七年做成的一夜成名：Bolt 如何在四个月做到四千万美元 ARR

Bolt, the product from StackBlitz, is a text box where you describe what you want to build and get a full-stack web or mobile app, with databases and hosting, in about a minute. Eric Simons says that when it launched the company was on the verge of going under after seven years of building web-based development environments. The team hoped it might add $100K of ARR over a couple of months.

Instead, Bolt went from zero to $20 million of ARR in its first two months, crossed $30 million, and was approaching $40 million at the time of recording. It added three million registered users in a few months, with around a million monthly actives. The company's forecast for the year is $100 million of ARR, and Simons says the only comparable startup in AI coding is Cursor.

Simons admits he was skeptical himself. For about a month he woke up expecting the surge to end, and people he spoke to kept saying the number could go back to zero. The data never turned, and he argues that rational buyers are paying because the tools let them do far more for far less cost, which may become the new normal as AI improves.

The team is only 15 to 20 people, with a three-person support team facing close to 100,000 customers. Simons describes the scale of demand as mind-boggling and says everything had to be rebuilt quickly to keep up.

In the demo, Simons types "Make a clone of Spotify" and has a working app within about 60 seconds. He says Bolt's biggest difference from rivals like Lovable, v0 and Replit is subtle: speed and reliability. Other products run development environments on cloud servers that can take minutes to boot and can get stuck, while Bolt runs an operating system locally in the browser using the user's CPU.

That browser-based operating system is WebContainer, which took five to seven years to build and make dependable. Bolt's AI agent talks to it in both directions: it writes code, runs the dev server and sees what happens. Several engineers on the team came from machine learning backgrounds, so they already had the expertise to merge the agent with the operating system.

Bolt also has built-in integrations with Netlify for hosting and Supabase for databases. One click on deploy runs a production build inside the browser, at no cost to StackBlitz, and publishes the app to a real URL. Simons realised that people building wedding or donation sites could use this instead of Wix or Squarespace, because those tools are complicated and cannot produce a functional app.

Simons demonstrates a feature launched the week before, built with Expo, the React Native tooling company. Bolt boots the Expo toolchain, generates a mobile app and shows a QR code, so the user can scan it and run a live version on their phone. Both he and Lenny Rachitsky end up with an app that looks like Spotify on their devices.

As the user keeps prompting, the app reloads with new functionality, so two people on opposite sides of the planet can touch the same working app rather than look at a screenshot. Simons says this changes product development, since it can be faster than designing many Figma frames.

Most of Bolt's audience are PMs, designers and entrepreneurs, people who have always been good at building products but previously needed a developer to turn ideas into code. Simons says that for the first time they can produce production-grade full-stack web and mobile apps without developer knowledge.

Simons says WebContainer was the bet the whole company was founded on: a WebAssembly-based operating system that boots in a browser in around a hundred milliseconds and runs full development toolchains. The idea came partly from watching Figma's early days. Its founders first pitched a graphics engine built on WebGL, with no design tool yet, and it took years to create.

Around 2016-17, Simons and his co-founder Albert saw the same pattern for development environments. New browser capabilities like WebAssembly, shared memory and service workers made it theoretically possible to run Node.js and npm on top of a WebAssembly operating system. That would remove the painful setup that makes beginners quit and that eats the first month or two at companies like Netflix.

He argues that productivity apps that scaled to a billion users, such as Figma and Google Docs, use the user's own CPU and memory. Cloud IDEs and today's text-to-app tools need a VM per user, which is a problem at scale and with free tiers, where people mine Bitcoin or run DDoS attacks on your servers. Doing everything on the end device avoids that.

The harder part came after WebContainer was built. Developers loved it but did not use it in ways they would pay for, and Simons says the team had to accept that cool technology alone is not enough to build a venture-scale company. The company kept taking shots on goal until one connected.

Rachitsky points out three lessons: the team built technology first and found the problem later, the moment resembles the arrival of AJAX, and founders should survive as long as they can. Simons agrees. He and Albert had churned through many unsuccessful startups, so they had conviction that the web, which has no way to build the web, needed a tool like Xcode or Visual Studio.

Their previous company was bootstrapped all the way to acquisition, which taught them to live hand to mouth and make each dollar stretch. StackBlitz did not raise money for its first two or three years, and when it did, it barely spent it. Simons's rule is that until people are pulling the product out of your hands, the default answer to spending should be no, including asking vendors to sell at half price.

He says the company also resisted the advice in 2020 and 2021 to triple headcount and raise burn, and that following it would have meant going out of business long before Bolt. Entrepreneurship, he says, means making judgment calls that are not the consensus view and having confidence in your convictions.

Bolt launched with a tweet. On day one it added about $60K of ARR, roughly 10% of the company's $600K base, and the next day about $80K. Simons says the product had been built in 90 days and was missing basics, to the point that it reached $20 million of ARR without a mobile-responsive view.

StackBlitz had only ever had one $9 plan, which people burned through in 48 hours before asking how to buy more. Within a week the team rolled out new pricing tiers, which Simons says other companies in the space have since copied. His reasoning is that all-you-can-eat models force inference costs to stay low, which rules out agentic experiences, while users who see clear ROI will pay for more.

