There’s a sentence from Adam Wathan that struck me, that I’ve been thinking about for two days since he posted it. He wrote it the day after laying off three quarters of his engineering team. It was a reply to a pull request on GitHub, one of those apparently innocuous technical requests that normally pass unnoticed. But Wathan’s reply wasn’t technical. It was a cry of pain disguised as a rational explanation.
For those who don’t know him, Wathan is the creator of Tailwind CSS, one of the most-used frameworks in the world for web development. If you’ve worked on the front-end in recent years, you’ve probably crossed paths with Tailwind. It’s everywhere, with 75 million downloads a month. It’s the dream of open source realised: a free project, loved by the community, that had found a sustainable model to exist. Had. Because now that model is in pieces, and the reason should alarm us and make us ask questions.
Tailwind Labs’ revenue collapsed by 80%. In one year. Not because the project is in decline—on the contrary, it has never been more popular. It collapsed because AI changed how developers use software. And that, when I think about it, frightens me much more than any other consequence of AI we usually talk about.
To understand what happened you have to go back a few years and look at how Tailwind’s business model worked. The framework is open source, free, accessible to all. But the team also built Tailwind UI, a premium component library they sell. It’s one of the classic open source models: offer something free and valuable, build a community, and a percentage of that community ends up buying the paid products. In 2020, Tailwind UI made $500,000 in the first three days after launch. Within five months it had passed two million. In 2024 the team had grown to eight people, with engineer salaries of $250,000–$300,000. It was the living example that sustainable open source is possible.
Wathan’s Cry of Pain
Then AI arrived.
Since 2023, traffic to Tailwind’s documentation has dropped 40%. Not because developers stopped using Tailwind. On the contrary, usage is at all-time highs. But because when you need a Tailwind component, you no longer go to tailwindcss.com. You open Copilot, or Claude, or Cursor, and you say “make me a button with Tailwind”. And the AI generates it for you. Instantly. Without ever touching the official site.
Here’s the point that obsesses me: the documentation was the funnel. It was the moment developers discovered that Tailwind UI, the premium version, also existed. But if developers never visit the documentation, they never discover the paid products. The funnel has been completely short-circuited. AIs have removed the intermediation, that productive friction that allowed open source projects to monetise.
Wathan says it with a clarity that almost hurts:
Tailwind is growing faster than ever and is bigger than ever, and our revenue has dropped almost 80%. At the moment there’s no correlation between making Tailwind easier to use and making the development of the framework sustainable.
I’ve reread that sentence several times. There’s no correlation between improving your product and making it sustainable. It’s the death of the model. It’s the end of an era.
The Funnel Short-Circuited
The pull request that triggered all this was a request to add an llms.txt file to the Tailwind repository. It’s an emerging standard, a way to make documentation more easily readable for Large Language Models. Other projects have already adopted it. It looks innocuous, almost obvious. But Wathan closed it with an explanation that became an involuntary manifesto of the open source crisis.
We have more important problems to face, like raising funds to keep the business afloat. Making our documentation more readable for LLMs will only further reduce visits to the documentation, and fewer users will discover our paid products, further reducing the sustainability of our business.
Some criticised him. They told him the message he was sending is selfish, that he puts money before service to the community. But this is where the situation becomes really tragic, because Wathan is right. Completely right. And there is no winning choice.
No Winning Choice
I tried to think about the options he has, and none of them works.
If you block the LLMs, add aggressive rules in robots.txt, prevent OpenAI and Anthropic and Google crawlers from scraping the documentation, what happens? The LLMs stop knowing your project. Developers using AI assistants no longer get suggestions based on your documentation. Your project becomes invisible in an era when 84% of developers use AI tools. You lose userbase. You lose relevance. You lose everything.
If you don’t block LLMs, you let bots scrape freely. The LLMs train on your content. They get great at generating code that uses your framework. Developers love it. But they never visit your site. They never discover your paid products. Revenue collapses. You have to fire the team. The project becomes unsustainable.
