Summary

Bolt.new reached $20 million in ARR within eight weeks after pairing frontier coding models with infrastructure that had been developed over roughly seven years. Eric Simons says ARR kept rising, churn was around 25% versus 60-70% for some consumer AI products, and the company had more than 2 million registered users and roughly 50,000-60,000 active customers. His practical signal is that “we’ve crossed this chasm where it’s crazy valuable.”
The technical moat is WebContainers, an operating system that runs inside the browser, boots in roughly 100 milliseconds, and avoids per-user cloud VMs. Bolt uses the customer’s CPU, memory, and GPU, avoiding cloud latency and VM provisioning; Simons notes that scaling to a billion users with one VM per user would require a billion VMs, while a billion user devices already exist. He calls the roughly five-year WebContainers project “the crown jewel of our company.”
AI is expanding software creation beyond developers, with roughly 60-70% of Bolt users coming from groups such as product managers, designers, and entrepreneurs. Developers report up to 10x leverage, while PMs can replace Jira-to-engineer handoffs with direct prototypes or production changes: they are “saying basically the same words and getting an instant result.” One freelancer reportedly spent $7 on Bolt inference and billed the resulting project at $9,000.
The sharpest customer examples suggest that software-production costs and delivery times can collapse by one or two orders of magnitude. A nontechnical Thai PM built Viral Hooks in two weeks on a $50 plan after receiving an approximately $3,000-$4,000, three-month contractor quote; another nontechnical creator built a full CRM for $200-$300 in about a month after an agency quoted $30,000 and six months. Both products launched and were making money.
Browser delivery and generative AI are multiplicative rather than merely additive. Simons says forcing users to download and configure professional software can create “90% or something” drop-off; local environment setup adds one order of magnitude of complexity, and learning to operate the tool adds another. Six months before Bolt, StackBlitz was considering winding down after failing to pull developers away from local tools.
Bolt’s open-source strategy turns model competition and developer enthusiasm into product R&D while keeping its long-built core technology as the stated moat. Bolt.diy exposes its system prompt, supports user-supplied providers, can run DeepSeek locally and offline, and is being adopted behind enterprise VPNs and by AI labs testing models on real applications. Paid plans run from $20 to $200 plus additional tokens; volume discounts on Sonnet 3.5 and other models let Bolt price aggressively.
Long-running autonomous coding agents could unlock another step-change, but Simons believes dependable value today still comes from a human steering short loops. His concern is that every additional autonomous step creates another chance for errors or misunderstood requirements to cascade; Nathan Labenz wants an agent that implements, tests the happy path, and returns with a concise status report. Simons thinks trustworthy longer-running work may be only “a model or two releases away,” while describing model releases as tsunami waves that products must catch without being wiped out.

Deep dive

1. Software is becoming AI’s first major tectonic market

Rather than anchor his worldview on an AGI date or p(doom), Eric Simons takes a pragmatic view of software, where code is text-based, abundant for training, and deterministic enough that a model can write a program, execute it, judge whether it worked, and turn the result into another test case.
That feedback structure helps explain why Simons expects coding to keep advancing faster than domains such as law, where outcomes depend on historical context and “how this judge felt at the time.” His call, stated with a hedge, is that “the world of software is getting rewritten.”
Nathan Labenz adds the o1-to-o3 transition and the R1 moment as evidence. The reported o3 competitive-programming result—placing the model among roughly the top 200 coders globally and ahead of nearly everyone at OpenAI—looks less like generalized hype than a concrete capability jump.
Simons sees reasoning as especially potent when multiplied through software development: “Applying reasoning to software development—boom, that’s huge.” He expects capable engineers to gain 5-10x leverage while new creators enter a market where “the world needs more software; we’ve just been constrained.”

