An AI wrote 90% of its own code, and then its valuation increased by 25 times.

With Claude Code and Codex gaining popularity, and Cursor being acquired by Elon Musk's SpaceX, the era of the independent Vibe Coding platform seems to be over.

But is this really the case? At least Devin doesn't think so.

On May 27th local time, Devin's team, Cognition AI, announced the completion of a financing round exceeding $1 billion, with a valuation set at $26 billion — just eight months ago, the company's valuation was only $1.02 billion.

This is not a small step forward; it's a vertical takeoff.

Behind the 13-fold Growth

Let's start with the numbers.

Cognition's Annual Recurring Revenue (ARR) was $37 million in May 2025. By May 2026, it had reached $492 million, a 13-fold increase in one year.

The number of enterprise customers has increased by more than 10 times since the beginning of the year. The customer list now includes Citi, Goldman Sachs, Mercedes-Benz, as well as the US Army and Navy.

Putting these names together means that this is no longer just a "developer's toy." Financial institutions, automotive giants, and military organizations are starting to integrate AI agents into real production environments — a leap whose significance many have underestimated.

But the detail that really gives this story a "science fiction" feel is something Cognition itself admits:

Over 90% of their own codebase was written by Devin itself.

An AI programming tool writes its own code and then raises $1 billion in financing based on that code.

This cycle is a bit dizzying, but it's logically consistent.

What Did Cognition Do Right After the Windsurf Acquisition?

To understand today's valuation, we must go back to the acquisition in July 2025.

Cognition acquired Windsurf — a very delicate timing. Around the time of the acquisition, Google was also quietly poaching Windsurf's executives and researchers. The fact that both companies were competing for the same group of people shows that the market's judgment on this direction is highly consistent.

After Cognition acquired Windsurf, the company's ARR increased by more than 30% month-on-month within seven weeks. This speed is an outlier in the SaaS field.

Windsurf brought not only users but, more importantly, it broke down the wall between "developers' daily tools" and "enterprise-level AI agents". Cognition's Devin is positioned to "complete an entire task autonomously" rather than "assist humans in writing a piece of code" — this difference in product philosophy determines the scenarios it can enter.

The case of Mercedes-Benz is the most illustrative one so far. They had a project to modernize a legacy system. The original estimated project duration was eight months. After using Devin, the project was completed in eight days.

The comparison between eight months and eight days — even with some discount — is enough to make any enterprise IT director excited.

The Survival Space for Independent AI Programming Tools

This financing has another meaning, less obvious but more worth pondering.

In the past year, all the signs have pointed in the same direction: large model companies are getting involved in programming tools themselves. Anthropic launched Claude Code, OpenAI restarted Codex, and Google has its own programming agent, Jules. The three top AI labs have all listed "helping programmers write code" as a core battlefield.

Under this pressure, can an independent AI programming startup still raise $1 billion and support a valuation of $26 billion?

TechCrunch's judgment is that this is a "vote of confidence by top venture capitalists in the development space of independent AI programming tools" — especially when all signs indicate that model manufacturers are about to dominate this market.

Scott Wu, the CEO of Cognition, partly explains why they can remain independent. His core logic is: The future of AI programming does not depend on a single large model but on an "orchestration layer" that can intelligently schedule multiple models and tools.

To put it simply, Devin is not just a shell that calls GPT or Claude. It is an "agent system" that can understand tasks, break them down into steps, call appropriate tools, and handle exceptions. The underlying model can be replaced.

This logic is somewhat similar to the rise path of Cursor.

Cursor also packaged the capabilities of large models into the developers' familiar workflow and reached a valuation of $6 billion — it was finally acquired by SpaceX. This news objectively added fuel to Cognition's current financing round: investors saw that even SpaceX was investing in AI development tools, so their confidence in the entire track naturally increased.

The lead investors in this round, Lux Capital and General Catalyst, along with Peter Thiel's Founders Fund, believe that an independent AI agent platform has the potential to become the infrastructure for enterprise software in the next decade, rather than just a plug-in.

The Cost of "Autonomy"

Aravind Srinivas, the CEO of Perplexity, said that Devin is the first AI agent he has seen that "seems to have truly crossed the human-level threshold." The CEO of Ramp also said that Devin's demonstration was the "most impressive single demonstration" he has seen in the past decade.

But crossing the threshold also means new risks.

Concerns in the security field revolve around the same question: What are the consequences if an AI agent that can autonomously execute commands, modify the file system, and advance tasks without continuous human supervision makes a mistake?

It may inadvertently spread malicious code, introduce serious defects into real-time software, or do a series of "unchecked" things while executing "seemingly correct" instructions.

This is not a theoretical concern. How many lines of code in the Mercedes-Benz project completed in eight days were manually reviewed line by line? When such tools start to be used in military systems and bank core business systems, the weight of this question is completely different.

Another concern among industry observers is the sustainability of the valuation.

An ARR of $492 million corresponding to a valuation of $26 billion means that the market has given a revenue multiple of over 50 times. This number is not low — if the adoption rate of the enterprise slows down due to security concerns or budget constraints, the valuation will face great pressure.

A high valuation requires high growth rates to support it, and high growth rates require continuous investment from enterprise customers — whether this flywheel can keep spinning is the most core question for Cognition in the future.

It has been less than two years since Devin was launched and was regarded by many as a "demonstration gimmick" to today, with a valuation of $26 billion and an annual revenue approaching $500 million.

AI wrote 90% of its own code, then helped banks write code, and then raised $1 billion in financing.

This would have sounded like science fiction five years ago. But it's happening today, and the pace is accelerating.

The next question is not "Can AI programming tools succeed?" but "When all software starts to be rewritten by AI agents, what exactly are the software engineers in this world doing?"

*Source of the header image: Medium

This article is from the WeChat official account "GeekPark" (ID: geekpark), written by Hua Lin Wu Wang. It is published by 36Kr with permission.