Petrobras and OpenAI: the right question isn't whether to sign

The wrongly framed question

Imagine the following scenario, purely hypothetical: a state-owned oil company like Petrobras is evaluating a modernization agreement with a leading artificial intelligence provider like OpenAI. For the argument that follows, it does not matter whether this is already happening somewhere in the region, whether it is under negotiation, or whether it is entirely speculative — I use it as a device for reasoning about a pattern that will repeat, under these names or others, at every state-owned energy company in the region over the coming years. My interest is not in confirming or denying a specific fact. It is in showing why the public debate surrounding any agreement of this kind is usually mis-framed from the very first question.

The question that dominates the coverage and the legislative discussion is binary: is it dangerous for a strategic state-owned company to hand its technological modernization to a foreign AI provider, yes or no? It is the wrong question, and it is wrong for a structural reason, not an ideological one. It treats "signing with OpenAI" as a single, indivisible event, when in reality an agreement of this kind is a bundle of dozens of separable decisions, each with its own level of risk, its own reversibility, and its own jurisdiction. Asking "yes or no?" forces an answer about the entire bundle in a single word, which guarantees a bad answer — either a refusal that forfeits real operational gains, or an acceptance that hands over more than needed to be handed over.

The error is not exclusive to the press or to legislators. Companies themselves fall into it when they structure the negotiation as a single purchasing decision instead of as a layered architecture. A board that approves or rejects "the OpenAI deal" in a single vote is, without meaning to, accepting that everything that provider touches falls under the same contractual regime — the same level of access, the same data clause, the same ambiguity about what happens if the relationship ends. That uniformity is the real risk, not the provider's origin.

The right questions

The Sovereign Architecture framework starts from a distinction that is not exclusive to energy but becomes especially sharp in state-owned energy: every modernization operates on a decision layer and an execution layer, and the error of the public debate is treating them as if they were one thing. In a state-owned oil company, the decision layer includes which fields are explored, which reserves are declared, which seismic and geological data are considered strategic, and who has the authority to decide on national energy security. The execution layer includes processing that same seismic data with machine learning models, predictive maintenance of pumps and compressors, optimization of logistics routes, early fault detection in refineries. These are technically adjacent and legally distinct domains, and that distinction is exactly what a well-structured agreement has to preserve and what a headline cannot convey.

The correct operational question is not "should the state sign?" It is a series of granular questions that any technical and legal team can answer before signing a single line. What specific data will leave the national infrastructure to be processed, and what stays on servers under domestic jurisdiction? Does the model that optimizes a pump's maintenance need to see the same data as the model that interprets a seismic profile of an undeclared field, or are these separable workloads with distinct access perimeters? Under what jurisdiction do the audit logs of each query live — can they be subpoenaed by a foreign court, or are they subject exclusively to the law of the country operating the field? What reversibility clause exists if the provider changes its terms, unilaterally raises the price, or the political relationship between the two countries deteriorates?

None of these questions is answered with a yes or no to the whole agreement. Predictive maintenance of a pump is, in the vast majority of cases, pure execution: delegable, reversible, with no bearing on sovereignty over the resource. If the model fails, it gets replaced; if the provider raises the price, you migrate to another one; losing that relationship compromises no strategic asset of the country. Processing seismic data from an unexplored basin is a different case — not because the technology differs, but because the data itself is the strategic asset, and once it leaves a controlled perimeter, no contractual clause can truly guarantee it returns to exclusive control. The correct architecture separates these two cases in the technical design of the agreement; it does not blend them under the generic label of "AI modernization."

The capacity to negotiate the distinction

Framing the distinction correctly is not worth much if the state-owned company does not, in practice, have the capacity to impose it against a provider with more technical and commercial negotiating power than almost any state counterpart in the region. The question that follows, then, is not normative but operational: what would actually give a state-owned oil company or any government the real capacity to negotiate this distinction without depending on the provider's goodwill?

Three concrete contractual elements, verifiable before signing, make that difference. The first is an explicit reversibility clause: the demonstrated — not merely promised — capacity to export one's own data and discontinue the service within a defined period, with the provider obligated to certify the destruction of any copies it retains. Without that clause, any theoretical separation between decision layer and execution layer is paper, because the real cost of exiting the relationship ends up so high that reversibility is never exercised. The second is retention of copies of the models fine-tuned with proprietary data: if a model is trained or adjusted using the country's geological or production data, that adjusted model — not just the raw data — should remain under custody or joint control, so that the knowledge accumulated in the tuning does not disappear or end up exclusively on the provider's side once the contract ends. The third is independent auditing: a mechanism, run by a third party with no commercial interest in the relationship, that periodically verifies what data actually crossed each access perimeter — not what the contract says should cross it.

These three elements are falsifiable in the most literal sense: they can be verified in the text of the contract, not in either party's stated intentions. An agreement that includes them allows operational modernization to proceed aggressively — predictive maintenance, energy efficiency, accelerated processing — without that implying ceding the layer where sovereignty over the resource is actually at stake. An agreement that does not include them is indistinguishable, in its real risk exposure, from the binary warning the public debate already issues about "handing the data to a foreign company" — only with the added varnish of appearing to be a carefully structured technical decision. The difference between modernizing and surrendering control is not in whether you sign. It is in what you sign, clause by clause, and in whether anyone on the state side had the technical capacity to read those clauses before they turned into the wrong question on a front page.