much faster than a developer could because the developer, if you think about that keyboard shortcut, that was the developer trying to do research really fast. I can make my agent, especially with the growing ecosystem of MCP servers, I can make my agent to do that research 10 times faster, 100 times faster, and go through those cycles of iteration really, really quickly.

How do you think about the concept of developer love, which Vercel is one of those rare products that has achieved massive amounts of developer love in a context precisely where the AI craps out on a regular basis and probably always will. So I guess you just mentioned a minute ago that it abstracts away some of the errors, but equally it's a mind shift.

Yeah.

So how do you make people happy?

Yeah, it's really interesting. So the way to think about agents and the things that the models and v0 need are not too different from what humans need. For example, if you're a developer tool builder, something that you would have been known for in the past generation was you put a lot of effort into writing really good error messages. They were being processed by a human. Now it's the same, but they're being processed by an agent. Things— some things are changing. I think the errors that we would give humans needed to have a lot of, like, visual context. AIs just need a stream of bytes. They just need, like, raw raw information straight into their brains, right? Like pure signal, no noise. So that's kind of the big difference. I've been talking to a lot of researchers that have been working on deep research for things like OpenAI, Anthropic, et cetera, because Vercel is very invested in the future of the web. And we were kind of like coming up with what is the best framework for how you would present your content, your documentation, et cetera, for the next, for the future of the internet, which, you know, today if you write a documentation website, you're putting a lot of like navigation and pixels and whatever, and that's great for humans. You still want to have that, but AIs just need raw signal. So there's a new emergence, there's an emergent convention, which is HTML for humans and .txt, literally .txt, like It's like 1950s and we're in BBSs and Unix terminals. LLMs.txt or raw Markdown is better for agents. So if you're building developers still today, you have to have that duality in your head. You need to think, okay, how do I make my content, my errors, my developer tools great for agents? What's actually kind of neat is I've always been a geek of like the Unix days and command line interfaces and terminals.. And those things are actually amazing for agents. So a lot of the things that Vercel have been working on, because kind of my fascination with that have become super relevant. There is almost like a circularity in relevance. Uh, we started Vercel with a really simple command line application. You would type Vercel, enter, and that would deploy.

Mm-hmm.

And then we built a lot of graphical user interfaces because we needed to attract more and more people.. And I was told repeatedly by everybody, investors, advisors, users, everyone's like, are you sure you want to be just in the command line interface? Because that was kind of my bias. But nowadays, when you ask Claude Code, Anthropic's agent, to deploy, they use Vercel because we had, you could argue, accidentally created the perfect tool for an agent to deploy. So long story short, you need to be thinking about agents as your customer and developers as your customer. There is a lot of commonalities, not necessarily though. And so you have to really pay attention. But that's the new world, right? You're not just catering to the developer choosing their own tools, you're catering to an agent that chooses tools.

So developer love and agent love.

Agent love, yeah.

You mentioned taste, which is both an exciting concept but very amorphous. And I heard you talk about how in order to increase or improve taste, you suggested increasing exposure. Yeah, but what does that mean in practice? If you're a developer at Vercel and you want to do great at your job and Guillermo says, well, you need to have better taste, what do I do?

Yeah, maybe just to set the background, I believe that every big internet or technology wave has come with a trend and an anti-trend. So when the internet came out and when we all got email, we got spam. So it's like the tale of two cities, email and spam. Email, awesome, connects the world. Spam, oh my God, I'm getting phished by a Nigerian prince. I believe the next one probably happened with Facebook, right? Like, oh my God, the world is getting connected. But then people started getting concerned about privacy and maybe how they were spending their time. Certainly this is the case with TikTok. So you get incredible, like a creator platform, then you get, you know, and TikTok is actually a good example of like, you get a lot of negative externalities. And I think with AI we have the problem of slop. It's kind of like TikTok in a way. Like TikTok could be great. Like it could be enhancing our IQs. It could be teaching us Gödel, Escher, Bach all day long. But in reality we're just like, we're seeing like twerking or whatever, right?

And so I'm glad you and I have the same feed.

Yeah, exactly.

I thought that was supposed to be personalized.

I tune in carefully. And so with AI, I think it's really important that the AI creators, the companies, the model makers, etc. We set the bar for what people are going to create with it. I think TikTok could have set a different bar. And so what we've done with v0 is we've embedded instructions, we've embedded context. Context is king, really, right? Like the things that you make the model aware that exist or you want are really important. But I think there's still Because the models have, like, I mean, even if you look at their size, right? Like, they have so much information, so many parameters that in some ways you have to discover. People call it actually discovering the latent space. If you pictured it as, you know, Ready Player One, where in this infinite metaverse do you go to? What do you pick from this tree of life? Like, what do you manifest into the world? And so what I call exposure hours is you have to— one, you have to understand how people interact with your own products today. If you already have a product, you have to really pay attention to, okay, how are people exactly leveraging your product? Because I think as creators, sometimes we get confused with an idealization of our products. But also you can increase your exposure hours of other people's products, like literally go and use all the things that are available. And that's going to build up your context in your, in your head. It might inform really interesting crossovers. So if you have two things that you've absorbed about the world— oh, look at how cool the interface for Runway AI is— and then you go and use, I don't know, like Operator by ChatGPT. Okay, like, is there an intersection between these two ideas that could give me my new idea? That's what v0 is really good at., but you first have to have given it that sort of initial prompt or input so that you can create something useful and novel and hopefully high taste. There's something really interesting about Web 2.0. A lot of people believe that AWS started out because it had spare hardware. If you listen to the AWS episode on the Acquired podcast, the story is actually way more interesting. The reason that it's called Amazon Web Services is that there was a hypothesis in the world that when APIs and web services started coming out, people were going to create mashups of applications. Remember even the term mashup? That was like a big principle of Web 2.0. And so Amazon believed that they needed to create web services for Amazon.com inventory and products, etc., and that was going to be Amazon Web Services. In fact, it sounds— that explains perfectly the acronym. Yeah, yeah. It later turned out that people didn't want to create those mashups. People actually wanted to create their own products. And so you have the AWS that you have today, which is known for infrastructure. But I think the idea of a mashup is really, really, really interesting today. Again, why? Because you're going to see soon that the world is reconfiguring itself as MCPs. If the listeners are not familiar, the way to think about MCP is like the HTTP for AIs. It's a protocol that allows AIs to talk to other AIs. It's a way of creating integrations between AI products. And in a more technical way, it's like calling a tool over the internet, calling a tool that a particular agent or AI system didn't previously have. So I think what's going to happen over the next 10 years is that people are going to start creating cross-pollinations between products and integrations between products and novel interfaces to products that weren't— I mean, they were possible before, but just very, very difficult to do. And so what we're going to end up is with an internet that is fully generative, where there is no fixed interface that is very rigid. People have different needs, different levels of expertise. The world is constantly changing and the context is getting richer. And so this new world that we're going into is one where, again, if you're paying attention to like, what are the right MCP servers, what are the right things that we can connect, there's going to be alpha there for new products and new creators.

