StackBlitz Almost Shut Down. Then Bolt Hit $20 Million ARR in Eight Weeks.

The numbers StackBlitz has disclosed still sound like a glitch. In the first eight weeks after launch, Bolt went from zero to $20 million of annual recurring revenue and kept growing. Under the product is an operating system that runs natively in the browser, boots in about 100 milliseconds, and never spins up a cloud VM. Eric Simons, founder and CEO of StackBlitz, tells Nathan Labenz on *The Cognitive Revolution* that this is what seven years of a web-based IDE were for. Add an AI coding assistant, point it at bolt.new, and the curve went vertical. The company later announced a $105.5 million Series B.

Labenz wants the worldview before the demo. After ChatGPT, the industry talked as if AGI were a year away. Simons is more interested in what is moving on the ground. Models got good at software fast for two boring reasons: code is text, so it is easy to train on, and code is deterministic. Train a model on law and you are reading cases back to the 1700s and guessing how a judge felt. Write code, run it, see whether it did the thing, and that run becomes a new test case. His pragmatic view is that the world software order is getting rewritten, one-to-one, and that software may be the first major tectonic shift where AI replaces the old stack at scale.

He does not carry a p(doom) into the conversation. Every prior technology, he says, has been an enabler; things get reconfigured. What he sees now is people who could not code shipping real production products, and strong engineers leveraging themselves five or ten times. The world needs more software and has been constrained by headcount. Labenz adds the o3 result: a model in the top 200 competitive coders on Earth, maybe leaving one person at OpenAI with a higher score. He also likes "luxury software" or "selfish software": apps for yourself or your team that burn an unreasonable number of tokens because they are worth it, even if they could never go to a mass market. The tension is that homemade tools still bug out, and polished UIs still win.

Simons agrees the long tail of UX and support is not solved. Equilibrium depends on how far out expensive models get cheap enough to cover that tail. For internal tools and actual startups, it already makes sense, because the people succeeding on Bolt tend to have a keen eye for product: PMs, designers, entrepreneurs. They used to write code through someone else's fingertips. Now they say the same words to the model and get an instant resume. General end users who have never been in the product business will need more inference and more agent capability under the hood. When things break in Cursor or anywhere else, non-technical people still do not know how to fix them. That first big market unlock, he says, is already happening, and it will explode over the next few years.

Today's mix is roughly 40 percent traditional software developers and 60 to 70 percent PMs, entrepreneurs, and other people who do not write code. For developers, Bolt is a 10x on prototypes and UI. They drop in Figma, hit enter, run a couple of prompts, pull the result to Git or to disk, and keep going. Fortune 500 teams use it to prototype instead of the old Figma-plus-engineers loop. The ROI story Simons cannot shake is a freelancer named CJ: a dashboard web app, about $7 of inference on Bolt, $9,000 billed to the client. Search Twitter for Bolt arbitrage and a lot of freelancers have connected the dots. Labenz says he cannot measure how much of that is happening in the real economy, but he can confirm some of it is real. Businesses used to paying tens of thousands are, in some cases, effectively paying thousands per hour without noticing. The market, including the human help StackBlitz wants to bolt onto the product, will pull prices down.

On the PM and founder side, people are launching side projects and startups with payments, authentication, and a database included. The job of a PM has been writing Jira tickets, assigning them, reviewing the work, and sending feedback. The new loop is tell the model, see it immediately, and commit when it looks right. A developer pulling the branch cannot tell whether a coworker or an AI wrote it, and it does not matter. Hard functionality that models still cannot zero-shot stays with the engineers. Aligning buttons, fixing colors, and wiring the click to the right page can ship from the PM's chair.

Labenz walks through the barriers that still sit in front of his dad, who has ideas and no product career: underspecifying what he wants, not knowing how to fix a break, and never making it to a real deployment. Simons does not pretend those vanish. Building a great product is still a skill. What generative tools change is that you can learn it by sculpting with the model instead of getting mired in tooling. He and his co-founder grew up down the street from each other in Chicago, learned to code at thirteen, and did it because they wanted to ship products. If Bolt had existed then, he is not sure how deep they would have had to go as engineers. The end goal was never the toolchain.

A couple of weeks before launch he had his mother, 73, the least technical person he knows, try the product with no guidance. She built and deployed her first website. The interface is a text box, an enter key, and a deploy button that returns a live URL. People have used it for weddings and, in one case, a father raising medical donations for his daughter. Simons almost told the man that Wix exists. Then he remembered his own 2021 wedding site. He had wanted to write it himself, burned a weekend, quit, and got marched onto Squarespace. Those drag-and-drop builders, aimed at less sophisticated users than a professional developer, still felt terrible. That whole category, he says, is going away. There is no reason for it once a text box can do the job.

