On August 12, a two-year-old company in Stockholm with around 300 staff announced that it had raised $400 million at a valuation of $13.3 billion. Lovable does not make chips, or models, or data centres.
It makes software that writes software, on the instruction of people who cannot code. Menlo Ventures led the round. So did the Scaleup Europe Fund, a European Union vehicle managed by EQT that exists to stop the continent’s best companies from moving to California. Tencent came in from Asia. Kaszek came in from Latin America.
Eight months earlier the same company had been worth $6.6 billion. Two years earlier it did not exist.
That is the shape of vibe coding in 2026. A phrase coined almost casually in February 2025 has become one of the fastest-capitalising categories in the history of enterprise software, and the argument about whether it deserves the money has become just as loud as the money itself.
What vibe coding actually is
The term belongs to Andrej Karpathy, the former Tesla artificial intelligence (AI) lead and OpenAI co-founder. He used it to describe a particular way of working that he found both productive and slightly disreputable. You tell a model what you want in ordinary English, you accept what it produces, you do not read the code, and you keep talking to it until the thing works. The vibes, rather than the syntax, are doing the driving.
Karpathy meant it narrowly, for weekend projects that nobody would ever have to maintain. The industry took it much further. Within a year the phrase had been stretched to cover an entire product category, from consumer app builders aimed at marketing managers to agentic development platforms sold to banks.
It is worth holding on to the distinction that the marketing collapses. AI-assisted coding, where a professional developer uses a model for suggestions, autocompletion or refactoring while still reading everything, is now close to universal.
Survey work from Stack Overflow and JetBrains puts adoption somewhere between 84% and 90% of working developers. Pure vibe coding, where entire applications are generated from prompts and shipped with little or no line-by-line review, is a much smaller phenomenon inside professional engineering. Roughly 72% of professional developers say it forms no part of their working day.
The growth, then, is not mainly coming from developers replacing themselves. It is coming from a new population entering software creation for the first time. Around 63% of vibe-coding platform users are not developers at all.
They are operations managers, teachers, small retailers, analysts and founders without technical co-founders, and they are building things that used to require a hired team.
How it functions
A vibe-coding platform interprets. A natural language request is turned into a working specification, usually by a frontier model, and often expanded silently into decisions the user never made about frameworks, database schema, authentication and hosting.
The model also writes the code, typically in whole files rather than fragments, drawing on repository context so that new code fits what already exists. This is where the last two years of progress have concentrated. Repository-aware agents can now hold a large codebase in view rather than answering each question in isolation.
Third it executes and corrects. This is the real dividing line between the 2023 generation of code assistants and the 2026 generation of agents. Modern systems run the code, read the error, rewrite the code and run it again, in loops that can continue for many minutes without a human touching anything.
They open terminals, browse documentation, click through their own user interfaces in a test browser and file the results. Some platforms run several agents in parallel on different parts of the same task.
Fourth it deploys. Replit, Lovable and their peers do not simply hand back a folder of files. They provide the database, the storage, the authentication and the hosting, so that a person who has never configured a server ends up with a live web address.
The commercial architecture matters as much as the technical one. Almost none of these companies own the intelligence at the centre of their product. They rent it, mostly from Anthropic, OpenAI and Google, and increasingly they route between several suppliers to protect margins and hedge against a single vendor. Cursor went further and trained its own model, Composer, to handle a growing share of requests in-house. That tension, between owning the interface and renting the brain, is the strategic question hanging over every valuation in the sector.
The start-ups
The category’s defining event arrived in June. SpaceX, four days after the largest initial public offering (IPO) in history, signed an agreement to acquire Anysphere, the maker of Cursor, in an all-stock transaction valuing it at $60 billion. It is the largest acquisition of a venture-backed start-up ever recorded, and regulatory clearance was reported to be near completion in early August.
Anysphere, founded by four MIT students in 2022, reported $100 million in annualised recurring revenue in January 2025, $1 billion by November 2025 and roughly $4 billion by June 2026, with about two-thirds of it coming from enterprise contracts. Slack took five years to reach its first billion in revenue. Cursor took around three.
Lovable is the European counterweight. Founded in Stockholm in 2024 by Anton Osika and Fabian Hedin, it reached $100 million in annualised revenue within eight months of launch, $500 million by June 2026 and is tracking towards $600 million. Its Series C makes it Menlo Ventures’ largest single position after Anthropic.
Replit, the oldest name in the group, spent a decade as a browser-based coding environment before the agent era rescued it. Revenue went from $2.8 million to $150 million in under a year, its valuation tripled from $3 billion to $9 billion in six months, and it has aimed the product squarely at non-technical staff inside large companies rather than at engineers.