While we're on the topic, you started releasing product there? Maybe walk us through the Vercel offering and what you have on the roadmap.

Yeah, so in the old— we're going to call it old, but in the old days of the web, you were creating applications and websites that returned HTML or JSON. So you were creating interfaces for humans. I believe that the great next demand will be to create interfaces for agents. And so Vercel is in the business of giving you the infrastructure to minimize the time that it takes you, the cost that it takes you to deploy these applications. So we've basically now made it like 5 lines of code to deploy your first MCP server. So this is useful if a company already has a lot of data and they want to express that data in a way that will be useful for agents. It's actually that you're in so many ways, it's like your minimum viable AI product. That's why it actually has so much traction. If I walk into any company in the world today and tell them, all right, I'll help you create your next big AI product that is going to compete with ChatGPT, that might not happen overnight, but expressing your capabilities and data through MCP is a way of participating in the AI economy. And Vercel has basically reduced that line to handful of lines of code. And we have companies like Zapier using Vercel to host their MCP servers, and companies that are doing significant volume, by the way, which actually kind of surprised me.

Solana, for example, to make it super basic. So if you're target.com and you want to expose yourself to outside agents, potentially OpenAI or whoever, then you create your MCP server, which lives on your end. Correct. And then that allows you to deploy on Vercel. Yeah, on Vercel. And then so that your website is super fast.

This is a really great example. So if you're Target, you don't want to miss out on being embedded in ChatGPT. If people are searching for products, what is that agent-to-agent communication protocol? MCP, Model Context Protocol. So if you're Target, you write these 5 lines of code, You connected your existing APIs and data backends, and now you're interoperable with ChatGPT. But even cooler, you're also interoperable with v0. So if someone at Target wants to create a new application to increase internal productivity, like a better system for backend logistics or quality control systems or support systems, now you start typing in English and the v0 agent, of course, speaks MCP is like speaking a language and it can talk to the target MCP. These MCPs can be internal, external, they can be authenticated. They can, you know, in many ways I think MCP will be the new business development, but it's going to be at 100 times the speed. It's not going to be people meeting and it's going to be agents meeting.

Everybody speaking in text, in .text. That's right. As you were saying. Okay. And presumably that piece of software over time becomes thicker and thicker. Presumably you need security, you need observability, you need control. This agent can come and do this, but not that agent.

That's right. And that's why, as part of our AI Cloud offering, we're working on security primitives that are novel for this new age, right? So we offer the more traditional web application firewall to protect your pages in existing systems. But I believe that you need to have more and different kinds of systems to protect this emergent set of protocols, to observe them correctly, because an MCP doesn't have the same kind of like exactly the same kind of URLs that the existing internet has. But your team will have to understand, okay, what are my most popular tools that are being called? Is there harmful context being passed in? Are people trying to make my MCP server misbehave? There is new kinds of attacks like prompt injection attacks. In fact, there's already been quite a few incidents with MCP servers where depending on what chat message was sent, the MCP tool that it was called was made to misbehave. A simple example that was fascinating to me that I saw a couple of months ago was someone can encode information in an image that the human would not see, but the AI model will understand. So I can actually embed a prompt injection in an image that then calls a tool to exfiltrate data through that MCB server. So it's a brave new world and new security infrastructure and tools are required. And that's why our AI security research team is very focused on, all right, what are the things that we can repurpose? Because I think the traditional firewalls are still helpful, but what are the things that are net new that need to be supplied to these developers? And of course, the frameworks for these developers.

What's your experience so far in the you know, we're all talking to customers about their willingness to embrace that world of MCPs and agents. Because certainly if you spend a lot of time building a very nice responsive reactive interface, the idea that all of a sudden you can just be a feeder for an automated agent may or may not be exciting or maybe scary because you're losing control. What is the spectrum of reactions you've seen?

Yeah, the enterprise appetite for both v0 and our AI cloud infrastructure has been, to me, extremely surprising. I think it comes from a place of personal recognition of the decision makers of, look, my entire life I'm spending in tools like ChatGPT. A lot of CTOs and CIOs are starting to vibecode and like, teaching their kids, like, this is how you create an application. And so when you realize that in your personal life you're using so much AI, but you go back to work and you're like, oh, so everything happens in 4 meetings and 3 quarters and everything is slow, et cetera. I do think you start realizing, like, hey, if I don't jump on this AI train, I might just be entirely left behind.

And that applies to MCP as well. It's like I own Target.com, again, to take that example. It's in my best interest to open it up to agents.

Yeah, to expose your inventory in ChatGPT, for example. Right. And so that MCP server would allow that to happen, whereas if you didn't have it, you're excluding yourself from that economy. Great, great. In new economies. But the other thing that's really interesting is it's you creating your own AI interfaces. The best example is that if I go to target.com, the best way to explore their inventory will not be what we came up with 10 years ago, right? Like, we now have all these agents, tools, integrations, specialized models. So the way that I can— I think the web will evolve to have this hyper-personalized experience. It's probably going to start with search. So when you arrive to these websites, you kind of have an idea of what you need. One of the key inventions there from a UI perspective that actually required machine learning and dynamic web rendering, etc., was when Amazon started recommending—when you buy this, people also buy this, this filtering, right? And so I think there's going to be new primitives like that in the world of agentic commerce, where the agent jumps alongside you in your navigation journey, kind of like having a personal shopper, a personal assistant. Like today, we've not been able to scale that for stores because imagine just, you call them on demand when you need them. But I think it's going to be possible now to have for returning customers, signed-in customers, etc. They can have a super personalized experience. And typically when I talk to the decision makers of these companies, they want to cater to their returning customers. They want to, they want to cater to the people that are loyal to these brands. And I think that's most likely where AI will play an outsized role. Memory being one of the key components. AIs get better the more context they have.. But the other really cool thing is that AI can take down notes of what you like, what you've spent more time in. Even by just giving them the log of events of what you've done on Target.com, we can extract probably like 5 very important facts that will improve your next navigation. This person likes black. This person is a male. This person is so-and-so, this person lives in San Francisco. And those 5 facts and carrying them along in your subsequent journeys can create, I think, subtly mind-blowing and customer loyalty earning experiences.

