Recently codeberg passed two motions regarding LLM usage:

https://blog.codeberg.org/protecting-our-floss-commons-from-llms.html

The first vote is about codeberg not using the hosted code to train LLMs. The second one is the one I want to focus on; it prohibits vibe coded projects on the platform.

They mention in their post the “digital divide”. Large cloud companies are the only ones able to afford the infrastructure to handle all the costs associated with using LLM and being part of the 2026. Smaller actors can’t afford to exist on their own and instead they have to rent from the cloud providers.

I strongly resonate with this divide as I perceive it constantly. For example, I don’t think it would be possible for any startup to create github today if it didn’t exist. The amount of noise of slop projects would make too hard to serve legitimate users while keeping it free or with a harmless “pay for private repos” that github used to have.

So codeberg had the option to try to accommodate to these forces and become bigger and more corporate-like, or to stay small and serving the people that want the community above the market or energy forces.

The ecological case

Creating and modifying software involves effort. Developers have to put the thought to understand and translate the requirements into whatever needs to exist.

Cloud availability has accelerated the conversion from thought and reading into literal computer effort. I am talking about pipelines, deployments and other verification tools. Some of the effort has shifted to a computer running in the cloud, reducing the amount of human effort. Simply push the code and the pipeline will run.

LLMs are a big step in the same direction. A set of prompts and documents (a specification) plus a lot of energy and materials in a data centre produce plausible code. Leaving aside how good the code is, it reduces the amount of human effort to generate code. I wrote on this trade off recently:

[[2026-03-15-the-llm-code-bet]]

Both in the cloud revolution and in the AI revolution, the human cost of creating software goes down. Jevon’s paradox seems at play in here as these changes have increased the amount of code produced.

At the same time, LLMs have reduced the effect of trust and human networks in the process. The proof of work is not there. So I believe that people are less trusting of new projects unless they come with some guarantee in the form of robust tests, verification or guardrails.

The new ways of work require more energy to generate code, more energy to verify the code and also produce more software in general. It makes sense if energy is very cheap.

The bill please

So who is paying for this extra energy cost? I would say originally the investor class paid for it, looking for future profits. As LLMs improved, working people are starting to pay directly by losing their jobs. We are also losing by losing access to personal computing.

Energy wise the extra demands come from other sectors of investment. All the capital that go into data centres is not available for something else, like infrastructure, health care or education.

Ultimately the capital cost reflects the cost to nature so we are all losing out on that one, although we call it externalities.

This is why I think codeberg is right not to be part of this extra cost dynamic, and wait until we find the amount of LLM usage that actually increases productivity without increasing energy or human collaboration costs.

P.S.: I haven’t mentioned the likely bubble bursting nor that the multiple wars around oil will likely impact the price of energy, making LLM software development fragile. Maybe another day.