Ten years ago, Karina Portugal was discussing campaigns and branding at advertising festivals, from Cannes to São Paulo. Today, she sits across the table from some of the largest banks in Brazil, Europe, and the United States, tackling a far less glamorous and much more concrete problem: how do you know whether you can trust who, or what, is making decisions worth millions, balance sheets, and reputations? Her trajectory is a barometer of a seasonal shift. Marketing could afford AI as smoke and mirrors; banking cannot. And the question Portugal is teaching people to ask is not whether a Large Language Model is accurate or fast, but whether it is reliable in a context where an error is not a misplaced banner but a regulatory hallucination.

The core of the issue is not technical, at least not in the traditional sense. Language models are capable of probabilistic reasoning at human scale, but they lack guarantees of determinism, explainability, and data confinement. For a bank, this is an existential short circuit. Regulators demand traceability, auditability, and control over sensitive data. A public cloud deployment, with models managed by third parties, introduces a sovereignty risk that is hard to accept: where do prompts end up? Who accesses logs? Can you prove that the model has not incorporated biases prohibited by anti-money laundering or privacy regulations?

The right question that Portugal poses to executives is: “What is your tolerance for an error you cannot explain?” This shifts the center of gravity from efficiency to governance. And here a structural consideration kicks in: if the answer is “none,” then multi-tenant cloud is not the path. Direct control over inference is needed, often on-premise, in air-gapped environments or on hybrid infrastructure where data never leaves the corporate perimeter. This is not ideological; it is the concrete consequence of a context where Total Cost of Ownership is also measured in fines, reputational damage, and loss of a banking license.

Who wins and who loses, then? Big tech pushing cloud APIs lose ground the moment a regulated customer takes Portugal’s question seriously. Conversely, those providing hardware for local inference, frameworks for on-premise fine-tuning, and orchestration solutions that enable full auditing gain strategic relevance. It is no coincidence that banks are exploring architectures based on servers with high-memory GPUs, models quantized to INT8 or FP16 to stay within accessible VRAM limits while running inference without ever exposing data to external services. Quantization is not just a resource-saving trick: it becomes an enabler of sovereignty.

The lesson Portugal embodies is that on-premise deployment is not a nostalgic retreat to physical data centers, but the precondition for asking the right questions. When you own the infrastructure, you can choose what to log, what to trace, what to demonstrate to an auditor. You can fine-tune on proprietary data without sharing it with a provider. And you can decide when a model is reliable enough to go into production, without delegating that assessment to an external vendor. It is the overturning of the “API-first” paradigm: governance first, then architecture.

Unsurprisingly, in the AI ecosystem, figures like Portugal are becoming central. The right question is not formulated by a prompt engineer, but by those who understand the difference between winning a Lion at Cannes and losing the trust of a central bank.