Cognition sits at the other end of the spectrum, selling autonomous software engineering to enterprises through its Devin agent and the Windsurf editor it acquired after Google hired away Windsurf’s leadership in a $2.4 billion licensing arrangement. It more than doubled its valuation in May to $26 billion.
Beneath these are dozens more, including Bolt.new from StackBlitz, Vercel’s v0, Magic Patterns and Tessl, plus the specification-driven tools attempting to make the output auditable. Analysts who add up the private marks across the whole agentic and vibe-coding cohort now put the cumulative figure well north of $150 billion.
The Big Tech response
Every large platform company has now shipped something in this space, and the strategic postures differ sharply.
Google has been the most aggressive. At it’s I/O conference in May it launched “Antigravity 2.0,” a standalone agent-first development platform that deliberately removes the code editor from the centre of the experience and replaces it with a control room for managing multiple agents.
It also added native Android app generation from prompts inside AI Studio and began moving developers off its Gemini command line tool and into the Antigravity ecosystem. The reaction from professional developers has been genuinely split, with a vocal group objecting that Google has replaced a familiar workflow with a dashboard.
Microsoft has taken the opposite approach through GitHub. Rather than betting on one agent, Agent HQ opens the platform to rivals, allowing Copilot subscribers to assign work to Anthropic’s Claude Code and OpenAI’s Codex alongside Microsoft’s own tooling, inside the same repositories and pull requests. With more than 20 million Copilot users and adoption across 90% of the Fortune 100, Microsoft is betting that owning the workflow matters more than owning the agent.
Amazon has gone for rigour. Kiro, its agentic development environment, is built around specification-driven development, forcing a requirements and design stage before any code is generated, with automated hooks that trigger tests and reviews on defined events. It is a direct pitch to regulated enterprises that find unsupervised generation unacceptable.
OpenAI and Anthropic are competing from the model layer downwards. Codex now claims millions of weekly active users, with OpenAI reporting that more than 20% of them are not developers and that this group is growing faster than the developer base. Claude Code became one of the fastest-growing products of the cycle after crossing a billion dollars in annualised revenue within months of launch.
China has built a parallel ecosystem, and the momentum there accelerated after security concerns pushed some domestic firms away from American tools.
ByteDance’s Trae reported more than six million registered developers and the largest domestic market share. Alibaba’s Qoder, spun out of its Tongyi Lingma assistant, claims more than five million users across editor, command line and mobile.
Tencent’s CodeBuddy covers roughly 12,000 of its own engineers, and the company has said more than 90% of its developers use it, with about half of newly added code at Tencent written with AI assistance. These tools run on domestic models such as Doubao, Qwen, Hunyuan and DeepSeek while still allowing users to plug in foreign ones.
Apple’s involvement has been the least willing. App Store submissions reached 235,800 in the first quarter of 2026, an 84% rise on the same quarter a year earlier, and the store added roughly 560,000 new apps in the first half of the year, close to the whole of 2025.
Apple pulled some code-generation apps in March and blocked updates to others, citing a long-standing rule against apps that download and run code which changes their behaviour after review. Developers reported waits of several weeks.
Apple disputes that, saying 90% of submissions are reviewed inside 48 hours and that average review time is about a day and a half. Meanwhile it has been adding agentic coding to Xcode, its own development tool.
The money and the investors
Precise market sizing is unreliable here, because different research houses draw the boundary in different places. Estimates that exclude professional developer assistants and count only prompt-first platforms cluster around $4.7 billion for 2026, growing at about 38% a year.
Wider definitions that fold in coding assistants reach $7.1 billion for the same year, with a ten-year forecast of $34.4 billion. The broader AI coding tools market, on the loosest definition, generated around $12.8 billion in revenue in 2026 against $5.1 billion in 2024. Treat all of these as directional.
The funding numbers are firmer, and they are extraordinary. Average deal size for AI coding start-ups rose from $7.4 million in 2022 to $527.8 million by the end of 2025, a 71-fold increase. Cumulative venture money into AI coding tools passed $5 billion in 2024 alone, faster than any previous developer tooling category reached that mark.
Add the disclosed rounds of the past eighteen months across Anysphere, Lovable, Replit, Cognition and their peers and the direct equity total comfortably exceeds $10 billion, before counting the $60 billion acquisition that now sits on top of it.
That is happening inside an even larger distortion. Crunchbase recorded $510 billion of global venture funding in the first half of 2026, more than the whole of 2025, with more than 70% of second-quarter capital going to AI companies. OpenAI and Anthropic between them accounted for $217 billion, or 43% of every venture dollar deployed worldwide in the half.
