Okara
Get Started
Back to Blog
Published by Jordan Reese · September 11, 2026 · 11 min read

How Cursor Grew to $4B ARR

How Cursor used dogfooding, rapid product launches, developer word of mouth, and bottom-up enterprise adoption to build an AI coding business.

Cursor's first Hacker News launch did not look historic.

The March 2023 post received nine points. Developers called the product buggy, questioned why it needed to be a separate editor, and asked why the team had not built a VS Code extension instead.

Two years later, Cursor had passed a reported $500 million annual revenue run rate. By June 2026, that figure had climbed above $4 billion, according to Forbes. The company had spent little on paid marketing during its early rise.

That last detail is often retold as "Cursor grew without marketing." I don't think that is accurate. The founders made launch videos, posted on social media, ran a waitlist, wrote detailed product updates, and built a public forum. Later, they added enterprise sales.

What Cursor skipped was a conventional paid acquisition engine. Developers tried the product, used it at work, and showed it to other developers. Product quality carried most of the distribution burden.

Cursor began with the wrong product

Before Cursor, the four Anysphere founders worked on AI autocomplete for computer-aided design software used by mechanical engineers.

GitHub Copilot for CAD was the rough idea. Several problems surfaced quickly: the founders were not mechanical engineers, good training data was scarce, CAD software was fragmented, and AI models were still weak at reasoning about 3D objects.

Aman Sanger later put the lesson plainly: it is better to be a user of the product you are building. The team discussed the failed CAD attempt in interviews with Latent Space and Lex Fridman.

Writing software was different. The founders coded every day. They could notice a bad suggestion, slow response, or awkward interaction without arranging a customer interview. Every hour of work also became a product test.

That did not guarantee a good product. It made the feedback loop much shorter.

Copilot proved the habit before Cursor existed

The founders were heavy GitHub Copilot users. Copilot had already taught developers that AI autocomplete could belong inside a normal workday.

Then the team saw an early version of GPT-4 in late 2022. The model could do more than finish a line. It could explain code, help debug it, and make larger changes. CEO Michael Truell came away believing that much more of programming would pass through language models.

The existing editor interface was built around typing code by hand. An extension could add chat or autocomplete, but the Cursor team wanted control over how developers reviewed diffs, selected context, moved between files, and handed work to an agent.

So the company decided to build an editor.

They tried building the editor from scratch

Cursor did not begin as a VS Code fork. The first version was a custom editor.

Early users quickly found everything it lacked. Mature editors contain years of unglamorous work: extensions, keyboard shortcuts, language tooling, remote development, debugging, and thousands of small behaviors developers no longer think about.

The team moved Cursor onto VS Code's open-source base. That choice removed a large part of the switching cost. A developer could bring over familiar extensions, themes, shortcuts, and habits while gaining an interface the Cursor team could change more freely than a normal plugin allowed.

This was a useful compromise. Cursor felt familiar on the first day, but the company still controlled the application around the code.

Truell explained the decision in interviews with Lenny Rachitsky and Lex Fridman.

The launch created users, not a breakout

One founder had built a small following by posting technical explanations of AI models. The team used that audience to share a polished Cursor demo and collect names for a waitlist.

The demo was better than the product. Truell later called it "movie magic."

That brought in enough developers to start learning, but the public response remained mixed. In the first Hacker News thread, people complained about bugs and questioned the editor strategy. A second launch thread received two points.

By October 2023, the tone had improved. Developers in another Hacker News discussion praised parts of the experience while still reporting broken extensions and unreliable codebase context.

Those threads are more useful than a clean origin story. Cursor was not born with obvious product-market fit. The company released an unfinished tool into a skeptical technical community and kept fixing what users disliked.

The team spent 2023 in "monk mode"

After the initial launch, the founders stopped chasing large signup numbers. Truell said the team "lived like monks" through much of 2023 and focused on the product.

They tried short growth-engineering pushes along the way. In his telling, the effects washed out compared with improvements to the editor.

Cursor used a stricter internal metric than daily or monthly active users. The team watched for paid power users who worked with the product four or five days each week.

That measure filtered out people who tried an impressive demo and never returned. A professional developer paying for Cursor and opening it most workdays was a much stronger sign.

The October 2023 post "Our problems" shows what the team was working on at the time: better codebase retrieval, larger edits, faster diffs, and agents that could act without becoming unpredictable. Cursor said it had already indexed 150,000 codebases and stored 1.4 billion vectors.

Monk mode sounds romantic in hindsight. It was a small team grinding through context errors, latency, broken extensions, and edits that looked correct until they touched the wrong file.

Each product jump gave developers something to show

Cursor improved in visible steps.

Code-aware chat removed some of the copying between an editor and ChatGPT. Codebase indexing made it possible to ask questions about an unfamiliar repository. Cursor Tab began predicting larger edits and the next place a developer might want to change. Fast Apply made generated edits appear quickly. Composer coordinated changes across files. Agent could choose context and use the terminal.

The company also trained and served custom models for parts of the experience. Truell told Lenny Rachitsky that every "magic moment" in Cursor involved a custom model in some way.

These features were easy to demonstrate to another developer. Someone could show a colleague a multi-file edit, a useful Tab prediction, or an agent fixing a bug. The output did the explaining.

Word of mouth in software teams is practical. A developer sees a coworker finish a task faster, asks what they used, and installs the same tool. If it helps, the expense moves from a personal card to a team budget.

Cursor's launches gave that process a push. They did not create it from nothing.

