From Zero to $20M ARR in Two Months: Bolt's Moat Lives in the Browser

Sarah Guo opens by saying that in all her years investing she has never seen growth like Bolt.new's jump from zero to $20 million in annual recurring revenue in two months. Eric Simons, co-founder and CEO of StackBlitz, agrees it has been surreal and far beyond anything the team expected. For anyone who has not seen it, he describes Bolt as close to a ChatGPT or Claude chat box, except the job is to build a full-stack web application from a prompt.

Landing pages, blogs, and apps with real authentication are all in scope. Instead of going to a web-development agency, you put in an idea, hit enter, and get a production website. Simons puts the global developer population at about 25 million. In the previous week, he says, nearly 200,000 people the company calls software composers used Bolt to build web apps.

He thinks that number should be 100 million, and the current growth clip makes that horizon feel closer rather than theoretical. The distinction he keeps drawing is that Bolt produces real software — login, functionality, something you can actually use — not the drag-and-drop static sites of the previous era of how the web was made. Developers, designers, and people who have never written code are all doing it in the browser.

Plenty of tools generate code, including the frontier chat products themselves. What is special about Bolt, Simons says, comes from the origin of the company: they have written an operating system in WebAssembly that runs in the browser. If you want a real web project, you need to install arbitrary packages and run toolchains such as Next.js or Vite. Doing that on servers is complicated, slow, and billed by the minute.

Doing it in the browser is extremely fast, has no round-trip latency, and does not create a cloud invoice. Bolt's product is that runtime married to frontier models. Claude Artifacts, he grants, was one of the first things on the market that did a good job of "build me a UI," and he still uses it weekly for graphs. The break comes when you want something more meaningful: a landing page where people log in, with Firebase or Supabase installed, wired up, and actually deployed. That full loop, with no extra setup, is what Bolt is uniquely set up to do.

Guo raises the standard pushback: a lot of code-generation companies are a system prompt and a well-SEO'd website. Bolt's unusual move was to open-source the system prompt and a lot of the product code. Wrapper companies, Simons argues, tend to disappear when the next model ships or when the labs fold the capability into their own chat products. The five-year WebContainers stack is the part that is actually hard. The prompt, he adds, was going to leak anyway the moment someone got the model to dump it.

Coming from web development, they also felt the open-source world was missing a real working AI product that people could fork, improve, and contribute to. If Bolt becomes a great business, it will be because the team grew extremely quickly and built the best end-to-end experience that actually works — not because it hid glue code. He compares the leak to View Source on a webpage: you can go to google.com and read the markup, and so far nobody has used that to build another trillion-dollar company that took Google out.

StackBlitz has poured years into open source, but shipping Bolt still felt like a test of whether anyone would find the code valuable. They did. For AI tools in general, Simons says one of the hardest jobs — even for the labs — is teaching people how to use a non-deterministic model. Showing someone Gmail is simple: hit compose, type, send. Prompting is not, and every message costs money, so users who cannot find a working pattern simply leave.

Bolt treated community as product work. People share what actually works, and the company has had to admit it is no longer the expert on its own tool: power users know more. The team brings those users onto livestreams, then tries to upstream the prompts and workflows back into the product. Skip that loop, he says, and churn on early AI apps sits at 60 to 70 percent, which looks like the default for this generation of products.

Guo draws a parallel to David Holz and the Midjourney team: a cluster of obsessive, capable power users teaching the whole community how to use the models, and showing what can be created, is a large part of Midjourney's position among many image generators. Simons calls that the best example he has seen, and notes there are surprisingly few visible ones. For startups that want to win with this kind of product, real community investment is more critical than ever.

Most people on Bolt are already non-technical. The ones who succeed tend to have managed development teams — entrepreneurs, PMs — because managing an AI turns out to be extremely similar to managing software developers. Existing eval suites, meanwhile, are good at well-specified coding problems and bad at "build me a landing page." The open-source Bolt Local build has become a way people at Hugging Face and elsewhere drop in new code models, including a recent NVIDIA release and Qwen, and ask how they compare to Claude 3.5 Sonnet. It reminds him of the 2000s PC question: can it run Crysis? Now the question is whether a model can run Bolt.

Online just under two months, Bolt has already seen the first startups launch from the tool and charge customers on Stripe. One of Simons's favorites is a product manager at a banking software company in Thailand who moonlights on ViralHooks.ai. The product helps short-video creators write the opening hook that keeps people watching, and reverse-engineers how other successful creators did it, using models she trained on top of OpenAI and similar APIs.

A week before Bolt launched, she listed the project on Upwork and got a $5,000 quote from a developer in Ukraine, with a two-to-three-month timeline. Guo says that actually sounds cheap and realistic for the app described. The next week she signed up for Bolt's $50 plan and built and launched the entire product in two weeks. That is a 99 percent reduction in cost, from $5,000 to $50, and five times faster delivery.

It was not a one-off. A man named Paul, who has run web-development agencies for about twenty years, launched a fully featured CRM called Chilled CRM, with calendar, contacts, and a built-in AI chatbot. Building that CRM used to be a $30,000 quote. He did it on a $200-a-month plan in one month. Agency shops, Simons says, are now punching out client apps faster than ever and still charging the old price. One tweet he saw called it the most incredible arbitrage opportunity in web development ever.

