v0 Hit 3 Million Users. Vercel’s CEO Says Pages Will Be Generated On Demand Within Five Years

Guillermo Rauch, CEO of Vercel and creator of Next.js, joins Sonya Huang and Pat Grady on Sequoia’s Training Data to talk about how AI is reshaping software development. He has spent years obsessing over the distance between an idea and a live URL. With v0, Vercel’s text-to-app generator, he thinks that distance is collapsing for people who are not even developers.

The name he gives the destination is the generative web: applications created on demand for individual users, with traditional download-and-install software slowly leaving the stage. Along the way he shares a number that still sounds strange for a company that sells cloud infrastructure and frontend tools. Signups arriving from ChatGPT have been growing exponentially, and at the AI Engineer conference in San Francisco, people walked up to the Vercel booth and said they knew the product because ChatGPT told them to use it.

When ChatGPT launched, Vercel noticed something specific: the model was extremely good at writing React and Tailwind, the stack most web developers already used. The usual reactions to that kind of discontinuity are ignore it or fear it. Rauch chose a third path and went all in. They built v0, a text-to-app product that takes a description and returns a running frontend.

A couple of years later, he calls the growth astronomical: more than 3 million builders, high engagement and retention, Fortune 10 companies on the enterprise tier. The bigger discontinuity, in his view, is who showed up. AI coding tools are attracting what he calls “dev-adjacent” profiles—designers, marketers, “basically anybody with a keyboard.” His running joke is that instead of yapping into a team chat app, you can yap into v0 and create more value.

The cost collapse is concrete. Seed rounds used to exist so founders could build a first prototype. Rauch now hears that v0 prototypes are replacing pitch decks. Huang had just sat through a Series B whose frontend was built in v0; by the time you get to a pitch, he says, it is rare not to have a working frontend. Iteration velocity changed with the price: you can go through hundreds of prototypes before you settle on an idea. Vercel open-sourced the underlying AI SDK, published how they build the product, and recently released a v0 model so other founders can spin up their own site-building agents.

Pat Grady asks how you systematize quality when so much of it looks like taste. Rauch’s first answer is that AI products come with a feedback loop built in. GitHub Copilot had a single gigantic metric: completion acceptance rate, 51% this week, 52% the next, then swap the model and watch again. v0 has more of those signals: does the code run, does the page render, did the user deploy into the real world, and did anyone actually use the app afterward.

The subjective half he still calls taste. On iOS, the Safari theme bar should match the page background; most people cannot name the mismatch, but when the colors line up the canvas takes over the whole screen. Frameworks used to encode that kind of detail through upgrades and docs. Now Vercel writes it into the model: when you generate a landing page, just do it. Frameworks worked hard to put you in a pit of success on performance and security if you stayed inside the rails. LLMs, he argues, can give that guidance to a much broader set of people than the single-digit or double-digit millions of developers.

That is why he calls this the thing after frameworks. Frameworks hit diminishing returns; he used to reduce Next.js talks to a math problem—minimum steps, minimum characters, from a folder to a first React component. The new question on the v0 prompt is simply “what do you want to ship?” People entering software today, he says, leapfrog a generation that earned gray hairs the hard way, because they inherit what everyone has already learned. Vercel itself started from his own pain: putting a site online meant configuring cloud from scratch and stitching tools to infrastructure, which is why Next.js existed. Chapter one was infrastructure on autopilot, with developer experience as the Trojan horse. Chapter two is automating the writing of the software itself.

The most practical advice in the episode sits here. If you sell developer tools, you have two customers: the developer, and the person that developer is trying to serve. It is easy to sell plus-ten happiness points—an extra flag, a relaxed quota—and watch the developer cheer. Two quarters later their users are unhappy, and the business did not pan out. Unlimited quotas feel generous until every downstream service has to deal with an unbounded entity; Rauch admits he got that tension wrong in the early days. His meta-belief is not to hold too many strongly held beliefs: whenever someone says AI cannot do a thing, he tries to be on AI’s side, because three, six, or nine months later the situation has usually changed. Most weekends he starts something from scratch, both to dogfood the platform and to ask what friction can still be deleted.

Looking at Vercel’s last two years, Rauch names two dramatic shifts. The first is that the entire user base doubled year over year at nearly zero incremental cost. It took years of being a hyped, successful company to reach the old numbers; AI opened the top of the funnel so wide that the doubling just happened. v0 itself is a full-stack app running on Vercel with no special access and no tricks, which made the company its own most demanding customer.