If you add llms.txt to “help” the LLMs, you make the documentation even easier for AIs to digest. The LLMs get even better at answering without sending users to your site. You accelerate your own economic obsolescence. It’s like sharpening the knife for whoever is stabbing you.
There’s no winning choice. It’s literally a deadly vicious cycle. And what makes me shudder is that it isn’t only Tailwind.
It Isn’t Only Tailwind
Stack Overflow, the Mecca of programming questions for twenty years, has seen submissions collapse from peaks of 200,000 a month in 2014 to less than 50,000 at the end of 2025. 47% of daily active users have simply disappeared. Now 81% of developers use ChatGPT (or alternatives) for the questions they once asked on Stack Overflow. Traffic kept dropping even after Stack Overflow blocked GPTBot and then lifted the block for a partnership with OpenAI. Nothing changed: users had already changed behaviour and habits.
Business Insider has recorded traffic drops between 40% and 48%. “Zero-click searches”—queries where the user gets the answer directly on the results page without clicking any site—now make up 62% of all searches. 2025 has been called “the organic traffic crisis”.
Read the Docs, GNOME, SourceHut, LWN, Fedora—all open source projects, all under siege from AI crawlers. GNOME GitLab had significant downtime. They had to implement reverse proxies with proof-of-work challenges to block the most aggressive bots. But it’s a battle lost from the start, because bots become more sophisticated, pay for accounts to look human, falsify TLS fingerprints.
Some are trying to find ways out. Cloudflare launched “Pay Per Crawl” in June 2025. The idea is elegant: instead of blocking or allowing for free, you can charge AI crawlers per request. You use response code 402, “Payment Required”, dormant for decades. Cloudflare acts as intermediary and you set a price per request. On paper it could shift more than two billion dollars from AI companies to publishers by 2027.
But there’s a problem: it works only if a critical mass of publishers adopts it. If you’re alone in setting up the paywall, AI companies ignore you and go elsewhere. It’s a classic coordination problem, and historically the open source ecosystem has never been good at coordinating on this kind of thing.
The Next Chapter: Model Collapse
There’s an aspect of this story I find particularly disturbing, and few are noticing. If developers stop posting questions on Stack Overflow, if they stop writing detailed technical documentation because “the AI will generate it anyway”, what will future AI models train on?
It’s a circularity problem. Current LLMs trained on decades of high-quality human-generated content: carefully written tutorials, questions and answers on forums, detailed technical documentation. But if that source dries up because it’s economically unsustainable, what happens? AIs start training on output generated by other AIs. It’s the “model collapse” phenomenon, and it leads to progressive degradation of quality.
Wathan himself experienced something like this. He tried to use Claude to add dark mode to over 600 Tailwind UI components. He said the results were so inconsistent that reviewing and fixing the agent’s work took longer than doing everything by hand from the start. Even for apparently simple tasks, current LLMs have obvious limits in the absence of quality documentation.
The more I think about it, the more I’m convinced there’s no win-win solution here. Or at least, none that preserves the current model of sustainable open source.
Traditional monetisation models—open core, SaaS hosting, premium add-ons, professional services—were all based on an assumption: that there was a moment of contact with the user. A moment when the user came to know your project, visited your documentation, saw your brand, discovered your commercial products. That moment was the funnel. It was where you could convert free users into paying customers.
AI tools have completely short-circuited that funnel. Developers don’t “visit” anything anymore. They ask Copilot and Copilot generates the code. AI has become a total intermediary between users and open source projects. And that intermediary doesn’t pay commissions.
If improving your product no longer leads to greater economic sustainability, the project is destined to die or be absorbed by a big tech that can afford to maintain it without monetising.
The End of an Era
While I was reflecting on this, I found myself with more questions than answers. And they’re heavy questions, the kind that keep me awake.
Can open source survive without big tech backing? If projects can no longer monetise, only those sponsored by Google, Meta, Microsoft will survive. But is it really open source if it’s controlled by corporations? Or is it just “source-available” with an aura of community?
Do we need a new social contract? Some talk about treating open source as “digital public infrastructure” and funding it with public money. But which government wants to fund thousands of software projects? And who decides what deserves the funding?