Guo has long been skeptical of traditional no-code: a GUI editor in a closed ecosystem that traps you, cuts you off from open source and frameworks, and never makes the system's bounds clear. Simons shared that view six months earlier, and even three months earlier. What changed was the models. There was a tipping point with Claude 3.5 Sonnet, a chasm he says has been crossed: the systems became good enough to write production-grade applications, and they are only going to get better.

Inside baseball, he says they had the idea for Bolt around February and tried to build it with the frontier models then available. It was not possible. The models did not spit back accurate code; things constantly broke, which ruins the experience, so they put the project on the shelf. An early preview of the new Sonnet work felt like it "changes everything." Classic site builders exist, in his telling, because a custom WYSIWYG canvas was the only middleman that could turn a non-technical idea into code. That path locked users into walled ecosystems. Now the best interface is a text box.

The week they launched, a salesman in Dallas tweeted thanks: he had made a website for his daughter, who has a medical condition and needs to send information to donors as she travels. Simons's first thought was to tell him Wix exists. Then he remembered that Squarespace is itself complicated. In 2021 he tried to build his own wedding site as a developer, spent a Saturday, did not finish, and his wife bought an account. On Bolt you describe the date, the details, and the RSVP, hit enter, and get a production site. His 71-year-old mother built and launched her first website three weeks before the interview.

The code coming out is Next.js, Remix, Astro, or Vite — the same stack developers already use. When people get stuck, they raise a hand in Discord or on Twitter, and other developers book hourly help. A professional can pick up the repo because it is a normal codebase. Existing engineers use Bolt the way they used every other tool of the last twenty or thirty years: skip punching out a UI, iterate fast, pull in data, and if they need to, drop into Cursor and bring the work back. A large number of people downloading Cursor are not developers; they click Accept Change. After Bolt launched, YouTube filled with "Bolt kills Cursor" videos. Simons finds that funny. They are different products that happen to solve "I cannot code" for the same non-technical audience.

Overnight success, Guo notes, usually sits on a long grind: Notion spent five years finding Notion, and StackBlitz has been at this for five-plus years as well. Simons co-founded the company with his childhood friend Albert. They grew up in a Chicago suburb, building PCs, and for their thirteenth birthdays in the mid-2000s asked for $200 O'Reilly books instead of an Xbox. There was no Codecademy and little in the way of free tutorials. Coding was simply how you brought product ideas to life, and they have been building together for fifteen to twenty years.

Around 2016 and 2017 they realized browsers had hit an inflection: you could run an operating system in a tab, fast. Every successful platform can build its own apps — Windows apps on Windows, Mac apps on Mac — but the web had no built-in way to do that. They had watched the same story with Dylan Field's first Figma pitch: not a design tool, a WebGL demo of a 3D ball dropping, and the claim that browsers were finally powerful enough for meaningful 3D, which meant a design tool could live entirely in the browser. WebContainers took about four years, with Dominic Elm in Germany leading the engineering. StackBlitz now sees about three million developers a month; the original product was VS Code in the browser, powered by that runtime.

Guo, an early Figma investor, remembers Googling Simons and finding that he had been living in an AOL building for the free food and showers. He was nineteen then and is thirty-three now, with a wife, a daughter, and a dog. He had come to Silicon Valley with literally zero dollars, through Imagine K12, an education incubator that later folded into Y Combinator, with access cards into AOL as the company tried to host startups. After the money ran out he slept on couches, showered at the AOL gym, and ate leftover catering from the fridge. A contemporary article quoted his burn rate as a dollar a day. He thinks that story is why AOL shut the program down.

About seven months before the interview, the company was in a tough place: not about to run out of cash, but without a clear growth-oriented revenue product, while peers were getting acquired or shutting down. His daughter was born in April. Shortly after, he woke up and decided to do a full Ironman that year — a two-and-a-half-mile swim, a 112-mile bike, and a marathon — having never done any of those. He ran a marathon with his brother-in-law as a warmup, hired a coach who had been on the US Olympic team, and two months out the coach said it was a bad idea. In October he finished in the top 25 percent of finishers. Race day was about two weeks after Bolt launched, while the company was scaling at a crazy rate. The line he keeps is the military one: fate whispers to the warrior that he cannot weather the storm, and the warrior whispers back, "I am the storm." His advice is that if the universe is going to try to crush you, make it try hard.

Looking six to twelve months out — past that, Guo and Simons agree, AI forecasting gets mushy — he is convinced code generation is moving past tab completion. Products like Cursor already took line completion further; agentic workflows are here, and Bolt is one of the most visible cases that has actually worked. Engineers will instruct models at a higher level than "complete this line": go do this, and send it back. Software composer, the phrase Bolt uses internally, is the name he thinks that job starts to take.

The other conviction, and a big reason they bet on Bolt, is that models will keep getting better at code generation specifically. Training needs more data, and it has to be accurate, which is hard in domains that are not easy to grade. Software is different: the code either executed without errors or it did not. A landing page either generated or it did not, and you can capture an image and analyze it. Frontier labs, in his telling, are trying to put every permutation of every application into the training data.

People have been worrying that the models are hitting limits. He does not think so, at least in this slice. The past year already showed a lot of improvement, and he expects more of it, rapidly. What can be done with the systems we have now is already unbelievable, he says, and it is only going to keep getting better. That is the note he wants to leave on.