Do you think we're far from a world where all of this communicates, including internally, meaning that in interaction with customer service, could lead to actually changing the code for a certain product based on a certain amount of feedback?

Absolutely. I think everything needs to go into the context. A good example for us is on Vercel, our customer support team is actually one of the most important teams at the company because we deployed AI, we took the underlying v0 model, and when you go to vercel.com/help, we give you this cutting-edge help of everything you could possibly imagine that relates to Vercel, from, I have a problem with my code, to I don't know how to solve this problem, to can you explain this usage? All the things that are typically hard about the cloud, we have that AI that can just automate and explain. But there's always a class of problems that an AI cannot solve. I can call them the frontier problems. And so it's made the job of our customer service team even more interesting and challenging. Now, there are customer support engineers. When something arrives to them, it's gone through the filter of our cutting-edge models. That means that, you know, these are the people that are going to inform the next iteration of the product. Our customer support team uses v0 very extensively because they, they want to produce suggestions of, hey, if we make this part of the product better, if we improve this area of this, of the service, we will avoid this category of issue that is arriving to me. So it's like you're kind of endeavoring to like make the model smarter, make the product better so that you don't have that ticket. But doesn't mean that there's not another frontier ticket coming. And so their job is still extremely important. I think we've just removed, you know, our numbers are kind of staggering in terms of how many redundant tickets we're answering every single day. And our customers are actually happier. I was a little skeptical a couple of years ago when we rolled out customer service AI because our product is so technical, we're so well regarded in the community that I didn't want, I didn't want to run the risk of like someone wants to talk to a human and they get this AI thing and they might get upset.. But people tell us, this is amazing. Thank you very much. Thank you. Thank you, bot. I love you. So we pay a lot of attention to— I use the term disengagement, as in self-driving. What is the disengagement event from the product? And then we pay a lot of attention to where are we leaking alpha? Where do we need to go and improve our models and our product?

Let's talk about the AI Cloud, the big announcement of this week. What is it? What does that do?

Yeah, I kind of talked about the origin of Amazon Web Services. It was created in many ways to create websites. And I see this really big transition happening from what I'd call pixels to tokens. Pixels are going to be very important. The human-centric interactions, you want to have fast pixels, beautiful pixels. So Vercel will continue to help you with that. But we're entering a world where so much is going to happen agentically, where so much is going to happen because you're serving tokens rather than, you know, the more traditional like SaaS pixels or GUIs. And so there's an opportunity to create a cloud that is purpose-built for enabling the creation of AI products. But also using AI to solve some of the most painful traditional problems of the cloud. One of them being, for example, that if you use the cloud today, if you use AWS, Google Cloud, etc., you're showered with problems. If you use Datadog, for example, if you use Splunk, what do you get? Like going back to like how developers who are confronted with negative feedback. DevOps people and cloud operators are constantly being, you know, given like, the system is down, the API endpoint is creating 500 internal server errors. We believe that an AI cloud should have agents that automatically produce solutions rather than give you problems.. So today you're kind of seeing a glimpse of this because agents are creating pull requests. So if you're not familiar with a pull request, you know, when, when a developer needs to make a change to an existing large codebase, they can't just like YOLO the change into production. They create a pull request and other people review it. Because Vercel has this end-to-end intelligence, we see every transaction, every data point, every request. From our CDN through our firewall, through the rendering process, and through the backend systems, we can create AI agents that completely automate both the infrastructure side of things, but also the problem solving that happens in the day to day of operating really, really large websites and applications. Like, we host openai.com. It's one of the top 20 websites on the internet.. If there's a problem in their infrastructure, I don't want to tell them, hey, there's a problem, tough luck, buddy. I want to open a pull request and say, I've cooked on what I believe to be 95%, or in many cases, actually 100% of the solution. So it's a cloud that's completely autonomous. And of course, what are you going to be building with the AI cloud? Most likely you're going to be building AI products. You're going to be building MCP servers, you're going to be building agents, you're going to be creating conversational AIs. But very, very importantly, I believe that you're going to be creating a lot of AI platforms. I call it, you know, you're going to be creating v0 style agents, you're going to be, you're going to be creating coding agents, support agents, law agents, fintech. I believe that every single industry and vertical in the world is up for disruption. I think your best bet is to start with an AI cloud rather than like figuring out all of that infrastructure yourself.

Okay, so this is pretty spectacular as not just an announcement but as progress. So to unpack it, you're basically saying that the Vercel AI Cloud is making DevOps redundant. I mean, is that a self-healing system?

