$600M in annualized revenue on code nobody's engineering team wrote

Before vibe coding had a name, the assumption was simple and rarely questioned: software gets built by people who can write software. You either learned to code, hired someone who did, or you waited. Product managers wrote specs and handed them to engineers. Founders with ideas but no technical background spent years either learning Python on weekends or burning equity on developer salaries. The gap between having an idea for a digital product and shipping one was, by default, measured in months and thousands of dollars, at minimum.

That was not a law of nature. It was a consequence of the tools available.

What is vibe coding and who coined it?

The phrase "vibe coding" was coined by Andrej Karpathy in early 2025. Karpathy, a founding member of OpenAI and former head of AI at Tesla, posted a description of a new way he was working: describing what he wanted in plain language, letting an LLMLLMA Large Language Model is an AI system trained on vast text data to predict and generate language, enabling tasks like writing, summarizing, and answering questions.View full definition → generate the code, and largely ignoring the underlying implementation. He called it "fully giving in to the vibes." The post spread quickly because it named something developers had already started doing informally, and because it arrived at exactly the moment when code-generation models had become good enough for the workflow to actually hold together.

The timing was not coincidental. Models like GPT-4 and Claude 3 had crossed a threshold where they could generate coherent, functional code across a wide range of tasks, not just autocomplete snippets but whole components, full routes, connected logic. That capability had been accumulating for a few years, but it took a recognizable name and a credible voice to crystallize it into a practice people could talk about and build products around.

Lovable, the Swedish startup formerly known as GPT Engineer, was already moving in this direction before Karpathy's post. The product lets users describe an application in natural language and receive a working web app. No IDE, no local environment, no version control to configure. By mid-2026, Lovable's co-founder Fabian Hedin reported that the platform had crossed $600M in annualized revenue, and that apps created on it were receiving nearly a billion monthly views. Those are not prototype numbers. That is production traffic on software that, in most cases, was built without a single line of hand-written code.

Where the real friction in software development lived

What makes the Lovable trajectory interesting is what it reveals about where the friction actually lived in software development. The assumption had always been that the hard part was writing code. Vibe coding platforms suggest the hard part was everything else: knowing what to build, describing it clearly, iterating on feedback, getting something in front of users fast enough to learn from it.

Those are skills that exist in abundance outside engineering teams. Product managers, consultants, marketers, operations leads, and founders without technical backgrounds have spent careers developing exactly this kind of thinking. Vibe coding does not give them a shortcut around software; it gives them access to a process they were already equipped to drive.

The category has also matured faster than skeptics expected. Early vibe coding outputs were easy to dismiss: toy apps, visual demos that broke under real conditions, UIs that looked fine and did nothing. That critique still applies to some portion of what gets built today. But the billion monthly views on Lovable's platform suggest a non-trivial share of what ships is useful enough that real users keep returning to it.

Understandinghow agentic coding systems actually operate behind the scenes helps explain why quality has improved. The better platforms now run multi-step agent loops: generating code, testing it, catching errors, and revising, without user intervention at each step. That is architecturally different from a single prompt returning a static code block.

The question of output quality has not gone away, though. A billion monthly views says something about volume and reachreachThe number of unique people exposed to your message in a given period. Unlike impressions, reach counts each person once, no matter how often they see it.View full definition →; it says less about security posture, data handling, or what happens when an app built in forty minutes starts processing real customer information.Knowing how to evaluate whether what an AI system produced actually works is a distinct skill from knowing how to prompt it, and it is one that vibe coding platforms cannot substitute for.

Why COBOL, 4GLs and low-code never closed the gap

The origin story here is not really about Lovable, or even about Karpathy's 2025 post. It goes back further, to decades of failed attempts to make software creation accessible to non-programmers. COBOL was partly motivated by the idea that business users should be able to write their own programs. Fourth-generation languages in the 1980s made similar promises. Low-code tools in the 2010s got closer but never escaped the gravitational pull of complexity: sooner or later, you still needed a developer.

What changed this time is not the ambition but the mechanism. Large language models can hold enough context, generate enough coherent structure, and iterate fast enough that the gap between description and working software has collapsed in a way earlier tools never managed.

For a professional who is not an engineer, that shift has a practical implication worth sitting with: the bottleneck in building digital products has moved. It used to be technical skill. It is now clarity of thinking and quality of judgment about what good output looks like. Those are learnable, and they were already your job.

Lovable's $600M run rate was built on people who had ideas, described them well, and shipped. The engineering credential was not the constraint. It never was.