2. Product judgment remains scarce even when code becomes abundant

Nathan’s “luxury software” or “selfish software” thesis asks whether people will build for themselves and small teams rather than mass markets. His reservation is practical: personalized applications remain unattractive if they are buggy, poorly supported, or visibly less polished than established products.
Simons agrees that generated code does not eliminate the long tail of UX and support. Today’s strongest nontechnical users are product managers, designers, and entrepreneurs who already know how to “sculpt” an experience; previously, they could write code only “through someone else’s fingertips.”
For a general end user with an underspecified idea, reliable creation will require more inference and more capable models and agents under the hood. Simons calls the current tooling “very rudimentary,” especially when something breaks and the user lacks the technical knowledge to diagnose it.

3. Product managers are the first major nondeveloper unlock

Bolt’s user mix is roughly 30-40% traditional developers and 60-70% nondevelopers, including PMs and entrepreneurs. Developers use it as a 10x prototyping and UI accelerator: drop in Figma material, iterate for a few prompts, then export to GitHub or a local environment.
Simons’s PM workflow removes an organizational loop. Instead of writing a Jira ticket, assigning it, waiting for implementation, reviewing the result, and sending feedback, a PM can instruct the agent directly and commit the finished change; the downstream developer need not care whether “a coworker or an AI” wrote it.
The division of labor matters: engineers retain difficult functionality that models cannot reliably produce zero-shot, while PMs handle alignment, colors, navigation, and UX polish themselves. Fortune 500 teams are already using Bolt for prototypes in place of Figma, according to Simons.
Simons also describes an arbitrage example: a freelancer spent about $7 on inference to build a web app and billed the client $9,000. Developers and freelancers are searching for “Bolt arbitrage” because the demand-supply curve has not yet normalized.

4. Nontechnical adoption requires education and rapid human rescue

Simons rejects the idea that first-time builders should immediately know how to specify a great product: “Everyone has to learn.” He and his co-founder learned programming at 13 only because they wanted to make products; had Bolt existed, they might not have needed to become software engineers.
His strongest accessibility test was his 73-year-old mother, whom he describes affectionately as “probably the least technical person you’ll ever meet.” Without guidance, she typed what she wanted, pressed Enter, clicked Deploy, and published her first website.
Personal sites reinforce the same point. One man created a website related to medical donations for his daughter, while Simons remembered abandoning a hand-built wedding site and finding drag-and-drop builders overly complicated. His conclusion is blunt: “There’s no reason for drag-and-drop builders when you have this sort of stuff.”
Bolt is attacking the remaining gap through community education and on-demand human help. Inspired by Midjourney users sharing prompts, it cultivates experienced builders; Simons said the company was rolling out a program in which a stuck user could raise a hand and be connected with a certified expert for quick debugging, although the program had not yet been announced at the time of the conversation.

5. Revenue and retention serve as Bolt’s practical success-rate proxy

Asked how often users reach the application they intended to build, Simons answers with behavior rather than a task benchmark: Bolt went “from zero to 20 million of ARR” in its first eight weeks, continued growing afterward, and sees customers returning rather than merely sampling the novelty.
He puts churn near 25%, compared with 60-70% for some direct-to-consumer AI products and an “acceptable” range he describes as roughly 20-50%. That does not prove every project succeeds, but it indicates sustained value as more people use the product.
In Simons’s market snapshot excluding Microsoft and GitHub, Cursor ranked first at approximately $100 million, Bolt second, and a larger group clustered around $5 million-$15 million in ARR. His framing is that “free markets are a good way” to measure value, while acknowledging that more breakout competitors will appear.

6. WebContainers remove the cloud-VM bottleneck

A full-stack agent needs somewhere to run the application it is editing. The conventional cloud-IDE answer gives every user a VM, but every keystroke and result must traverse the network, broken environments require support, and somebody must absorb the compute bill.
Scaling that design compounds both cost and scarcity: “If you want to scale to a billion people, that’s a billion VMs,” and Simons notes that a billion rentable VMs do not exist. A billion user devices do, which is why Google Docs and Figma render through local resources exposed by the browser.
StackBlitz therefore wrote a browser-native operating system that uses the device’s CPU, memory, and GPU. It boots in about 100 milliseconds, has zero latency according to Simons, requires no VM spin-up, and gives the AI an immediately available environment for installing dependencies, writing files, and displaying results.
Simons calls WebContainers “the crown jewel” built over roughly five years inside a company that had spent seven years on browser development infrastructure. Matching the environment’s speed, fidelity, reliability, and scale is the deeper engineering challenge.