So there's two parts I could call infrastructure and code. Infrastructure, we've been automating for the past 10 years. And of course, it's all automate, it's on autopilot, it's autonomous, etc. What's been harder for people is the operations of that. Right? For example, someone attacks your website. I'm talking about a particular targeted attack. We host a really, really popular sneaker website on Vercel., it constantly has bots trying to do credit card stuffing attacks, like trying out stolen credit cards. They're trying to scrape their inventory and they're anomalous in nature. Let's say that, you know, 9 to 5 business hours, you're operating your, your online store. You kind of know for the US what your hot hours are, what your peak traffic looks like. It's really easy, and Vercel already does this, to detect an anomaly. But what we're doing now is we're deploying the Vercel agent to investigate that anomaly. We're going to tell you what the recommended action is to do. Maybe the attack originates from a country that you don't even do business in. It's like, why did traffic from Argentina— just to throw myself under the bus, I'm from Argentina, because I always say, like, what country— like, I don't want to say a country that's going to get people upset. All right. Your traffic from Argentina on that.com increased by 100x. I don't want to give this problem to the website, to the developer, to the website builder, etc. And so Vercel agent is going to tell you, you know, you've never received traffic from Argentina. This is 100x. We are going to actually— this is really cool. Obviously we showed this demo on stage because we understand the network so well. We can decode who the user agents are. Are they trying to fake their user? Are they trying to do spoofing? And so we're going to give you all of the facts first of all, and then we're going to give you the decision. And so you're still in the loop, you make the decision. But imagine the other world, a world of a more traditional cloud where maybe there's some anomaly stuff, but typically an operator is going to get paged with zero context. Literally, the pager goes out. I don't know if you heard this story from Patrick Collison, the CEO of Stripe. In the early days of Stripe, he was holding the pager and it was a duck sound because he was trying to make it as stress-free as possible. And he had this really funny tweet of like, anytime he's walking around in a lake in a new city or whatever, he hears ducks, his cortisol levels go up automatically because it's the pager sound.— but this is the real world of the cloud today. You get paged, you get your ducks, and then you're like, okay, what do I do? All right, begin the investigation process. Maybe an hour later, if you have someone that's really on the ball, they kind of have an idea of what's going on. So we believe that the AI cloud will be self-healing as well as completely autonomous from an infrastructure point of view. And we believe that agents and humans will collaborate.. So it's not going to be just like off the rails autonomous, it's going to be with you in the loop. And that's why the pull request is such an important tool here where like we can create code changes. You know, by and large, the cloud has, I would say, like 3— like the dark side of the cloud would be 3 things. Number 1 is what I talked about. The internet is a very nasty place. Attacks constantly happen. And that's why we've been investing so much in our firewall capabilities, DDoS mitigation capabilities, etc. The other one is stuff goes down. So we— you saw the Google Cloud outage the other day. For hours, a lot of the cloud was down. So I'll give you an example. Vercel didn't go down, and this is not me throwing shade at Google, but if you were using a Google Cloud service in Vercel, your endpoint is now down because you would— you took a dependency on Google Cloud. Luckily for us, a lot of our customers did not. We actually looked at the numbers, but there were, there was quite a few customers that had increased 500 errors. This is an error that happens when your website is crashing, but many times your website crashes because of somebody else's stuff crashing. Vercel Agent will also help you investigate and repair that. Attributing the cause is so important and liberating and empowering.. Because if I get an investigation from Vercel Agent that says the reason that you have an elevated error rate is because this outgoing host is down, knowledge is power. I can now— I know, I now know what phone to pick up and who to yell at, or I could help you fail over that system. So that's, that's the second category. It's like when things break. Will help you. The other one is optimization. So frameworks like Next.js help you make the web a lot faster, but you can— you still have— like I always tell people, like, this is Turing-complete systems. We don't yet have the quantum computers that would allow us to simulate every possible future. Things can still get slow in ways that it was hard for the developer or agent to anticipate.. And so we can help you. We can emit pull requests that optimize your systems. We can say, look, big e-commerce website, this image that you're rendering from a mobile phone is just awfully slow. And this is the stuff that I was telling you, like I would do door to door. I literally open a website, I can immediately spot, yeah, that image you're not prioritizing and optimizing correctly. So the Vercel agent will not just tell you, oh, you're slow loser, you have a 50 score. It's actually going to be a part of the solution. So that's kind of like the— on the autonomy side. The other thing that an AI cloud needs is services and SDKs purpose-built for AI. So we announced the AI Gateway in beta. We announced the Vercel Sandbox. Obviously, we announced the Vercel Agent., and we announced improvements to Fluid, which is our compute platform that make it better and faster and way more cost efficient to host things like MCP servers and AI workloads.

And Fluid itself is a reasonably recent launch as well, right? That was a few months ago.

Yeah, it's very recent. One of the things that this Cloud 1.0 to Cloud 2.0 transition has done is that the profile of workloads that run in the cloud has changed dramatically. I cannot emphasize enough just how different these programs are. By and large, the big difference is you can call it a transition from frontend to backend. In the past, everything was pixels. They need to render really fast and you could have short bursts of compute. Amazon invented the 100-millisecond rule, is a very important rule for the world to know. For each 100 milliseconds that Amazon takes longer to load, they would lose 1% in conversions. And so we optimize all of our systems around that world of pixels, the world of tokens. I don't know if you've used the new model by OpenAI, O3 Pro. Yes, it can think for 15 minutes. What compute platform was optimized for a request comes in, something gets launched and it cooks for 15 minutes? Basically none except for Fluid. So what we announced is Fluid now only— this is actually only a billing model change. We already did the sort of like compute optimizations for like running longer, etc. But we changed the billing model. If you're waiting for 15 minutes for OpenAI to respond and you do a tiny bit of compute, maybe to like transform it into HTML, HTML, or transform it into an MCP response or, you know, call other systems. We're only going to charge you for the actual CPU cycles that you use. This is amazing because the cloud started out— maybe we can call it Cloud 1.0— started out with pay for what you use, but you're actually paying for the idle time. It's kind of a a white lie. You're paying for what you allocated, and you most likely overallocated the heck out of your system. With Fluid Compute, you're actually paying exactly for what you use in the sense of what you compute. So it's a really, really, really high efficiency CPU that complements your GPU. So you're still gonna call models like Anthropic's and OpenAI's and Mistral's, et cetera., but you need to do the computation on top of them. And so that's where Fluid really shines. Fascinating.

So it's a reinvention of serverless for the world of AI infrastructure. Correct. In particular around response time, idle time. Okay, what about Sandbox?

Yeah, so one of the fascinating things in the world of agents is that Fluid needs to run the code that humans wrote to power your applications. But what's happening is that increasingly more and more code is being created by agents on demand. A good example is when you're working on things like deep research. Let's say you're a big pharmaceutical company or a bank and you want to create your own deep research system. So first of all, you're going to need this sort of high efficiency compute like Fluid. You're going to need obviously some of the foundational infrastructure we talked about, like CDNs, firewalls, et cetera. But now you need a new thing that actually quite doesn't exist in the world, which is you need a sandbox, meaning a secure place for compute to run that is generated by the model. So when agents are doing research, they might decide, hey, I need to run a quick Python script to compute some numbers or to produce a visualization of the data. Or to make up my mind or whatever. So you can think of Sandbox as the Amazon EC2 of AI. It's not for the code that your developers wrote, it's for the code that gets emitted by the LLMs. And it allows you to create this incredible new products. I mean, you could create your own Bee Zero and Lovable with this Sandbox primitive, but also you can create systems where the code runs behind the scenes in the service of an outcome like a deep research report.

