The Swiss labor market is confronting a figure that sounds an alarm for anyone watching the concrete impact of Large Language Models. A team of researchers sifted through roughly 7.3 million online job advertisements and found that positions aimed at first-time entrants fell in 2025 by almost a third compared with the average preceding the explosion of generative AI. This is not a marginal adjustment but a contraction that redraws the map of required skills.

The drop in entry-level openings

Conducted on Swiss data but with implications that cross borders, the study shows the phenomenon is far from uniform. The steepest declines cluster in tasks that AI tools can perform relatively easily: drafting standardized texts, analyzing structured data, administrative support. In these areas, the number of ads for junior profiles shrank far faster than in roles demanding physical dexterity, complex human interaction or non-algorithmic creativity. The signal is clear: automation is not merely supplementing expert work; it is eroding the base of the professional pyramid.

What it means for on-premise stack builders

Those evaluating self-hosted deployments of language models – perhaps to retain data control or reduce TCO at high volumes – should read these numbers carefully. The disappearance of entry-level figures is not an abstract problem: in many organizations, young professionals represented not only low-cost labor but also the pool from which to nurture internal skills for infrastructure management. If companies start replacing precisely those tasks with LLM-based agents, the know-how necessary to train, fine-tune or secure an on-prem model risks becoming concentrated in ever fewer hands.

The sovereignty paradox

Internally managed infrastructure – be it on bare metal, in an air-gapped Kubernetes cluster or at the edge – promises to preserve data confidentiality and avoid lock-in with cloud providers. Yet to function effectively it needs personnel able to manage inference pipelines, optimize latency, administer GPU nodes and intervene when a model suffers performance drift. If the labor market stops producing junior figures to train, those running on-prem stacks may soon find themselves with state-of-the-art tools and an increasingly skeletal team, or forced to compete with exorbitant salaries to attract a handful of senior profiles already trained elsewhere.

Beyond the alarm: a reading for real-world deployment

The Swiss research offers no ready-made solutions, but it illuminates a variable often overlooked in total cost of ownership analyses: the availability and evolution of human capital. For those designing on-premise inference environments, this probably means weaving into TCO calculations investments in internal training and mentoring, or choosing architectures that simplify daily management – for example, pushing aggressive quantization to reduce the hardware footprint, or adopting serving frameworks that automate part of the maintenance operations. It is no longer just a matter of GPU, VRAM and throughput: the real game is about making the coexistence of automation and human skills sustainable over time.