7. AI rescued infrastructure that professional developers would not adopt alone

Six months before Bolt’s launch, StackBlitz had admired technology and devoted developer users but no compelling venture-scale business. Professionals would not leave their local tools, and after seven years the founders were discussing whether they might need to wind the company down.
Nathan recognizes the pattern from his browser-based video company: years spent achieving TV-quality, 30-frames-per-second HD rendering did not make nonprofessionals create more videos. Users still lacked ideas or became frustrated writing them; adding AI shifted the product from “DIY to done for you by AI.”
Simons sees a broader opportunity among browser-native professional tools. Local setup alone can drive “90% or something” drop-off; learning to configure a development environment adds one order of magnitude of complexity, and learning to code or operate a professional application adds another.
AI addresses the creation barrier while the browser addresses setup. Even professionals benefit because they can generate a scaffold instantly, share it through a URL, and pull it local only when needed—avoiding the familiar dead end of “it’s done, but where can I use it?”

8. A Spotify clone demonstrates the infrastructure more than the prompt

Simons enters only “make me a music app that looks like Spotify.” Bolt returns a polished, zero-shot interface with relevant licensed or royalty-free placeholders while its browser OS boots, installs dependencies, and streams the generated code into the live application.
Underneath is a Unix-compliant environment implemented in WebAssembly. Adapting a Docker image could produce a WebAssembly artifact of 100 megabytes or more running about 10 times slower; Simons says the operating system instead had to be kept to roughly 2 megabytes, perhaps 1 megabyte, to satisfy web users’ expectation of a near-instant page load.
Simons compares that work with Figma’s 2012 ball-dropping-into-water demo. WebGL and WebAssembly or its predecessor made browser design newly possible, but Figma still spent its first years writing a rendering engine from scratch; StackBlitz likewise had to build directly against constrained browser APIs.
A second prompt attempts to activate the play button and seek bar through the HTML5 Audio API. The agent reports that no audio file exists rather than pretending playback works; Simons says a Spotify API, OAuth flow, or another music API could turn the mock-up into a functional streaming application.

9. Real applications compress months and five-figure budgets

One creator connected Bolt to Suno for a meditation app that generated calming music and guided breathing. Simons said the app was built in about 10 minutes and that he pinned it to his phone, calling the result “insane.”
A PM at a Thai software-banking company first listed Viral Hooks on Upwork. A developer quoted roughly $3,000-$4,000 and three months; after Bolt launched, she used its $50 plan and, despite a full-time job, shipped in two weeks.
Viral Hooks reverse-engineers techniques used by popular creators and generates opening lines for a user’s script. Simons characterizes the comparison as roughly a 99% cost reduction and 10x faster delivery, while preserving the crucial point that this was a functioning startup rather than a disposable landing-page demo.
Another noncoder, Paul U., built Chill CRM with an embedded agent that logs activities conversationally. He spent, by Simons’s estimate, $200-$300 and had it done in roughly a month after receiving a $30,000, six-month agency quote; Simons says both this product and Viral Hooks launched and were making money.

10. Usage pricing captures value that flat subscriptions suppress

Bolt rejected the familiar $20 “all you can eat” AI plan that quietly throttles heavy users. Subscriptions begin at $20, extend to $200, and permit additional token purchases; users typically exhaust the free tier, upgrade, experience more value, and keep moving upward.
Nathan observes that Bolt’s effective token price appears lower than direct Claude 3.5 Sonnet pricing. Simons confirms heavy Sonnet 3.5 use alongside other models and says Bolt’s position as roughly a top-three-to-five customer of upstream providers gives it volume discounts unavailable to individuals.
Subscriptions also make reserved demand predictable enough to negotiate forward inference commitments. Simons acknowledges that inexperienced builders may perceive the product as expensive, but contrasts that with customers saving 99% against an agency—or turning $7 of inference into a $9,000 deliverable.