Truell described growth as a fairly consistent exponential curve that accelerated around some launches. Public evidence cannot tell us how much revenue came from Tab, Composer, Agent, or any single release.

"Vibe coding" arrived after Cursor already had traction

In February 2025, Andrej Karpathy posted about "vibe coding," a style of programming in which a developer lets an AI model generate much of the code and pays less attention to each line.

He named Cursor Composer and Claude Sonnet in the original post. The phrase spread far beyond Cursor and gave the behavior a memorable name.

It was a large piece of free distribution. It was not Cursor's starting gun.

By January 2025, Cursor had already announced more than $100 million in recurring revenue. Karpathy's post amplified a product and behavior that developers were already using.

The free plan made trying Cursor easy

Cursor's early pricing was easy to understand. The free Hobby plan included limited completions and premium requests. New users received a two-week Pro trial. Pro cost $20 a month, while Business cost $40 per user.

Developers could test the full experience before paying. Heavy users hit the free limits and had a clear reason to upgrade.

This model also separated curiosity from habit. Someone opening Cursor once could remain free. Someone using AI every workday would run into the paid boundary.

Some widely repeated figures claim Cursor reached 360,000 paying users and converted 36% of free users. Those numbers come from Sacra estimates, not an audited company disclosure, so they should not be treated as fact.

The forum kept technical feedback in public

Cursor opened a searchable Discourse forum for tutorials, bug reports, feature requests, and support.

That was a good fit for the audience. Developers could find old answers through Google, read how other people used the tool, and watch employees respond to problems. A Discord server would have made much of that discussion harder to find later.

The forum was not always flattering. Users documented regressions, pricing confusion, failed edits, model limits, and features that did not work as advertised. Those complaints also showed that people cared enough about the tool to keep pushing on it.

There is no public number tying the forum to revenue. Its role was simpler: it gave Cursor a public feedback archive and gave potential users a large body of real product discussion.

Individual use opened the enterprise door

Cursor's first enterprise users often arrived from the bottom up.

By August 2024, Forbes reported that engineers at OpenAI, Midjourney, Perplexity, Scale AI, Shopify, and Instacart were using the product. These were individual developers and teams adopting Cursor before a company-wide purchase.

Enterprise sales became important later. Cursor added privacy controls, single sign-on, admin tools, analytics, audit features, model restrictions, and spending controls. Salespeople could then formalize usage that already existed inside a company.

In June 2025, TechCrunch reported that individual subscriptions had supplied most of Cursor's revenue until recently, while enterprise licenses were beginning to grow quickly.

By November 2025, Reuters said sales-led revenue had increased one hundredfold since the beginning of the year. Bloomberg later reported that corporate customers supplied about 60% of annualized revenue in February 2026. Cursor told Forbes that the share had reached 75% by June.

The sales team did not need to introduce an unknown tool. In many accounts, developers had already done the first sale internally.

The growth story gets too neat if you stop there

Cursor's product-led rise had real mistakes.

In June 2025, the company replaced predictable request limits with compute-based credits and used the word "unlimited" in a way that confused customers. Some users received unexpected bills. Cursor apologized, clarified the pricing, and offered refunds in a post titled "Clarifying our pricing".

Reddit filled with angry threads. The pricing megathread captured the main complaint: customers could no longer predict what their subscription bought.

Developers also continued to report reliability problems on complex work. Cursor could produce a convincing edit that still needed careful review. The better the output looked, the easier it was to trust too much.

And "almost no marketing" became less true as the company moved upmarket. Enterprise growth required sales, security material, customer education, launch communication, and proof that large companies could control how the tool handled code.

None of this erases the growth. It explains what the simple version leaves out.

The useful lessons are less glamorous

The failed CAD product may be the most useful part. Cursor's founders moved into a market where they were users, so they could test the product all day without manufacturing feedback.

The VS Code decision mattered for a similar reason. They kept the part that needed to feel familiar and rebuilt the part where they wanted an advantage. Developers did not have to relearn their editor before seeing value.

I would also steal the paid-power-user metric. Signups can reward a good demo. Repeated paid use is harder to fake.

The part I would not copy blindly is "ignore growth and build." Cursor already had a social audience, a waitlist, public launches, and a market educated by GitHub Copilot. Product work happened inside that distribution context.

Cursor grew because developers kept returning and brought the tool to work. The launches, posts, forum, and later sales team helped more people notice behavior that was already happening.

Product stories still needed distribution

Cursor did not need a large paid marketing team early on, but it still had a steady stream of product stories to distribute: changelogs, technical posts, launch demos, founder interviews, and answers to community questions.

Okara can help a small team turn that raw material into search content and social posts without making the founder rewrite the same launch five times. It cannot create the underlying product moment. Cursor had to earn that part.

Frequently asked questions

How did Cursor grow so quickly?

Cursor improved the product through heavy internal use, made switching easy by building on VS Code, offered a free trial, and spread through developer word of mouth. Enterprise sales later converted bottom-up adoption into larger contracts.

Did Cursor really grow without marketing?

Cursor spent little on paid marketing during its early growth, but it still used social media, launch demos, a waitlist, technical content, a public forum, and later an enterprise sales team.

When did Cursor launch?

Cursor's earliest official release record is dated March 14, 2023. It passed a reported $100 million recurring revenue run rate in January 2025.

Is Cursor's reported revenue audited?

Public audited statements are unavailable for the private company's run-rate figures. Later numbers were reported by Cursor or by publications citing people familiar with its finances.

Sources