The second shift is who the new users are. They still care about the end user, but they are less sensitive to whether the generated code is long or short, or to the shape of the APIs. That leads to a sharp claim: LLM strengths and weaknesses will inform the runtimes, languages, type checkers, and frameworks of the future, because your customer is no longer the developer. Your customer is the agent that person is wielding. Stripe’s API will still be beautiful, he says, but you now have to ask whether something should change so the LLM as user can call it more reliably. These users also have a shorter fuse. Developers learned to live with type checkers screaming at them; the new cohort flips the table. That pressure, he thinks, is healthy: you want 99.99% reliability.

He describes a tale of two cities. Underlying model reliability is still low, and AI providers go down more often than people like to admit, yet consumer adoption is flying. It took a long time to get everyone on team internet; team AI did not wait. Infrastructure is being built while demand is already unprecedented. On the product side, Vercel had spent years moving the web from static to dynamic—streaming responses the way Amazon computes recommendations just in time. That work became Fluid Compute, which happens to fit LLM streaming. The next representation is a web for agents: `llms.txt` as a simple alternative document, and MCP as the protocol that lets an agent talk to another agent.

Companies, he tweeted, go from no AI, to an AI prototyping team, to a production AI product team, to every company being an AI company. Support is the lowest-friction on-ramp. Vercel went further and built an expert model of its own stack, then used it twice: it powers vercel.com/help, and it is the product they sell as v0. With more than 100,000 customers, professional services did not scale; selling the agent is a way to wholesale access to his brain and his CTO’s. He borrows Jensen Huang’s language: a token factory, a human amplifier. Firewall, security, and observability are next—an agent that sees 500s and files a pull request with a fix. Fully self-driving infrastructure? “Practically almost there.” The day of the taping, Google Cloud and Cloudflare had wide outages; Vercel’s systems could already investigate autonomously, feed logs into v0, and name which vendor to call.

Why ChatGPT sends so many Vercel signups is less mysterious once you watch the models. While building the AI SDK—Rauch calls it the Next.js of AI—they shipped a playground that talks to several models at once, the Coca-Cola versus Pepsi demo. Ask how to deploy a React app and the models often pick Vercel, because the training data is full of posts, GitHub issues, and write-ups. Marketing used to mean podcasts and articles; now people ask their AI buddy. Models still ground themselves in search, so ranking still matters, but the content should be organized around full questions rather than keywords. FAQ-style pages that match what people actually ask have worked well for Vercel.

On the competitive chessboard he offers two frameworks. The first is virtual coworkers: virtual designers, marketers, and developers, which is to say expert agents. People will still start with ChatGPT or Claude for broad questions, then graduate to a specialist they keep coming back to. OpenEvidence looks like ChatGPT but is grounded in frontier medical data for physicians. GC AI does law, in a space that also includes Harvey; finance has Hebbia, FinTool, and Perplexity. He sees a world of millions, even hundreds of millions, of agents. Every digital-native company has to ask an uncomfortable question: if you were reinvented today, would your interface be agentic?

Agents themselves split into two generations. Synchronous ones—ChatGPT, OpenEvidence—ask and answer. Asynchronous ones work for a long time, collaborate with other agents and humans, and are more interesting over the long run. v0 is already blurring the line: ask for an optimal ecommerce interface and it may research first, then build. An AI IDE is a synchronous surface; lying in bed and telling an agent to cook and come back with a pull request is the async version. Frontends still matter on both sides of that loop. There is an output artifact—a site, a PR, a confirmation message—and an input that has to capture attention. Studying UI, Rauch concluded that agents are their own modality and are hard to merge into existing products, which is why v0 lives at v0.dev, the way Google carved out AI Mode as a separate product, like Maps.

Developers are splitting too: people who create apps, and people who create platforms. App builders get a conversational onboarding. Platform builders still get templates, but they roll up their sleeves. The metaphor is Waymo. v0 is the self-driving car that ideally never disengages; if it does, you eject into Cursor or VS Code. The Vercel platform is the expert-centered driving, with a human still in the loop. As models improve, more of the top of the funnel will land on the v0 side. Rauch himself, with a CEO’s calendar, has not used traditional coding tools in a while. He just uses v0, obsessed with time from idea to online.

