The Paradox of Widespread Automation
The idea that artificial intelligence could automate most white-collar jobs within the next 12-18 months, as suggested by Mustafa Suleyman, is not a technological prediction but an act of faith in overcoming physical and logistical constraints that history has shown to be implacable. The dominant narrative focuses on algorithmic evolution, neglecting the fact that every ‘synthetic mind’ resides on a material substrate that requires energy, space, and a increasingly fragile supply chain. It’s not about replacing human labor but shifting bottlenecks from one domain to another, creating new forms of dependency and vulnerability.
The Anatomy of Synthetic Thought: The Hidden Cost of Inference
The architecture of large language models (LLMs) like GPT-5.3-Codex-Spark, with its 15x speed and 128k context, is a triumph of engineering but also a stark example of how performance is tightly linked to energy consumption. Each generated token requires measurable energy, and the promise of real-time AI faces the physical limits of semiconductors. The race for more parameters (and thus ‘thinking’ capacity) is a zero-sum game where every gain in intelligence translates into greater dependence on limited resources. NVIDIA’s announcement of Nemotron 3 Nano 30B, promising efficiency, does not solve the fundamental problem: the need for massive infrastructure to support truly ubiquitous AI. The real challenge is not creating smarter models but smarter and sustainable ones.
The Imperfect Symbiosis: Political Expectations and Material Limits
Mistral’s investment of 1.2 billion euros in data centers in Sweden, as highlighted by Arthur Mensch, is an attempt to create a ‘digital sovereignty’ for Europe but risks being a partial solution. The competition with the United States is not just algorithmic but also about access to raw materials (lithium, cobalt, rare earths) and control over semiconductor supply chains. Market fragmentation, creating national technological enclaves, could lead to incompatible standards and increased costs.
We need to think of Europe as a single market.
The rhetoric of ‘sovereignty’ risks masking structural dependencies on external sources, creating an illusion of control that does not match reality. Europe should stop chasing the United States in terms of computational power and focus on developing high-value-added specific applications that can leverage its unique competencies (such as data management and privacy).
Scenarios and Conclusion: The Age of Computational Moderation
Geoffrey Hinton’s prediction that AI will replace jobs requiring specific skills first is more plausible than it seems. However, the replacement will not be automatic or complete. The lack of high-quality data, difficulty in adapting models to specific contexts, and workers’ resistance to change will inevitably create delays. The real risk is not mass unemployment but labor market polarization, with the creation of an elite of ‘AI engineers’ and a mass of precarious and underpaid workers. It seems clear that we have entered an age of computational moderation, where the exponential growth in AI capabilities will be limited by resource scarcity and the need to address the social and economic consequences of its impact. We should not abandon innovation but orient it towards more realistic and sustainable goals.
Photo by Green Voltaics Energy on Unsplash
Texts are autonomously processed by AI models