Should AI providers pay for training data? Intuitively it seems right. But how do you implement it? With what legal mechanism? And if data is released under an open source licence, how do you impose payment without violating the principles of open source itself?
Is this a temporary or permanent transition? Maybe in a few years new business models will emerge that we can’t even imagine today. Or maybe not. Maybe we’re watching the end of an era, and what comes after will be radically different from everything we knew.
There’s an almost comic irony in this tragedy. Tailwind is a victim of its own success. LLMs are so good at generating Tailwind code precisely because Tailwind did an exceptional job of creating clear documentation, practical examples, an active community. All that high-quality data trained the AIs perfectly. But that quality is now the weapon killing it.
It’s as if you’d spent years creating the best programming course in the world, free, accessible to everyone. And then someone builds a robot that watches all your videos, memorises everything, and starts answering students’ questions in your place. The better your course, the more effective the robot. And you stay there, watching while students stop visiting your site, because they have the robot.
I don’t know if there’s a way out of this vicious cycle. I don’t know if Wathan will manage to save Tailwind Labs. I don’t know what will happen to all the other open source projects in the same situation.
What I know is that we’re at a turning point. AI promised to democratise software development, to make programming accessible to everyone. And probably in some way it’s doing that. But the cost may be the entire infrastructure of open source projects, quality documentation, communities of experts that for decades made that democratisation possible.
When the last sustainable open source project closes, what will future AIs train on? It’s a question I have no answer to. And maybe the very fact that there’s no obvious answer is the most worrying thing of all.
Maybe what we’re living through is the moment when the open source ecosystem discovers it was too generous, too open, too trusting. Our sector was the only one that decided knowledge and innovation should be distributed publicly. It gave knowledge to the world for decades, building the foundations on which almost all modern software rests. And now that knowledge has been extracted, processed, and turned into commercial products that are its direct competitors.
It isn’t anyone’s fault in particular, perhaps. It’s the inevitable result of misaligned incentives, of an economic model that rewarded openness without considering long-term consequences. But the result is the same: an entire generation of developers who built incredible things may end up without the means to keep doing it.
And that, for someone like me who has always believed in the power of open source, is simply devastating.
Key takeaways
It isn’t a choice between blocking and allowing crawlers: both options lead to the same end—it’s a deadly vicious cycle.
If developers stop writing documentation and Stack Overflow empties out, what will future models train on? Model collapse is the next chapter.
Only projects sponsored by big tech will survive—but is it still open source if it’s controlled by corporations?
Questions & answers
What's the vicious cycle AI sets off on open source?
LLMs were trained on millions of open source repositories without compensation or attribution. Today developers ask LLMs to write code instead of consulting the original projects. Visibility, donations, contributions, bug reports—the economic and social fuel of open source—shrinks. Maintainers burn through fewer resources but also less feedback, less engagement, less sustainability. The cycle that kept the community alive shuts down.
Why did Adam Wathan (Tailwind) sound the alarm?
Because he sees the data first-hand. Traffic to the documentation of popular projects is dropping while LLM queries grow. Developers get “good enough” answers without visiting the project site, without seeing the Sponsor button, without meeting the community that makes the project something more than a piece of code. For him it isn’t an intuition—it’s a graph in decline.
Can LLMs replace open source communities?
No, and that’s the most worrying point. LLMs reproduce code taken from repositories but cannot maintain it, evolve it, fix it. When you entrust knowledge to the LLM and the original collapses, the LLM loses its source of updates. It’s like sawing the branch you’re sitting on, but slower: damage shows up after two or three years of training on its own outputs.
What can developers and companies do to break the cycle?
Developers: when an LLM gives you a solution, go to the source, verify, star it, consider donation or sponsorship. Companies building on top of LLMs: explicitly support the open source projects used (Anthropic has started some initiatives, but it’s little). Regulators: consider attribution and compensation obligations for training data—as in other industries (music, publishing) where the creative cycle has to be preserved.
The author
