For years, the Sintra retreat was the thermometer of central banks’ obsessions: inflation, interest rates, labor markets. But this summer, when the world’s most powerful central bankers met in the usual hillside town near Lisbon, the script changed. Artificial intelligence stole the spotlight, not as a tool for macroeconomic analysis, but as a source of systemic uncertainty. As reported by The Next Web, the gathering had a single organizing subject: AI, and «the awkward fact that nobody in the room» had a clear answer on how to handle it.

That’s no minor detail. Central banks handle some of the most sensitive data on the planet, oversee payment systems, and regulate financial stability. The entrance of Large Language Models and opaque decision-making systems raises questions that go beyond economic theory. Who trains these models? On whose data do they run? And, crucially, where does that data physically reside? A question that shifts the focus from academic speculation to concrete deployment choices.

The trust short-circuit

The promise of AI in finance is huge: fraud detection, predictive analytics, reporting automation. But for a central banker, the risk is not a mere statistical error: it’s the domino effect across entire markets. A model that “hallucinates” on an inflation report could trigger unjustified moves. That’s why control becomes the key variable. A cloud-based LLM, managed by third parties with servers in multiple jurisdictions, introduces regulatory and security fragilities that an institution of this caliber cannot afford.

This is where data sovereignty enters the picture. It’s not a technical quirk: it’s the foundation for auditing, explaining, and, if needed, stopping a system. The on-premise approach, or at most a well-defined hybrid one, returns to center stage. Not out of digital isolationism, but because trust in models is built only when the infrastructure is under your own control. AI-RADAR, which dedicates ample space to trade-offs between self-hosted and cloud, has documented scenarios where the TCO of an on-premise solution, when compared with compliance costs and reputational risks, proves advantageous in the long run.

Hardware isn’t an afterthought

People often think of AI as a software matter. But when talking about local deployment, the node is hardware: video memory (VRAM), compute capability, inference throughput. To run medium-sized LLMs with custom fine-tuning on proprietary data, you need machines with enterprise-class GPUs, such as NVIDIA A100 or H100 80 GB, and architectures that minimize latency without sacrificing accuracy. This isn’t science fiction: several central banks are experimenting with internal proof-of-concepts, driven by the need for absolute confidentiality.

Quantization, a technique that shrinks models down to INT8 or FP16, is essential to maintain acceptable performance without requiring huge data centers. Still, throughput remains a constraint: serving real-time analytics on transaction streams demands software optimization and efficient pipelines. The open ecosystem (vLLM, llama.cpp, Ollama) is rapidly closing the gap, making on-premise not only viable from a security standpoint but also competitive in day-to-day operations.

What Sintra signals for the near future

The fact that central bankers shifted from worrying about inflation to worrying about AI is no mere agenda change. It’s a symptom of a growing awareness: AI is no longer a tool to delegate to IT departments or outsource to Silicon Valley. It’s a geopolitical and regulatory asset. The Sintra discussion, though lacking operational conclusions, marks a turning point: the question is no longer whether to adopt AI, but how to do it while preserving democratic control and systemic resilience.

For those in enterprises or government evaluating adoption paths, the message is clear. Assessments cannot stop at immediate cost convenience. They must factor in vendor lock-in risks, exposure to unilateral changes in terms of service, and the ability to intervene directly on the model in an emergency. AI-RADAR offers analytical frameworks at /llm-onpremise to weigh these variables without succumbing to fleeting trends. The game is open, and it’s played on the terrain of physical infrastructure as much as on pixels.