In a lot of what you've described, both for the AI Cloud and v0, it seems that you have a constant data flywheel going where you constantly learn and optimize. Is that how you think about it? And to which extent have you formalized it into a product or a part of the product?

Yeah, I think part of the AI Cloud is going to be what we built. Like, we build Fluid, we build the sandbox, we build the AI Gateway, but also the rich ecosystem of integrations. So for evals, we use a product called Braintrust, for example. And we want to make sure that every product that we don't build but you still need in the AI world is in our marketplace. Another really cool product that I think most people are building agents are going to need is you need to give your agent the ability to run code securely, like sandbox. You also need to give it a web browser, which is also another kind of sandbox in a way. You need to be able If you're an agent, you need to be able to browse the web. And so you can expect integrations like BrowserBase and BrowserUse. They're building browser infrastructure for agents. So the things that we don't build, you're going to have one click away as part of the Vercel Marketplace. Evalz, I can't emphasize sort of like that data flywheel. I just cannot emphasize enough how important that is. I think in the old world, there's companies that are more data-driven, there's companies that are less data-driven, right? Like Apple famously is, is always been a vibes company where their designers sit down, they look at the product, how it feels, the very demo culture. We do a lot of that at Vercel, but famously other companies like Facebook, for example, very data-driven. Oh, the user stopped at an ad for less than 10 milliseconds. Like, we should create another ad. We need to optimize this. And like, if you add 7 friends, you're going to be a long-term engaged user and whatever. Like those, that kind of thinking in the AI world is not optional to have that data-driven approach where every data point you need to synthesize, you need to pay attention to. How users are responding to generations, what feedback they're given, what's the error rate. Satya was just talking about the acceptance rate of when you propose something. If you're building an agent that proposes changes, you want to monitor very closely the acceptance rate of the things that the agent proposes. So this world is— I actually think it's nature is healing in a way. You cannot participate in the AI product economy without getting your metrics in a good place, which I believe could have been a blind spot for would-be entrepreneurs in the past. Luckily, the AI was sort of batteries included.

Let's switch to the business side of things, if you will. How do you adapt as a company to a changing ICP? Presumably because you you create by virtue of v0 a new type of users for Vercel who are going to be non-technical people who are going to need different kinds of interaction with the product. Then presumably over time, the difference between a frontend engineer and a backend engineer and perhaps a product people, all of that is going to blur. So from a business product and business standpoint, how do you adapt to that?

Yeah, one of the basic things was we talked to so many people using v0, and I use the famous Jeff Bezos thing. We doubled down where there was the light. So I think v0 has always been appealing to so many people. Like, everybody would use the product. It's just so easy to use, and like, everybody would have at least an okay experience. But we doubled down on the people that were telling us this is life-changing. And so, for example, within companies, the people that were dev adjacent, the people that were sitting outside of the developer room but still every single day work with developers, for those people, this product is a superpower. And so that helped us refine the persona because we're like, okay, everybody can use this. I'm never going to turn away a sign-up. To the product. But for people that are not professional developers, this is transformative. Even just before this podcast, I got another message from a father that was like, maybe probably developer, and saying like, I just cooked with my son an application. I could have never done this before. So double down where there is the light is one of my operating principles. The other one, again, is realizing how the world of AI is evolving and what are the models capable of. And there's another emergent principle of you have to design from the model forward. If you go too far into a world of science fiction, you actually might not have a working product. So the very first UI of v0 was almost too ambitious. So you would type in a prompt and then it would give you 3 generations of the application. And it was very, very hard to actually get into an agentic conversation where you were going to improve that product. We're almost aspiring in a world of in one shot, it's going to be perfect. And then we realized, well, the models are not there yet.. And so we actually have to redesign the entire product. We called it internally V1, which is kind of a joke. But when V1 came out, our traction improved by 100x because we designed for where the model was at. You still have to create enough leeway that you can— that you can— you still need to leave that door open into the science fiction future when you're working with LLM systems. But so much of what we've done is that again, recognition of who is the persona for which this is a 10x better experience, and what is the reality of the models, and how do you create a product that meets the models, not just where they are, but where they are in the next 6 months, right?

And then design for that. Sorry in advance for a couple of VC questions, but retention and gross margins. So To start with retention, it's been one of the open questions. In particular, when Cursor was raising, a lot of people were questioning whether retention was high or not. By all accounts, apparently it's high. But I'm curious what you think and what you see. And then on the gross margin side, same thing heard in the context of Cursor, that some of those products are have negative gross margins. So on a unit basis, people lose money as those things are getting used. Same thing. What do you see?

What have you— Yeah, v0 has positive gross margins, is a healthy business. The margins are improving, the business is improving. And I think it's rooted in the fact that we designed for an enduring persona because v0 is helpful to both companies and individuals via coding. We're not just left with the very volatile world of, you know, you saw it on X, you saw it on YouTube, you got really excited and rightfully so, right? Like you want to try new tools, you want to try everything that's out there. But, you know, your usage of the tool is more sporadic, is when the inspiration comes. Whereas, you know, to put it in a funny way, like at work, you're forced to be creative, right? Like your paycheck depends on your ability to, you know, produce value and continue to improve things and talk to customers and talk to your colleagues, et cetera. So I think something has improved. I think maybe perhaps I don't have the numbers of other products. You do this for a living. But I think something that's making vZero better in the world of retention and margins is that it's actually delivering value for businesses. There is like an economic equation attached to it. The thing that I'm excited about for individuals is that we want to give them ways to make money, right? I think very much like the App Store made it so easy. I've talked shit about Apple's App Store for a while because there's a tax, there's a review process. It's not like the web. The web is permissionless, et cetera. So that out of the way, what Apple has done really well, is that you implement in-app payments, you implement subscriptions, they have Apple Pay. Like, I believe that they're a money-making machine for their developer ecosystem. And so I think v0, you'll see same kinds of things come to v0 so that I think what's going to improve retention is that you're not just like vibe coding into the void, you're actually creating a successful economy. And because I sold my previous company to WordPress, I saw how Matt did this really well for WordPress. And I think I have a lot of the right inspirations and data points to hopefully have a good shot at this.

