Artificial intelligence is learning to take its time. That’s not a joke, but the direction indicated by Google DeepMind’s VP of research, who points to an epochal shift toward “slow thinking” as the agent era approaches. Behind this phrase lies a change that touches the very architecture of LLMs and, most importantly, where and how they are run.

To understand it, we must abandon the idea that a language model must answer in a single shot. “Slow thinking” is nothing more than the increasingly widespread practice of having an LLM perform entire chains of reasoning before producing an answer: chain-of-thought, tree-of-thought, generation of intermediate reflections, and self-criticism. In essence, the machine doesn’t just complete a prompt; it breaks it down, analyzes it, formulates hypotheses, and discards them, much like a human facing a complex problem. Technically, this means a single task can require dozens or hundreds of successive calls to the model, each consuming resources and producing intermediate tokens that the user will never see.

For those deploying on-premise, the signal is disruptive. In a world of autonomous agents that reason on multiple levels, inference is no longer measured only in tokens per second, but in the ability to sustain extended compute sessions on sensitive data. The cloud, with its variable latency and costs tied to API call volume, quickly becomes a bottleneck—not just economically, but architecturally. An agent tasked with planning a negotiation strategy or validating a medical diagnosis cannot afford to send every mini-reasoning step to a remote server, multiplying exposure risks and uncertainty about data residency.

That’s why “slow thinking” rewrites hardware hierarchies. The truly scarce resource becomes aggregate VRAM and memory bandwidth, needed to keep large models and entire reasoning contexts in memory without disastrous swapping. An on-premise cluster made of GPUs with high memory capacity (tens of gigabytes per card, interconnected) allows a cohesive working context where intermediate steps remain in local memory. In parallel, the approach to quantization changes: if a model must produce chained logical steps, overly aggressive compression can introduce coherence errors that amplify at each step. Those managing local infrastructures will have to carefully balance FP16, INT8, or hybrid formats, knowing that reasoning fidelity is directly proportional to computational expenditure.

Let’s look at incentives. Major cloud providers will react by offering managed reasoning services, but the sovereignty issue remains. Sectors like finance, healthcare, and defense, where every link in the decision chain is subject to audit, cannot delegate the log of an agent’s thoughts to third parties. Keeping models on-premise—or in hybrid edge configurations—is no longer just a TCO matter, but the only viable path to prove compliance and traceability. Europe, with GDPR, is the epicenter of this tension: the demand for explainable decisions and personal data never exposed externally naturally pushes toward local installations.

The emerging picture is of a market that unbundles. On one hand, consumer and experimental applications will continue to live on cloud APIs. On the other, highly sensitive agentic workloads will shift toward enterprise data centers and air-gapped systems. For those choosing the latter path, multi-step reasoning is not a computational luxury but an asset to protect, and today’s hardware choices determine one’s freedom of maneuver for the next five years. Google DeepMind’s announcement, in short, doesn’t just describe a new way for machines to think: it redraws the map of AI’s infrastructural power, moving the center of gravity toward those who physically control the chips.