The servers were melting, and Simons says Anthropic ran out of GPUs for them. Dario Amodei emailed to say there was nothing more to give. His chief of staff and he spent about 95% of the day on customer support. Normal playbooks for scaling to $20 million of ARR suggested taking six months to a year, which was not an option.

Asked how 20 people handled this growth, Simons credits the technology first and then the people. It is rare, he says, to find a startup where five, six or seven people have stayed for five years or more, and that stability is what allows long bets. Having fewer people with more context, a habit from bootstrapping, gives high trust and lets engineers fix a customer's issue, design the UI and ship it without involving anyone else.

When hiring, he looks for people who do not care about titles and who leave their ego at the door, rather than brilliant jerks. The company is fully remote, with most hires in Europe, and many early hires came from the StackBlitz user community. A tweet saying they are hiring an engineer is often enough.

On prioritisation, Simons says some bets cannot come from user requests because users do not know what is possible. Native mobile app support, launched the week before, had internal doubters but became the biggest launch the company has seen, and thousands of mobile apps are now created per day. He compares it to a chef balancing customer feedback against a dish they believe is a killer.

The whole company meets on Zoom at 8:00 AM Pacific every day for at least an hour. Critics say it is the most expensive use of time, but he says it means zero fidelity loss, so people can make correct decisions independently. He does not expect to keep doing it forever and says they will splinter into smaller syncs as the team grows.

StackBlitz uses Linear for engineering, Notion for product roadmaps and PRDs, and increasingly Bolt itself for design and prototyping. Simons keeps PRDs light, with only enough context to align everyone on key outcomes, and often a PRD is just a link to a working Bolt demo. He says a live demo is worth millions of words, and businesses now use Bolt mainly for high-fidelity prototyping because it is faster than dragging frames in Figma.

On what people have built, he cites Paul, a non-coder who built an AI-powered CRM with Stripe billing in three weeks for about $300 of Bolt usage, after an agency had quoted $30,000 and six months. Paul now sells it. For greenfield projects, public websites and marketing pages, Simons says Bolt can replace Webflow, and for existing products it speeds up the product development process.

Bolt can open existing repositories, but he says LLMs are not yet reliable on very large apps with a thousand files or more, where you still need something like Cursor and real developer skills. He expects that answer to differ in a year.

The most successful non-developers are great PMs, because defining scope, speccing clearly and helping debug is exactly how you work with an AI developer agent. About 67% of users are not developers. Simons expects org charts to change as PMs write code directly instead of Jira tickets, with developers handling harder problems and reviewing code, and he says this is not yet priced into software companies.

Rachitsky agrees, adding that product-thinking founders are in the same position and that engineering and design skills remain an advantage. Simons imagines pods with more product and design people and one or two engineers.

Asked what a child starting college should study, Simons says to learn how to leverage AI tools and get a basic understanding of programming, but to follow intrinsic interests. He and his co-founder skipped college, and he says he might not have learned to code as deeply had Bolt existed when he was 13, since coding was just a requirement for building products.

Rachitsky summarises the skills as figuring out what people need, articulating it clearly to AI, growing the product, and helping AI get unstuck. Simons notes that StackBlitz just announced Bolt Builders, a program of certified humans, comparable to the genius bar at an Apple Store, who help non-developers when the AI gets stuck, for around $50 an hour.

Simons says Claude Sonnet was the first model that flipped the equation, and Bolt's rapid growth started the moment it went online. The team had tried building Bolt about a year earlier with frontier models and it did not work: apps broke or looked ugly. After a preview of Sonnet in May, they took the project off the shelf.

He explains why coding is special. Law, for example, is not deterministic, so reliable training data is hard to make, but code either runs or does not. Anthropic, he says, ran reinforcement learning on permutations of every kind of app, so he is bullish that LLMs will improve faster at code than most other fields. Everyone is now chasing coding because rewriting the software world is a trillion-dollar opportunity.

Simons says Bolt will focus on fitting the tools PMs and designers already use. One feature, built with Anima, lets you put bolt.new in front of any Figma URL to turn the design into a full-stack or mobile app. Another is a Bolt Slack bot that acts like a developer on your team, reading a thread, converting Figma links and returning a URL where you can keep prompting.

His own engineers mostly use Cursor, plus Bolt for product development, with Claude and ChatGPT subscriptions. When asked about advice for a first-time user, he says to talk to Bolt like you would write a Linear or Jira ticket: be specific about what matters and leave room for creativity. For a first try, he suggests pasting your LinkedIn bio and asking for a personal website, then deploying it.

Finally, he tells the AOL story. In 2012, at 19, he was building a K12 education startup in an Imagine K12 office inside AOL's Palo Alto headquarters. When the $20,000 ran out, he secretly lived there for four or five months, using the couches, ramen, showers and laundry, until a guard threw him out at 4:00 AM. He says his burn rate then was about a dollar a day, spent on the McDonald's Dollar Menu.