Still on the business side of things, it seems that Vercel just had a super big year in terms of revenue. So I read somewhere that you were at $100 million-ish last year. So to put this in perspective, that was year 1 to 9, 0 to 100. And I also read somewhere that you were probably at like 180 at the beginning of this year. So like 80% growth. Why was that? Was that v0? Was it now you're evolving towards the enterprise? Was it now you have fluid compute, so you're taking more dollars because you're also involved in the backend? All of the above.

A lot of it is AI. I mean, just to put it simply, right, like on both sides, it's v0 is growing really fast and is a significant percentage of our revenue. And then you got the fact that like we are quick to meet people where they are because in order to best serve enterprises, I spend all of my time with the startups and innovators. And the startups and innovators, you know, when ChatGPT came out, it became a joke. I never made fun of it. I actually have— I'm on record on this. People were joking about the ChatGPT rappers. I was like, ChatGPT rappers are fucking amazing. Everything's always been a rapper of a rapper of a rapper. And we leaned in. So we created the AI SDK. The SDK, if you look at the download numbers, I wouldn't be surprised if it surpasses Next.js, right? Like in the trajectory, it's the second biggest module in the TypeScript ecosystem, which is the largest— JavaScript and TypeScript are the largest developer ecosystem in the world. It's the second largest AI module behind OpenAI. So we created the Next.js of AI because we want to respond to that trend really quickly. And The AI SDK and what we call the Chat SDK, which gives you the UI primitives to build your own ChatGPT-style products, that gave us a ton of momentum on the cloud infrastructure side because we realized that's what people want to build today. What people want to build tomorrow is these agents, these token factories, these MCPs, and we're meeting them where they are. But what's abundantly clear is that there's no going back. We've accidentally had product-market fit in certain areas in the past. I'll call two big ones that helped us actually go from like, the 1 to 50 was insane. It was like in a couple of years, right? And then to 100. COVID made e-commerce a— not a priority, a must, right? So we saw a lot of tailwinds of like the acceleration of e-commerce. Crypto. Crypto is this always perpetual, it's so over, we're so back. And like a lot of the largest crypto companies in the world are on Vercel. But AI is different in the sense of you can talk about temporary churn and like short-term churn and whatever. But even with our individuals trying out things churn, what we've noticed with v0 is that it's a mind virus. You buy it for a month and it's like, you know what, I'm going to churn. And then 3 months later, you're like, fuck, I need a prototype. I'm not going to— it's just like once you learn that way, like you don't go— I mean, obviously we could lose a customer to a competitor. Totally fair. But even that, I call it a win for like AI team. AI. What happens with AI is that once you use it, it's like a new muscle. It's a new efficiency. I wrote an article years ago called You Can't forgo efficiency. Once you've acquired or tried a higher efficiency system, it's impossible to go back. And my thesis is built on the fact that Albert Einstein wrote this beautiful piece on what it was like to be in America compared to Europe. Now begins the dunking on Europe phase of the podcast. No offense, but it was going to happen at any point, at some point. It's in the DNA of this country because he taught— this is before like anything, internet, etc. He talks about how America is huge, but we, we— that played into an advantage because we started creating, you know, high, uh, uh, cohesion networks. So the, the railways and the telegraph, like we needed technology to scale our enormous distances.. And, you know, for him, I think it was like literally a one-way street of, well, once you have that kind of efficiency, how do you go back? And that was my journey personally of like, once I came here, I would go back to Argentina and it was very, very hard to readapt because in the beginning, before I got my O-1 visa, I could only come here on business. And so I could come here to do conference presentations and meetings and things like that. And like, the lion's share of my work, I needed to go home and like, all right, you know, create my standing desk, my computer, my internet connection, etc. And just everything would be slower. Starting a company is slower, doing paperwork is slower, like getting a loan is slower, like phone calls are slower, like, you know, like restaurant services is slower. Literally, we have an unfair latency advantage. I love coming to New York because it's close to US-East-1, which is where the vast majority of cloud services are, right? So you feel the better latency to the internet. And so with AI, the efficiency win is so remarkable that I believe that if you have a good product on the consumer side, even churn that you might see is short term, especially for productivity tools, I think. And of course, you want things to be so high engagement. And this is what we see more on the company side, like when there's economic utility that you use it every single day of your life. Actually, these products, you can use them on your phone, which is remarkable, right? There's the v0 app. We've sort of spoiled, spoiler alerted on X, but the fact that you can take these agents on the go, this is going to open up a whole new world of if you have a desktop app, there's still a potential now to have a mobile app, even if the desktop app doesn't fit, because now you can control computers remotely with agents. And again, we're just at the beginning of that. I think there's very few people that have actually done this. There's a company called Manus AI that has like this, like on your phone, you are controlling remote computers, essentially. So you're going to see a huge wave of this. And this is why I talked about the AI cloud, because that world requires new primitives, new services that don't exist. You need sandboxes for running virtual computers, you need web browsers. So this has a lot of upside to mobile users. This is a lot of upside to consumers. And yeah, it's just, it's really exciting to be powering so many of this. And that explains a lot of our growth.

Fascinating. All right, I want to close on a couple of more personal questions, if I may. One, you're truly inspirational in how you communicate with the community through conferences like this week. You're super active on X. How do you think about the importance of that in a context where you have like a million other things to do?

Yeah, I think of it as number one, when I speak on X, I speak first to myself, second to my team, believe it or not, right? Like it's actually like a highly scalable communication channel to the almost 700 Vercelians. I speak to our community of builders. I want to give them ideas. I want to support them. I want to engage with them and get feedback from them about the product. And so that's kind of where, like, you can call it inspirational. You know, I tried really hard to bring positivity into the world, and I try to always recognize that, you know, in the short term, things look very unclear to people. There is a 'it's so over, we're so back' like side of the world. And especially for developers, this happens a lot. Like one day, you know, like everything's great, the next day everything is awful. So I try to bring that optimism into the world that translates into there's always more things to build, etc. And because I'm the first customer of what I say, I'm sharpening my thought process. I'm minimizing the number of words. You know, one of my internal principles at Vercel is we sweat every pixel. Not in the sense of like we want to make it look good, which we do. But my principle of sweat every pixel is every pixel is a liability. Every little feature, button, text, link, etc. is a huge liability. It's product surface area, it's things that can go wrong, it's things that can confuse users. If you look at ChatGPT, one fundamental advantage they have over Google is the simplicity of the UI. You go to ChatGPT and notice that there is a concerted effort to remove stuff. Go to Google and you have the search stuff is somewhat clean still, but they have a very uphill battle. You have AI mode and AI overview that shines in front of your face and links and ads. It's This is just overwhelming. So I try to do the same for words. I try to do the same for my thoughts, sharpen my thinking, speak clearly. It's almost like practicing for AI, like, you know, like I called it out recently. It's like, I do believe that in life there is a positive correlation between prompting and success. The more you prompt, The more you get your ideas out there, the more you hop on sales calls, the more you market. And obviously you have to do this with taste. You can't just do it like intention, every time 1% better, etc., etc. But by prompting and taking in that feedback, you're actually making progress. So, you know, the way that I communicate with the world is that every single time I'm trying to like refine what I say and make it a little bit better and then see what feedback I get from the world.