Making new users successful is three layers. One is education and community, Midjourney-style: people share prompts, and the heaviest users now know the product better than the core team. Two is a human escape hatch. For anything complex, you still cannot rely on the model with zero technical knowledge. StackBlitz is rolling out a program inside Bolt: raise your hand, get connected to a certified expert, spend thirty minutes, get unblocked. Simons is not aware of anyone else doing that well. Three is the models and the agent getting better, which multiplies the first two and reduces how often you need them. A year ago this was not valuable. They have crossed that chasm. Labenz is happy to retire the "no progress since GPT-4" line. Even before the o1 sequence, GPT-4-class models got much more consistently useful at the jobs people actually want, especially code.

Labenz asks for a success rate: if someone sits down to build X, how often do they get to X? Simons points at revenue that has not gone down, at more people arriving every day, and at retention that would look sickly in consumer AI. Direct-to-consumer AI products can churn 60 or 70 percent a month. The acceptable band, he says, is more like 20 to 50. Bolt hangs around 25. Exclude Microsoft GitHub, look at the coding startups, and free markets are a decent scoreboard: Cursor is number one at $100 million ARR, Bolt is number two, and a pack sits in the $5 to $15 million range. Two tools have broken from the rest. Something in the experience is working.

The question underneath every one of these products is where the full-stack app actually runs while the model programs it. The traditional answer, since Cloud9, is a cloud VM per user. That model has never produced a breakthrough used by lots of people. Every keystroke syncs out and back. Someone pays for the VM. A billion users is a billion VMs, which is a lot of capital, and there are not a billion VMs for rent on the planet. There are a billion devices. Google Docs and Figma already understood the move: render and compute on the user's CPU, memory, and GPU, and treat the browser as the interface. StackBlitz wrote an operating system that does the same for a dev environment. It boots in about 100 milliseconds. There is no cloud VM to get borked, no keystroke lag, and the AI punches code into a live machine you can see immediately.

They spent five years on WebContainer; this year makes six. A couple of the engineers who built it had come from machine learning, so when it was time to write the agent they knew how to instrument it, collect real usage, and fine-tune down the error rate. Labenz has a parallel story at Waymark, his video company for small businesses: years of an in-browser renderer forced to hit television specs for Spectrum Reach, 30 frames a second, HD, no shortcuts. Pros still wanted pro tools. Non-pros said the product was easy and then did not use it, because they had no ideas or got stuck writing a script. Browser accessibility did not automatically become end-user value. Layering on AI changed the trajectory. Simons says that is exactly what happened to them.

Rewind six months and the technology was awesome, developers loved it, and nobody would leave their local tools. There was no meaningful revenue business. They were a venture-backed company seven years in, and they were looking at how to wind it down. Then the same AI overlay that hit Waymark hit Bolt. Simons thinks there will be more of these stories. Most professional tools still live on the desktop. Video, graphics until Figma, development, editing, games: a list that has not come to the browser, because it is hard to pull professionals out of the rooms they already inhabit. Figma and Google's G Suite are the rare breakthroughs. AI lowering the barrier is what makes browser-native versions of those tools suddenly matter.

The founding insight at StackBlitz was environmental, not linguistic. Week one and two of learning to code are not coding. They are installing things and poking the machine until a dev server starts. Join Meta or Netflix and the first month is still onboarding the repo onto your laptop. Force a download before anyone gets to use AI and drop-off is insane, on the order of 90 percent or more. Native in the browser, that layer of friction disappears, and then the model can pick up. There is an order of magnitude of complexity in local setup, and another order of magnitude in knowing how to code. Collapse the stack and you open a market that did not exist, while professionals still get a scaffold they can pull down locally. Sharing stops being a nightmare of "it's on my machine" and a staging deploy you dread.

On camera he opens bolt.new, which he loves as a domain, and types an unbounded prompt: make a music app that looks like Spotify. The left pane looks like ChatGPT until the right pane is a Unix-compliant operating system, already booted, dependencies already installed, the model writing into it. They had to write that OS from the ground up. Convert-a-Docker-image-to-WASM tools exist; the files land at 100MB or more and run about ten times slower. A web page is expected to load in a second, so the system had to be around two megabytes, maybe one, and boot in 100 milliseconds. In 2012 he and his co-founder bumped into Dylan Field and Evan Wallace, whose first Figma pitch was not a design tool. It was a 3D ball dropping into water, ripples in real time, because WebGL had just landed. They spent the first year or two writing a renderer from scratch. WebContainer is the same script. If Gmail breaks you hit refresh. If this OS breaks, refresh boots a new machine, reinstalls dependencies, and brings the server back in seconds. A cloud VM that dies sends you to support and hope.

Zero-shot, the Spotify clone looks finished enough that he keeps prompting: make the play button work, make the seek bar work. The agent reaches for the HTML5 audio API, throws because there is no file, and still goes further than he expected. People have wired similar stacks to music APIs. One user built a meditation app on Bolt in about ten minutes, calling Suno for Zen audio, and Simons looked up Calm's market cap and pinned the result on his phone. A PM at a banking software company in Thailand shipped Viral Hooks, which reverse-engineers creator openings and drafts hooks. A week before Bolt she had an Upwork quote from a Ukrainian developer around $3,000 to $4,000 and three months. She bought the $50 plan, built it after work, and launched in two weeks: about a 99 percent cost cut. A man named Paul built a full CRM with an agent you can tell, "I just met this person, log it." An agency wanted $30,000 and six months. He spent two or three hundred dollars on Bolt and finished in a month. Both products were already live and charging. Neither builder was a coder.

