It’s not the recent graduates who are trembling, but fifty-somethings with high salaries. This is not bar-room speculation but research from Boston College led by Geoffrey Sanzenbacher, showing that workers aged 55 and over in AI-exposed occupations are leaving the labor market at higher rates than before ChatGPT launched. The debate about AI and jobs, so far focused on young graduates, has overlooked a segment that is hit first and hardest: well-paid seniors.

This dynamic is not merely demographic. It is a structural signal that directly challenges companies’ deployment choices. The spread of self-hosted Large Language Models — running on local servers, away from public clouds — is creating an environment where automating complex cognitive tasks becomes more economically attractive than ever. An on-premise LLM can absorb analysis, reporting, internal consulting, and even tactical decisions, all tasks once guarded by experienced professionals, without giving up data control.

Here’s the point: the hardware needed for local inference (GPUs with plentiful VRAM, modular architectures, low-latency networks) has matured to a point where the total cost of ownership tips the balance. Enterprises with adequate on-site compute can internalize AI, avoiding recurring cloud fees and, crucially, complying with privacy and regulatory requirements. This removes the friction — both psychological and accounting — that used to block the automation of high-wage roles. Investing in local compute capacity thus becomes not just an efficiency lever, but an instrument of organizational transformation that hits the top of the salary pyramid.

Sanzenbacher’s research tells us the phenomenon is already underway. ChatGPT’s launch acted as a detonator: not because it made previously unthinkable things possible, but because it made the concrete potential of language models visible to everyone. Since then, the exit of senior workers has accelerated, and not by voluntary choice. The data suggest that, in highly exposed sectors, firms are opting for forced generational turnover, conscious that three decades of experience can be replaced by an inference system running on consumer GPUs or corporate servers.

The implicit pact linking seniority to job security is now under pressure. If until yesterday accumulated knowledge was a shield against automation, today it risks becoming a marginal cost that is too high when a local model, trained or enriched with the organization’s data, produces comparable outputs at a fraction of the cost. This is where technological sovereignty comes into play: those who manage deployment in-house not only retain data but can also precisely calibrate the model on business processes, accelerating the moment when the machine outcompetes humans on a cost-value basis.

The trajectory is not inevitable, but it is heavily incentivized by the economics of on-premise. For those evaluating self-hosted deployment, there are complex trade-offs between flexibility, scalability, and in-house skills, but the vector is clear: automation is climbing the professional ladder, starting precisely from the highest rungs. The side effect is a labor market divided between those who design and train models and those who are replaced by them, with an increasingly narrow window of age to stay in the first category.