Investor behaviour within the coding vertical has shifted in three visible ways. Capital has moved up the stack, from application builders towards infrastructure, orchestration and the data layers that agents depend on, on the theory that interfaces commoditise faster than plumbing.
Strategic money has crowded in alongside financial money, with Nvidia, Google, Salesforce, Databricks and HubSpot taking positions in companies that are also their customers or suppliers. And the underwriting has become explicitly momentum-based. The $60 billion price for Cursor implies roughly fifteen times annualised revenue, which is defensible only if the revenue keeps doubling.
The candid version of the thesis, offered privately by several investors in the category, is that they are not underwriting current cash flow. They are buying a position in the layer that sits between models and the people who build with them, and they are prepared to be wrong about individual companies as long as they are right about the layer.
Where the ecosystem is growing fastest
The United States remains the centre of gravity by capital, holding roughly 88% of AI equity funding and around a quarter of platform activity, and it now hosts the category’s largest exit.
Europe’s position rests unusually heavily on one company. Lovable’s rise makes Sweden the second most significant country by enterprise value in the vertical, and the involvement of an EU-backed fund in its latest round turns a commercial story into a sovereignty story. Britain, Germany and the Netherlands follow on usage rather than on capital.
Asia leads on people. The Asia-Pacific region accounts for roughly 40% of global vibe-coding adoption. India is the single largest source of users at about 16.7% of the global base, ahead of Japan, Pakistan and Indonesia, and Indian developers report among the highest daily intensity of use anywhere.
That has begun to convert into companies. Emergent, headquartered in San Francisco but with 70 of its 75 staff in Bengaluru, tripled its valuation to $300 million in January on $50 million of annualised revenue.
Rocket, founded in Surat rather than in any established technology hub, raised seed money from Salesforce Ventures and Accel and has been in talks for a further round. Zoho, Composio and the cloud provider NxtGen have all entered with domestically hosted alternatives.
China is best understood as a separate market rather than a laggard, with four large platform companies shipping competing products into a domestic developer base that has strong regulatory reasons to stay local.
Brazil ranks surprisingly high on platform activity, and South Korea and Japan are growing steadily on the back of national AI frameworks and enterprise modernisation programmes.
The case against
The security argument is the strongest, because it is the best measured and the least improved. Veracode tests generated code from more than a hundred models against common vulnerability classes and has found a 55% pass rate that has not moved through two years of model releases, meaning roughly 45% of samples ship an exploitable flaw.
A scan by Escape.tech of more than 1,400 vibe-coded production applications found 65% with security issues and 58% with at least one critical vulnerability, including hundreds of exposed secrets and instances of exposed personal data.
Georgia Tech researchers tracked a rising monthly count of vulnerability disclosures attributable to AI-generated code through early 2026 and believe the true number is several times higher. Models are optimised to produce code that runs, not code that validates inputs and handles authentication edge cases, and getting better at the first has not delivered the second.
The maintenance argument is nearly as strong. Analyses of post-adoption codebases report sharp rises in code duplication and sharp falls in refactoring activity, alongside higher defect rates that erode the initial speed gains.
The debt is unusual because the code looks clean, passes linting and compiles, so it does not announce itself until someone has to change it. Forrester expects three-quarters of technology decision-makers to be carrying moderate to severe technical debt by the end of 2026.
The labour argument is real but frequently overstated. Junior software developer employment has now fallen for 33 consecutive months, entry-level postings are down sharply, and Forrester expects computer science enrolments to fall by around a fifth.
But the same period has seen machine learning engineering postings rise, the US Bureau of Labor Statistics still projects 15% growth in software developer employment to 2034, and several large employers are expanding graduate intake on the argument that judgement has to be trained rather than bought. What has changed is the entry price and the job description, not the existence of the profession.
The bubble argument is the hardest to settle. Twenty-plus well-funded platforms are chasing overlapping users with rented intelligence and thin switching costs, and the Bank for International Settlements has warned that this cycle carries echoes of earlier manias in which the technology was transformative and the investors were still wiped out. The counter-argument is that the revenue is real, recurring and enterprise-weighted in a way that 2021 vintage software revenue often was not.
The most telling number is none of these. It is the gap between 84% adoption and 29% trust. Developers are using tools they do not believe, because the speed pressure is too great to refuse. Karpathy himself declared vibe coding passe in February 2026, arguing for a more structured arrangement in which agents handle implementation and humans own architecture and review.
That is probably where this settles. Not unattended magic, and not a fad, but assisted engineering with the review discipline bolted back on. The companies that survive the coming consolidation will be the ones that make the code they generate legible to the people who have to live with it.