Seventy years after the Dartmouth workshop that coined the term “artificial intelligence,” the technology is at the center of an unprecedented adoption wave. John McCarthy, Marvin Minsky, and Claude Shannon envisioned machines capable of simulating human intelligence. Today we live in an ecosystem dominated by Large Language Models and autonomous agents, but for those deploying AI, the real question is no longer just “what can AI do” but “where and how do I run it.”

From early dreams to transformers: a history of promises and winters

The journey of AI has been marked by waves of euphoria and sharp cooling periods, the so-called “AI winters.” After early steps with neural networks in the 1940s and Turing’s test in 1950, 1956 established the field formally. Lisp, machine learning, and expert systems followed. Then expectations clashed with hardware and theoretical limits. The current “spring” exploded with the transformer in 2017 and ChatGPT in 2022, but the historical lesson is clear: every technological leap is tied to infrastructure constraints that determine real-world success.

Strengths and fragility of LLMs

Generative models excel at pattern recognition, generating text, code, and multimedia content. They automate repetitive tasks, reduce operational costs, and speed up decision-making. Yet they come with significant risks: hallucinations, bias, opaque decision-making, security vulnerabilities, and misuse of personal data. When a company uses an LLM via cloud APIs, it loses visibility into many of these aspects. Governance becomes a key issue, and with it the choice between cloud and on-premise.

The on-premise crossroads: beyond hardware

Running LLMs on your own infrastructure is no longer just about cost. For organizations handling sensitive data, regulated by GDPR, or operating in critical sectors, self-hosting offers direct control over privacy, audit, and model customization. However, on-premise deployment requires skills in quantization, VRAM optimization, and inference pipeline management. It is not a path without trade-offs: TCO can be high, but data sovereignty and performance predictability are hard to ignore.

Beyond the celebration: what the next seventy years hold

AI’s 70th anniversary arrives as IEEE and other institutions push for ethical standards and guidelines. History shows that without careful control, hype waves can drive excessive investments. For those evaluating deployment architectures today, the challenge is to integrate model innovation without yielding to a “cloud at all costs” pressure. The choices made now – on quantization, dedicated hardware, serving frameworks – will shape the ability to keep AI human-centered and aligned with the data owners’ needs.