Every thirty years, a new technology is presented as the definitive answer to humanity’s problems.
Electricity was supposed to guarantee general prosperity. Internet was supposed to democratize knowledge. Social networks were supposed to bring peoples together. Cryptocurrencies were supposed to liberate citizens from banks. The metaverse was supposed to replace the real world.
Today, it is artificial intelligence that takes over this endless promise: it would heal patients, save schools, speed up justice, modernize administration, reboot growth, and even resolve political crises. One might think that public policy could be replaced by lines of code.
Evgeny Morozov gives a name to this logic: techno-solutionism. He designates this tendency to transform political, social or economic problems into technical problems supposedly solvable by an innovation. Artificial intelligence is today its most accomplished expression. We no longer debate collective choices; we expect a technology that will spare us from making them.
In this narrative, three illusions overlap. First, the idea that AI would be truly “intelligent.” Next, the belief that it would be an immaterial, almost disembodied technology. Finally, the conviction that it constitutes an open revolution accessible to all, when in fact it is one of the most concentrated industries in history.
First, the illusion of intelligence. It is embedded in the vocabulary itself. To say “artificial intelligence” implies an ability to understand, to reason, to decide. Yet current systems, notably large language models (LLMs, i.e., statistical models capable of generating text by predicting the most probable words), do not “understand” anything in the human sense. They produce coherent sequences without intention or consciousness. As Emily M. Bender reminds us, they are statistical parrots: they recombine language based on massive correlations, without a representation of the world. Gary Marcus emphasizes a fundamental limit: these systems have neither causal understanding, nor robust reasoning, nor a true model of reality.
A simple example suffices: a model can draft a plausible medical diagnosis, then in the following sentence propose a treatment that is incompatible with that same diagnosis, without perceiving the contradiction. Or produce a seemingly coherent legal argument, but based on repealed laws. It is not an intelligence that errs: it is a system that generates plausible language.
The illusion comes from the fact that language gives the appearance of thought. But here, it is only a simulation of it.
Next, the illusion of immateriality. The second illusion is that of a “clean,” lightweight technology, floating in a digital cloud. The cloud evokes something dematerialized. In reality, artificial intelligence is a heavy industry.
Behind every response lie multi-hectare data centers, hundreds of thousands of specialized processors (GPUs, i.e., computing units optimized to process billions of operations simultaneously), massive cooling systems, dedicated power networks and global fiber-optic infrastructures. A single query can mobilize up to a liter of water to cool the servers required for its processing. A modern data center consumes as much electricity as a city of 50,000 inhabitants. A simple interaction with a conversational agent thus rests on an invisible but extremely material industrial chain. As Kate Crawford shows in her work, AI is an extractive industry: it draws on energy, water, minerals, human labor and large-scale data. Concrete example: chip manufacturing requires globally distributed supply chains of rare metals, ultra-precise factories and massive energy consumption even before the first line of code. The image of the “cloud” is thus misleading. It is rather a global industrial infrastructure, invisible but physically heavy.
Finally, the illusion of universal accessibility. This third illusion is the most structural: that of a democratic technology, accessible to all, and capable of redistributing power.
In reality, AI is becoming an extremely concentrated industry. Every industrial revolution relied on a dominant resource: coal, oil, semiconductors. AI accumulates all these dependencies at once: massive data, energy, infrastructure, talent and above all, capital.
The large language models (LLMs, models capable of generating text, code or images from vast data) require investments of tens of billions of dollars. Their training mobilizes hundreds of thousands of GPUs (specialized processors) and infrastructures comparable to those of entire regions. As models become more capable, their production cost increases. The result is paradoxical: a technology marketed as accessible becomes increasingly centralized in practice. A few companies simultaneously control the chips, data centers, models, cloud platforms and access interfaces. The AI economy does not diffuse power; it concentrates it.
We are thus witnessing a subtle transformation: countries and companies no longer necessarily produce artificial intelligence; they use it via subscriptions. They rent computing power, outsource their data and build their services on infrastructures they do not control. This dependency is economic, but also technological and strategic.
The consequence: a new economy of concentration. Artificial intelligence thus becomes the new space for capital accumulation. The more it advances, the more it strengthens entry barriers.
The effect is cumulative: more data means better models; more models attract more users; more users capture more data. This virtuous circle for a few actors is a closed loop for others. AI does not produce a distributed economy, but an economy of vertically integrated platforms.
And Tunisia in all this?
In this context, the temptation would be to want to “participate” in the global race by building heavy infrastructures. That would be a strategic mistake. Building massive data centers in a country facing water stress, electricity tensions and budgetary constraints would amount to investing in the most expensive and least specialized layer of this industry.
The question is not: how to rival the giants? But: where does accessible value creation lie? The answer lies in the upper layers of the value chain.
First, small, specialized models, trained on quality data, adapted to precise uses: health, agriculture, logistics or industry. In these domains, the precision and relevance of the data often count more than the size of the model.
Next, AI agents, i.e., systems capable of using these models to execute complete tasks: administrative automation, legal assistance, optimization of public procedures, business support, customer relations, document management. The value no longer lies in the model itself, but in what it enables to do concretely.
Finally, the structuring of sectoral data must be considered as strategic infrastructure: agriculture, water, health, energy, public finances.
Whoever controls data quality controls a decisive part of the value chain. The central misunderstanding around artificial intelligence is not technical, but political. It consists of believing that it is a universal, neutral technology, equally accessible to all. In reality, AI is less a revolution of intelligence than a revolution of concentration: concentration of data, of capital, of energy, of infrastructure and of power.
In this context, sovereignty does not consist of reproducing the tech giants, but of wisely choosing one’s place in the value chain. Modernity is not about participating in every technological race. It is about knowing which ones it is rational not to pursue.
Hédi Sraieb, State Doctorate in Development Economics