Huang asks the sharp question: if everyone can produce software, does the world get better apps or just more noise? Rauch says better, and he has a method. The team continuously analyzes categories of errors customers fall into and asks whether a model could have solved them. If yes, it becomes a strict eval and goes back into the product. The Safari theme-bar detail is one of those evals, now in the default experience of millions of people.

Security is the cleaner case. Vercel’s CTO found a vulnerability in an open-source framework and helped patch it before a broad release. They then turned the pull request into an exam for frontier coding models: do you notice something wrong? Several of them did, on a non-trivial bug. Rauch’s next thought was to let LLMs sit in existing codebases and burn tokens until they found problems like that. Not long after, news broke that a high-severity use-after-free in the Linux kernel had been found almost exclusively by OpenAI’s o3. Seasoned kernel engineers told him those bugs are everywhere; experts simply do not have the bandwidth. His conclusion: LLMs will create code that is secure by construction, and “the next Heartbleed won’t happen in a world of AI-written code.”

He is not only optimistic. LLMs will cheerfully ship database secrets into client-side code. v0 has prevented tens of thousands of those vulnerabilities; last he checked with a research engineer, it was about a thousand a day relative to what the model wanted to ship. The advantage is infrastructure primitives: generated code can go to the server instantly, interactive bits to the client, and the secure actions the docs would tell a human to take can run on autopilot. Performance is the same story. The web is a generator, streaming values over time, and everything that hits the retina has to land in the right place. Apple does that with a culture of human testing. Rauch wants agents rooted in production data to take those tasks, plus computer-use agents that click through the product and find the broken signup flow.

He spent a weekend trying the three hottest speech-to-text products. Two of them failed on the critical path—download, sign up, grant OS permissions. Entrepreneurs always think it is rock solid the day they tested it. Amazon’s CTO had the phrase for this: everything fails all the time. What you want is a QA agent watching the paths that actually matter: buying, signing up, contacting sales, downloading.

Rauch has always fought the idea that an app is a permanent thing. On a Mac you mount a DMG, eject it from Finder, drag it into Applications, and then it is your puppy: feed it, launch it, live with it. He is team web, where everything is instantaneous—open chatgpt.com, open v0.dev, done. App stores add their own latency: hundreds of megabytes and a long wait. The web, he thinks, is the right place to host AGI, and the place where everything will be generative, just in time, for everybody. You will not even notice that the thing in front of you is an ephemeral app.

He now has a physical aversion to searching for software, because generating it feels faster than finding and installing it. If generations keep improving on quality, performance, security, and reliability, “it’s going to be an unwinnable battle for the downloadable, installable, procurable” product. He is on team fully generative software, forever. Developers will make agents and agentic platforms: go to v0 to generate apps, come to Vercel to build the platform. v0 itself might become the generation engine of that future internet; browser integrations are already letting people call it just in time from existing apps. Jensen Huang said all pixels will be generated, not rendered. Rauch thinks the same of the web experience.

The timeline is specific. In five years he expects the biggest transformation of the web in its interfaces, in people’s habits, and in the number of creators. Five already sounds like a lot to him. In three years, kingdoms will collapse: internet-native companies that did not adjust fast enough, and AI-native companies rising to unprecedented heights very quickly.

The rapid-fire round is less about the generative web and more about how Rauch actually works. His favorite AI app outside v0 is the new speech-to-text category—Superwhisper, Wispr Flow—which is a self-disruption for someone who grew up as the kid who typed really fast. The person he admires in AI is Andrej Karpathy: vibe coding plus speaking applications into existence is a vision Rauch thinks is coming true, and Karpathy had already done a version of that future at Tesla with self-driving.

Asked for recommended reading, he points to an essay called “The Five Whys, and Why the World Is About to Get Weirder.” It is about airplane crashes that have become so rare that whatever still breaks the news is the wildest event, going from zero to taking over the world. He reads the same pattern in startups: growth is so fast, and the tactics so weird, that the article feels predictive. Underhyped, for him, is easy: LLMs themselves, a platform shift that requires you to rewire your brain. Kids growing up with this, he says, have psychic magic powers, and he is jealous. Overhyped is harder, because on a long enough view he was even willing to stay hyped on crypto through winters. All money will be digital; all software will be AI.