How do you manage negative feedback when you get it?

Because you're building and thinking in public. So my approach to negativity is like every other human being, I have my own. And so I try to catch myself, like, am I going to say something mean or negative about a person or a product? And literally just don't write it or leave it on your drafts, right? So try really hard to stay positive. And I think negativity is a useful tool. You can't just be Mr. Positive all day long and You have to deploy it when it's needed and when it's useful, especially with feedback with people, like, regulate the harshness of the feedback so that, yeah, hey, this really matters to me, and et cetera, et cetera. But I do think that it's really easy to get engagement through negativity. And so I try to resist that. And then when I meet negative feedback, it's always an opportunity, as long as it's not a troll, from a troll farm orchestrated by Cloudflare. Just kidding. You know, I engage in as much as good faith as possible. There's— I always say there's a kernel of truth in everything a customer says. Another operating principle of me, which is more of a traditional, I think, American business principle, is like, the customer is king, the customer is always right. And so if the customer is annoyed and hating on you, they're still right.. And so get to that kernel of right and truth. Sometimes it can be frustrating because when people are acting in bad faith, and this is a promise of recently on X, and I would say probably the larger internet, is that you don't really know who you're interacting with. And so you don't really know if the other part is engaging in good faith. So you don't know if they're an AI bot increasingly, and again, if they're engaging in good faith. But provided that I calibrate my compass to the good faith principle, I dig deep as much as I can to try to understand. I think this— what helps me is that because I've been a developer for so many years and because I build things, I understand the insane amount of frustration that these tools can produce in people. And so Again, try to meet with as much empathy as I can and then try to fix the product. At the end of the day, I'm in this world, I very firmly believe this, is I'm not here to create fans of Routagee. I'm here to create amazing products that make people's lives better. And the way that Routagee wins, I get a lot of joy when people say, I love Vercel. Like someone tweeted today, I love Vercel. Like, hey, that's awesome.. And the reply was, it was because of a positive product experience. It's not because they like me as a person, right? And so try to make the best possible product for ourselves and for our customers. And again, the negativity is part of the process.

How do you do it all? So you're a CEO of an almost 700-person company, you're 10 years into the journey, and here you are bursting with energy and passion. You tweet a lot, as just mentioned, by the way. You're also a very prolific angel investor with a lot of great companies. You're a dad of 5. How do you not lose your mind? How do you manage your stress? It so happens we're recording this on a Sunday, by the way. You could be walking around the streets of New York, but you're investing time in doing this, which we immensely appreciate. Thank you. How do you, how do you balance it all?

So number one is I owe it all to my team because going back to like, you can't forego efficiency. My team and my family create efficiency for me. So, you know, I try to remove all the stuff that would be a waste of my time so I can focus on the essence, on the signal. And so in order to produce signal for the world, I have to capture that signal myself and that requires, by definition, tuning out all the noise. The other one I can't just emphasize enough is just how much investing in your own health and well-being will help you create more energy. I'm not sure how this works, right? But like, it's ATP in the cells and whatever, but like, it's kind of amazing that spending energy in a hard workout somehow creates more energy for the day. And so today I had like a 1-hour run that— you know what's fascinating is every exercise, every workout is like a spiritual experience because I wanted to quit 20 minutes in. And you can start— I never done therapy, but I have a lot of friends that have done. In fact, it's huge in Argentina. Everyone has a psychotherapist. There is cognitive behavioral therapy, which is, if my understanding is correct, is the— it's rooted in paying attention with the help of the therapist to your thought patterns, using AI lingo, to your thinking tokens, your reasoning trace, whatever, and spotting bad thought patterns. What's interesting when you work out is that I'm always fascinated by, like, holy shit, it's a negative thought pattern after negative thought patterns. Like, every single workout, 25 minutes in, I'm like, why are you doing this? Like, you could just be eating breakfast. You could be eating a lox bagel right now. Uh, that's literally my thought today. It's like, I woke up, I was like, there's two paths today: lox bagel, or I'm gonna go to the gym. Uh, because one thing in SF we don't get in any remote, like, semblance of is a good bagel and lox. I'll give that to the city. Okay, one for New York. Yeah, so, so that's the thing is that, uh, I love, um, there's a David Goggins quote of like, every day we have an opportunity to find out who we are as people. And they're like, you know what's fascinating? This is also true of the world of startups, is that I'm competing with the next, you know, garage 20-something years old. And I take that extremely seriously. Like, it doesn't matter how big your company is, how much history you have, how much money in the bank you have. You're still competing with the seed of an idea of a vibe coder in a garage somewhere.. And so that vitality and recognition of the game and respect for the game is what actually makes you realize that every single day you need to show up your best. And that doesn't mean that I don't have weak days or like bad days and whatever, but like, you know, part of, part of this, this being in this world requires that. So it's kind of like a give and take. It's like, of course, I do all these things, but if I didn't do them, I think I'd be in a bad spot.

All right, last question. What advice would you give the 18-year-old version of yourself? I know you started coding super young, so maybe for you that's 15 or 12 or 8.

