Codeberg, a hosting platform for open source projects managed by a nonprofit association, has approved two AI motions through member voting. The first commits to never using code or data from hosted projects to train LLMs. The second, more divisive (358 in favor against 144 opposed, with 50% turnout), modifies the terms of service to ban vibe-coded projects: projects mass-generated by individual developers with LLMs, often supporting more platforms than they have actual users, consuming disproportionate resources.

The position comes at a critical moment in the training data debate. As we reported on July 22 regarding Ben Thompson’s fair use proposal, AI labs train on billions of pages without permission while forbidding others from learning from their models. Codeberg brings the voice of those who produce and freely share that data.

For those building with AI, this raises concrete infrastructure issues. Crawlers from AI labs overwhelm volunteer infrastructure: they read every page, every issue variant, the entire Git commit history. The costs fall on platforms. An SSD drive that Codeberg paid €700 for a few years ago now costs €3,700, often out of stock.

Codeberg raises a point few state clearly: free software is a common resource that the AI industry consumes without compensation. It remains to be seen whether this stance will remain isolated or become a model for other platforms.

In detail

Codeberg is not just any forge. It’s a hosting platform born as an ethical alternative to GitHub, run by Codeberg e.V., a German nonprofit association funded by donations, which owns its own hardware and offers free services to developers and open source projects. The two motions approved on July 23, 2026 come from the annual member assembly, where proposals are discussed live and voted asynchronously over 14 days.

The first motion builds on the existing privacy policy (“We do not want to need your data”) and extends it to LLM training. The text states that Codeberg will not use or permit the use of project code or data to train generative AI. The explicit reasoning is that these technologies are “incompatible with the responsible creation and maintenance of free and open source software”. It’s a principled stance from the association itself, not individual projects: the ban applies to the infrastructure, not the code license.

The second motion sparked debate. Approved with 358 votes in favor, 144 against, and 14 abstentions (50% turnout among active members, high for this type of assembly), it modifies the terms of service to ban “vibe-coded projects”. The term comes from the practice of rapidly generating code with LLMs iteratively, without a real community behind the project. Codeberg describes the phenomenon precisely: projects with a single developer, no users, but heavy CI/CD, large binary releases, and more supported platforms than people using them. These projects consume as much storage and compute as major community projects, without delivering shared value.

The practical issue is infrastructure under stress. Codeberg documents how AI lab crawlers systematically read every page: repository files and all issue filter variants, the complete Git history, files at every commit even when identical. This generates expensive database queries, degrades service for legitimate users, and forces the team to build defensive mechanisms (rate limiting, blocks) that worsen experience even for actual users.

Then there’s hardware costs. The LLM industry has driven up SSD and memory prices. Codeberg cites a concrete example: the disk type that cost €700 a few years ago now costs €3,700 and is often unavailable. Even owning your own hardware, and thus less exposed to cloud costs, renewal and expansion have become prohibitive. Money that doesn’t go to new services or the mission, but to hardware inflated by the AI market.

The thread connecting this story to the broader debate is training data and copyright. Beyond Thompson’s fair use proposal, on July 14 we reported how Satya Nadella accused AI labs of training on others’ data while forbidding distillation of their models. Codeberg brings the argument from the side of those providing data without compensation: free software is training material, but the ecosystem producing it receives nothing in return.

The limitations of this move are clear. Codeberg is a relatively small platform compared to GitHub, and its terms bind only those hosting projects there. Crawlers can extract code from public clones elsewhere. Banning vibe-coded projects is hard to enforce without arbitrariness: where’s the line between a project generated with AI by a solo developer and a legitimate project with few contributors? Codeberg itself acknowledges it will need to clarify the practical impact.