GitHub Copilot and ChatGPT trained people on $20 all-you-can-eat, then a throttle. Bolt let people pay for as much inference as they need, plans from $20 to $200, extra tokens à la carte, plus team plans still metered on monthly tokens. Users try the free tier, love it, burn it, climb. After a $7 cost against a $9,000 delivery, the ROI does the selling. Labenz notices the per-token price can undercut Claude Sonnet 3.5 and asks which model is hiding. Simons is open: they use Sonnet 3.5 a lot, plus others in spots. People complain it is expensive and threaten to go straight to Claude. He tells them to go ahead. It will cost more. As one of the upstream vendors' larger customers, they negotiate volume because a subscription lets them forecast reserved inference. That is how they price aggressively.

At launch they also shipped bolt.diy, an official open-source clone. Point it at Gemini, OpenAI, or, as of that week, DeepSeek, paste your own keys, run locally or behind a VPN. It still executes in WebContainer in the browser; DIY is just a way to host the app and aim it at any provider, including an offline DeepSeek box. Labs use it as a qualitative bench the SWE evals cannot see: does the zero-shot design look good, does the agent feel like it can build a real app rather than an unstyled todo list. People consolidate around software they can fork. Many try DIY and still buy a Bolt subscription. Cursor, by contrast, is closed source. Simons pulls up the GitHub repo, last commit nine minutes ago, and opens the full system prompt: months of work, forkable, with chained prompts behind it. Node toolchains have had five years in WASM. Python has a REPL and basic stdlib, not pip yet. The entire WordPress and PHP world, Laravel included, already runs. He expects most major languages in WebAssembly by the end of the decade, and jokes they will ask the superintelligence to finish the port.

The happy path is opinionated on purpose, especially for people who are not developers: Vite and React, Supabase for auth, database, and webhooks, Stripe for billing, Netlify for deploy. You do not even sign into Netlify. Click deploy and a production build runs on your own CPU, then goes live on a URL you can hang a custom domain on. Change the background to neon, click deploy again, refresh, and production updates. Labenz has lived deployment as a nightmare. He gives Replit credit for baking hosting into the product, and likes that Bolt partners with a company whose actual business is reliable hosting, real domains, and real teams. If something in the in-browser OS goes wrong, refresh is a fresh slate. That, more than the model name, is the part of the experience he calls magic.

Labenz's own loop is already a 5x, sometimes 10x: take a feature to a reasoning model for a plan, hand the plan to a coding agent, mostly Cursor, then spend the expensive hours testing edge cases. He wants to zoom out one more level, let the thing write, test, bang on the happy path five times, and come back. He paid $200 a month for Operator and still felt the missing piece was an AI user. He lined up the same app on Bolt, Replit, Devin, and Lovable. Devin was the alien. It kept going. He did not want to watch it, and when he tabbed back he did not know where he was. A good intern would have walked the fairway, sanded nothing maybe, and returned a summary: here is what I did, how I tested, what is still open. Nobody has nailed that interaction layer. He would wait longer and spend more tokens for bigger chunks if the update felt human.

Simons has been turning this over with Shawn Wang, known as swyx, an angel in the company. The problem with products like Devin is not a slight. Tell an agent to go do something that takes a while, and it does a lot of things in that while, spends a lot of inference, and the odds of collapse rise with every extra step. Wrong does not have to mean an exception. It can mean a misread at the root that cascades. His understanding is that longer-running agents have a pretty low thumbs-up rate. Bolt's ARR keeps ramping because people see a result immediately, and a mistake did not burn twenty minutes from a bad premise. Humans, for the foreseeable future, are still an important part of the creation process. Unit tests written up front help, until you ask who checks that those tests represent what a non-technical user actually wanted.

He still wants the intern who comes back with a status. He just does not think the fidelity is there yet. A model or two from now, some workloads will be trustworthy enough to run longer. Sonnet 3.5 is why the pricing model works at all: more money in, realistic value out. They tried to build Bolt in February 2024 on then-frontier models and could not. The code was unreliable. The design was ugly when it worked. Internally they stay laser-focused on making today's agent excellent, with R&D irons in the longer fire. The other constraint is the tsunami. Every six to twelve months a new model lands. Replit, v0, and Lovable were already in the category; when Sonnet 3.5 arrived, agent work for the previous models stopped mattering, and some of them had to reinvent or pivot the offering. The job is to ride a wave into the shore, paddle out, and catch the next one without getting wiped out.

A week before the taping they announced the Series B: $105.5 million. Tens of millions in revenue, more than two million registered users, something like 50,000 to 60,000 active customers, and a team of twenty. They are, he says, severely understaffed for the demand. Careers are at stackblitz.com/careers, or the hiring link at the top of bolt.new. A year ago, reliable production code from a model was not on the table. That bar is cleared. The contest now is who is still standing when the next wave hits.