Even if we had a private jet parked in our garden instead of a bicycle, hardly any of us would think of taking off and flying around the globe just to pick up bread rolls from the bakery next door. Yet, in a metaphorical sense, that is exactly what we are doing in the way we use artificial intelligence.
For simple tasks such as structuring a PowerPoint presentation – something we would have done ourselves until very recently – we now turn to one of the most complex technological systems ever built. Based on billions of parameters, trained on vast amounts of data, and operated in global data centres whose energy consumption exceeds that of a medium-sized city. The technical effort behind a single prompt is completely disproportionate to the complexity of the task we are asking it to perform.
Whether we are asking AI to structure texts or simply check our spelling, we remain unaware of the enormous machinery operating in the background. A request disappears into an input field, and a few seconds later an answer appears. Between our action and its consequences lies an infrastructure that remains invisible to us: power grids, cooling systems, fibre-optic cables, specialised chips, water consumption, global supply chains. Historically speaking, this disconnect between human action and systemic impact is unprecedented. We prompt as if there were no tomorrow.
The consequences of our digital actions remain largely hidden from view. Who is actually able to break down the infrastructure and resource costs generated when we casually ask “Chatty” whether Sunday will be good weather for a barbecue?
We do it because we can. Because the constant availability of generative AI has, within just a few years, created a new cultural practice – one that exists somewhere between maximum convenience and the outsourcing of our own thinking. Is there actually a word for this yet?
We talk a lot about the new efficiency that artificial intelligence promises to bring. Far fewer people ask a more fundamental question: How much AI is actually enough?
Not every task requires the most powerful language model available, and not every cognitive challenge justifies the use of a global AI infrastructure. In a very short period of time, we have become accustomed to using the most powerful tools available to us for everyday tasks. As if we were using a pneumatic hammer to turn every screw, or – returning to the metaphor above – flying a private jet to the bakery to pick up a few bread rolls.
What is often overlooked in all the enthusiasm surrounding artificial intelligence is this: the infrastructure we use so naturally today does not belong to us. It is operated and controlled by a small number of companies.
The history of the internet shows how these developments tend to unfold. Services start out as free, grow rapidly, become indispensable, and eventually disappear behind paywalls. Or they remain free because we are no longer the customers – our data and our behaviour become the actual product.
There is little reason to assume that this cycle will not repeat itself with large language models. The consequence would be that, in the near future, far fewer people will be able to afford the luxury of unlimited AI access.
For everyone else, the challenge is to invest energy and creativity into developing alternatives. The answer is not to reject AI, but to build a better-organised form of AI: one whose infrastructure is open, decentralised and collectively governed rather than controlled by a small number of companies.
This development is already becoming visible: small, specialised language models that can be operated locally, connected with one another and managed independently. Technically, this is no longer a vision of the future. Open-source models such as Llama, Mistral or Gemma can already be run locally on personal hardware or within private infrastructures.
Tools such as Ollama, Open WebUI or LocalAI make it possible to set up your own AI environments without depending on a central cloud provider. This opens up new possibilities for communities, companies and public administrations to build and operate their own AI infrastructures.
The idea is not to replicate a single, enormous model, but rather to create networks of specialised AI agents. One agent for legal questions, another for medical knowledge, another for local administrative data, or another for internal business processes – each contributing its own specific expertise. Through open interfaces such as MCP, these systems can communicate with one another and collaborate on solving complex tasks.
This approach also opens up new organisational possibilities: communities could share computing resources and infrastructure. Members could collectively decide which models are used and according to which principles these systems are developed and maintained.
The costs involved are more manageable than many people assume. Depending on the use case, local AI systems can run on existing hardware or on jointly operated computing resources. The main investments are related to hardware, maintenance, administration and ongoing development.
At the heart of this lies what we call digital sovereignty: the ability to decide for ourselves which technologies we use and how they should function. This also raises the question of whom these technologies ultimately serve - and how important data security and privacy are to us.
Local language models and collectively operated AI networks could pave the way towards a form of self-determination that we already know from cooperatives: collectively owning the means of production, democratically shaping the rules of use, and sharing responsibility, costs and benefits.
More than 200 years ago, cooperatives emerged as a response to the challenges of industrial capitalism. Today, we find ourselves at a similar turning point. This time, it is not about dairies, consumer cooperatives or agricultural credit systems. It is about the digital infrastructure on which our society will increasingly depend.
Disclaimer: This article was originally published on platformcoops.de.
This summer, I had the opportunity to speak twice about the question of how artificial intelligence can be organised and shaped through cooperative models: first at the Perspectives Festival in Hamburg and then at the AI for Good conference of the United Nations in Geneva.
Two very different contexts, but the same central question: What if AI infrastructures were no longer operated and controlled exclusively by a small number of large companies, but instead by small, interconnected communities? What could emerge if people collectively took responsibility for the technologies shaping our digital future?
This article summarises the key ideas from both presentations.