I'll give the advice to the 10-year-old. I was actually thinking about this today. So my first website, I believe it's so hard to actually pinpoint the first event, but I believe it was a website about Dragon Ball Z, an anime that I loved. I still love, actually. And it sadly looked like shit. It was on a platform called GeoCities, or might have been on Fortune City. There was another one. And I'm almost jealous because the reason that this thought popped into my head today is that I was playing with image generation models and video generation models, the quality of the first thing, your first X that you can create today is infinitely higher than what I could do back then. Infinitely higher. And so that means that when you are getting into this industry, if you're getting in today, you now have access to creating things that are infinitely better. So my advice would be, focus a lot on getting to know everything that's available in the new world. There might even be a temptation to learn the old way because it may have temporary, temporarily higher reputation and things like that. But where I found alpha in my life was in actually not immediately going to the thing that was like the obvious high reputation thing. When I started using JavaScript, it was considered to be a toy programming language and it couldn't be fast and it couldn't be correct and it couldn't be useful. And all of those things over time got proven wrong. And then because I started early, by the time I was 17, Facebook was trying to hire me, which is insane because like a recruiter reached out when I was in Argentina and I couldn't even legally hold a job. right? And so it opened so many doors to just be slightly ahead of the curve and not pay so much attention to, like, what people consider to be high status. My dad at the time wanted me to learn Java and other things like that, like C++ and Java. And, like, he would come up with a new idea every day, and I was like, nope, like, this little thing here one day will be pretty freaking big.

But you would still learn to code? You still learn the core principles of computer science? I would learn how things work.

That's the thing. Coding will be done by agents. So I would learn how it works enough that I can tell an agent, this is what I want. It's very important also to understand the limits of the system, knowing just how fast things can be, for example. Knowing what the hardware is capable of in theory, just ideally, you start with the universe, you start with like Planck distance, and you start with, you know, speed of light. And what are the limits of our simulation? Then you get into what are the limits of computers? Then you get into what are the limits of services that exist, like the things that actually make it plausible to use these computers. Then you get into the limits of LLMs, and that requires a lot of tinkering and exploration and reading and actually a lot of prompting, a lot of curiosity. I do think that there's a bunch of companies that have voiced, we will only— we will use AI literacy, or we will judge AI literacy in our screening when we hire. Shopify said as much. And you guys are in Shopify because I thought I saw a lot of Shopify. Zapier.

Nice as well. I mean, Zapier said that as well.

Yeah. So there's a bunch of companies that have said AI is going to be part of our screening. I actually think it's phenomenal. Not because it might get misconstrued as like, oh, if you're not on the latest fad, whatever, we're not going to hire you. But I think prompting is the biggest proxy for how would I measure someone's curiosity? Someone that's curious would be— I mean, this is again going back to that child, like, holy shit, you're gonna now understand everything in the world, how a quantum computer works, how derivatives in Wall Street work, how LLMs are trained. You could ask an AI, how would I build my own Vercel? You could ask, how do I build my own v0? And so if you're not prompting a lot, I'm sus. I'm sus that you're curious. I'm sus that you're learning all the time. And again, like coding is one of those fundamental skills. I think again, coding is not attractive in the sense of like you writing the words, it's you deploying the systems. This is why Vercel, like one of the key words in the lingo is deploy, deploy, deploy. Like the world moves, makes progress and moves forward one deployment at a time, not one line of code at a time. Line of code— code is inert. Code needs to run to be useful. And so I wouldn't take too much. I wouldn't get too married to what the code looks like. You know, when I started developing, like just what the code looked like was really important to people. No joke. Part of it was motivation because they told you like sitting down and developing and holding attention for long periods of time and getting the negative feedback of the compiler, etc., is like gruesome. So in an ASMR, a pain relief thing that people would do, I think it's just like making the code look pretty. Aligning your imports and the code and the syntax highlighting theme and just like the, the arrangement of your tools. So the reality is that, you know, formatting code, for example, completely irrelevant today, even before AI. Writing the code is going to become less and less relevant over the years. But the systems that the code is orchestrating, those are arguably more important than reality, right? Like the fact that you can, you know, deploy on Vercel and Fluid will help you scale from 0 to a trillion visitors and you'll never, you'll never see that hardware. You will literally never see what it looks like. So the world advances by creating those abstractions. And so just like code has been getting higher and higher levels of abstraction, the ultimate abstraction will be that you're thinking about it and then some outcome is getting produced and the code itself is not as important. My kids, they're on summer break. It's math summer break. So I still— and they love it, by the way. Just like I'm not making them. Uh, so again, math is a great way to develop a systems thinker. It's a great, like, math is built up of mathematical objects and there is a syntax for it and there is, you know, patterns to it and it's a tool for thought in many ways. It's one of the most fundamental ones. So like, learning that. And again, I would use the hell out of AI to learn it. And so I think it's still very important. You know, I'm very thankful that I went to a school that put a lot of emphasis on language. So the two first entry exams that I took for my high school in Argentina were math and language, and the Those were the two ones that would decide who would get into the school or not. Later we had geography and history, but it was well understood that geography and history, everyone could get 50 out of 50. With just like raw, like willpower, you could get that. But what would decide winners and losers was math and language. And so for whatever reason, I think in some intellectual communities, language has been frowned upon, looked down upon, whatever. Whereas I was like, holy shit. Like, first I loved that we would do like syntactic analysis and things like that. It was like grammars and whatever. I was like, I was so comfortable with that because of programming. I was like, nice. Uh, but even just like creative writing and, and the skill of weaving a story and the structure of stories, like the, the sort of like setup of, of the scenes and the development, the conclusion and I mean, if you're building a pitch for an investor or whatever, like, you need to be competent at that. So, yeah, so that's something that you can increase your exposure hours to really high quality writing, which, by the way, I think AIs are not there yet. I don't think AIs are at the point where they can write a Borges quality story. Maybe not even close. Right. They will get there, I think. But, um, you know, I'm a huge Borges fan and sort of, um, there's a, there's a class of, um, or, or branch of literature that happened in Latin America called magic realism with Guillermo Rauch and a bunch of others. Like, reading those things continues to be super important. Like, exposing a kid to, to all those, uh, things will be massive.

Guillermo, thank you so much for going super deep with us. That was fantastic. Thank you. Thank you. Hi, it's Matt Turk again. Thanks for listening to this episode of the MAD Podcast. If you enjoyed it, we'd be very grateful if you would consider subscribing if you haven't already, or leaving a positive review or comment on whichever platform you're watching this or listening to this episode from. This really helps us build the podcast and get great guests. Thanks and see you at the next episode. Next time! Next time!