This week, Konstantin Ryabitsev — the man who keeps the Linux kernel’s infrastructure from falling over — published a blog post with a number in it that should have made more noise than it did. git.kernel.org, the server that hosts the kernel’s entire commit history, now receives about six million requests a day for random commits. Of those, 66% are batted away by a proof-of-work challenge called Anubis. A third of them — roughly two million requests a day — are now solving the math and getting through. And after what Ryabitsev calls “a bunch of generous assumptions,” legitimate human traffic is about 2%.
Two percent.
The kernel’s git server is, for all practical purposes, no longer a tool for developers. It is a training-data mine. The humans are incidental.
The 2% Is the Story
Everyone wants to talk about the scrapers — the AI companies vacuuming up commit history because, as Ryabitsev notes, the full history of kernel commits is “worth its weight in gold as a source of training data.” But the scrapers are boring. Scrapers gonna scrape. The number that should keep people up at night is the 2%.
Think about what that means. The Linux kernel is the most important software project on the planet. Every Android phone, every cloud server, every router, every car’s infotainment system runs it. And the infrastructure that serves its source code is now, by volume, 98% non-human. The people who actually write the kernel — the ones who need to fetch a commit, check a patch, bisect a regression — are a rounding error on their own server.
A first-time contributor who spent three days trying to clone the tree through a university proxy put it plainly in a Libera.Chat channel that still logs to a text file: “I thought I was doing something wrong. Turns out I’m just outnumbered by robots by fifty to one.”
Proof-of-Work Is the Only Honest Price
Here’s the part nobody wants to say out loud: the Anubis challenge is functioning as a price. A real one.
The bots aren’t getting the data for free anymore. They’re burning actual compute — electricity, hardware cycles, money — to solve a math problem before every request. And a third of them are now willing to pay that price. That’s a market signal. The kernel’s commit history is so valuable to AI training that scrapers will spend real resources to get it, even when the maintainers have deliberately made it annoying.
But here’s the twist: the maintainers aren’t collecting that price. The compute the bots burn doesn’t go to Ryabitsev or to Linus Torvalds or to the kernel’s infrastructure budget. It just evaporates. The proof-of-work challenge is a toll booth where the toll is paid into a furnace.
That’s not capitalism. It’s not socialism either. It’s something stranger — a commons defended by a math problem, where the cost of extraction is real but the benefit of extraction accrues entirely to the extractor. The kernel developers get nothing except slightly less congestion.
The License Wars Are a Sideshow
For the past two years, the open-source world has been fighting about licenses. The Open Source Initiative spent months trying to decide whether AI models trained on open code can call themselves “open source.” Lawyers have drafted new clauses. Foundations have issued statements. It’s all very earnest.
And it’s all beside the point.
The license on the Linux kernel is GPL-2.0. It says nothing about scraping. It was written in 1991, when the idea that a machine would read every commit ever made and use it to train another machine was science fiction. No license amendment is going to stop a crawler that’s willing to solve a proof-of-work challenge two million times a day. The battle isn’t at the license layer. It’s at the network layer — in the Anubis challenges, the rate limits, the CDN rules, the infrastructure decisions that determine whether a public good stays public or becomes a free input to someone else’s model.
Ryabitsev’s post is framed as a status update: “we’re not quite overwhelmed.” But read it again and it’s a warning. The kernel’s infrastructure is holding, for now, because one person wrote a proof-of-work challenge and a third of the bots haven’t yet decided the data is worth the electricity. The moment that math changes — the moment the value of kernel history to a training run exceeds the cost of solving Anubis at scale — the 2% becomes 0.2%, and the humans are gone entirely.
The uncomfortable truth is that neither side of the open-source debate has an answer for this. The “charge for access” crowd wants a pricing mechanism, but the kernel developers don’t want to run a billing department — they want to be left alone to write code. The “keep it free” crowd wants the commons to stay open, but a commons that’s 98% bots isn’t a commons anymore. It’s a farm.
The only thing standing between the kernel and total extraction is a math problem that costs the extractors something and pays the maintainers nothing. That’s not a solution. It’s a holding action. And the 2% figure is how much time we have left.