11. Bolt.diy makes open source a model-discovery engine

Bolt.diy runs the open-source Bolt interface locally or on a private server while WebContainers still execute code inside the browser. Users can supply Gemini or OpenAI keys, connect other providers, or run DeepSeek locally for an offline experience; enterprises are beginning to deploy it behind VPNs.
That flexibility makes the repository a testing ground for production Bolt. Promising models can emerge through community use before being pulled into the commercial product, while AI labs use Bolt.diy to assess whether a model can build a real, attractive application—not merely pass a static software-engineering eval.
The repository includes the main system prompt and other chained prompts, representing months of work by Bolt’s team. Simons calls it “the whole shebang,” although he notes that some improvements on Bolt.new have not been backported and some may remain production-only.
Open sourcing the agent does not erase the moat because WebContainers took years to build. Developers can fork, use, and modify the tool, and some people try Bolt.diy before buying Bolt subscriptions; Simons describes the relationship as symbiotic, with open source becoming “a tide that lifts all boats.”

12. An opinionated stack creates the nondeveloper happy path

Node.js and JavaScript remain the mature WebContainers ecosystem, but basic Python already runs in WebAssembly with some standard libraries. Third-party package installation was not yet available and pip remained on the roadmap; PHP, WordPress, and Laravel had also reached meaningful browser compatibility.
The original StackBlitz thesis was that “the web should be able to build the web.” Just as macOS has Xcode and Windows has Visual Studio, browsers should have their own development environment. Nathan speculates that most major languages and ecosystems could be running in WebAssembly by the end of the decade.
For users who do not want to choose infrastructure, Bolt favors Vite and React for the application, Supabase for authentication, databases, webhooks, and Postgres backing, and Stripe for billing. Prescriptive defaults reduce the integration errors that strand less technical builders.
Netlify completes the workflow with one-click deployment that does not initially require a separate login. The production build runs on the user’s CPU, publishes to a live URL, and can later attach a custom domain, Git repository, or Netlify account; another prompt and deploy updates that same production site.

13. Human-guided agents win now, but the autonomy frontier is close

Nathan describes his desired outer loop: use o1 Pro to inspect the codebase and plan a feature, let Cursor or another coding agent implement it, then have an AI user or tester exercise the happy path and edge cases. Removing his remaining copy-paste and testing labor could turn occasional 5-10x gains into consistent 10x and perhaps 20x gains.
Simons’s pushback—worth keeping—is that long-running agents such as Devin make more decisions, spend more inference, and create more opportunities for one misunderstanding to cascade. These systems “just aren’t at the level of fidelity” required to “send them off into the woods,” so short loops with human verification currently deliver more reliable value.
Nathan does not want to supervise every step; he wants the equivalent of a good junior colleague who tests before reporting back, summarizes what changed, confirms what worked, and reports what questions remain. Simons agrees that experience is coming, but notes that even writing unit tests is hard when a nontechnical user cannot specify what the tests should guarantee.
Sonnet 3.5 was the threshold that made spending more inference plausibly return more value; a February 2024 Bolt prototype produced unreliable code and ugly designs. Simons now thinks trusted longer-running workloads may be “a model or two releases away,” while keeping Bolt focused on today’s agent and treating model launches as “tsunami waves” to catch. After a stated $105.5 million Series B, more than 2 million registered users, and tens of millions of dollars in revenue, its 20-person team was “severely understaffed.”

Software Supernova: Lovable's "Superhuman Full Stack Engineer" to Transform Idea to App in Seconds 23. The Prohibition 2.0 Era: DeepSeek—The End of Compute Supremacy, or Just Another Beginning?