On sovereign AI

 November 9, 2026

Although AI has become a very mainstream topic in the last three years, the concept of sovereign AI has not reached the mainstream yet. The problem isn’t only its low profile, but also that it is often used as a sovereignty theater rather than real sovereignty. The first thing to realize is that sovereignty is not a boolean characteristic; it is a multidimensional spectrum, and any conversation that treats it as boolean fails to be informative. Although the exact number of dimensions is arbitrary, and different people might consider different ones, I prefer a three-dimensional model.

The first dimension is deployment sovereignty: the organization’s ability to deploy an AI system for whatever it deems valuable. This obviously has to adhere to the legal framework where the organization operates. Another aspect to consider is whether a possible implementation option tangles the organization in more legal frameworks than required. If, for instance, the organization is under the European legal framework, any solution that adds additional legal frameworks (such as the Chinese one, the US one, or any other) should be viewed with skepticism. In fact, even if, when deployed, the frameworks are very similar, over time they might diverge and, since the legally valid uses will need to be compliant with both legal frameworks, their number will decrease as the two frameworks diverge more, with potentially a point where there is no overlap of the two groups, making it impossible to use it for any use case. This can further hinder the business if very different legal frameworks are chosen, or if the selected solution implies that the usage must comply with three or more legal frameworks. National legal frameworks are not the only relevant ones; in fact, if other companies are required to deploy an AI system, their policies will also need to be considered as additional contractual or organizational constraints. So, the more companies and the more dispersed they are, the more legally bound the implementation will be, and therefore the less sovereign. Remember that deploying an AI system requires more than models and hardware; it also requires people and processes, which should be considered in the same way. Another aspect of this dimension is the freedom the organization has to optimize models for its needs and understand output quality. Closed models generally provide less freedom to modify their weights and evaluate their behavior. Open-weight models can provide more freedom for both activities, depending on their license and available artifacts.

The second dimension is inference sovereignty: an organization’s ability to perform inference at will on the deployed AI system. This aspect focuses on the ability to continue using the deployed system and the limitations that may be encountered. All aspects of the previous dimension apply here as well, but failures in this dimension are more immediately disruptive. If deploying an AI system becomes impractical or impossible, it will inhibit future use cases; if inference becomes problematic, current systems are hindered, and the consequences might be huge for the organization. In this dimension, it is also important to understand the model’s pricing because, as I discussed previously here, it has huge implications for the solution’s sustainability. It is important to consider both the companies and countries involved in the provisioning of AI capabilities and in the inference itself, if they differ. Although it could seem theoretically nice to have zero dependencies (in addition to the organization’s national legal framework) in either of the first two dimensions, this is not reasonably achievable, since it would require producing within the company the design for the hardware, the silicon, the AI models, the AI tooling, and everything in between, and that is neither reasonable nor desirable. Also, it is critical to understand the providers’ business model to evaluate their probable future behaviors.

The third dimension is data sovereignty. This is often the most considered one, since many legal frameworks are starting to take a stance on this aspect, though often more from a compliance perspective than a real risk analysis. Often the solution to every problem in this dimension is to sign a deal with third parties that grants certain assurances on the matter. The problem is that once data leave your premises and other parties have access to them, many things can happen that put your data at risk. A third party can change the contract unilaterally, violate it, be forced by its regulator to share data, or suffer a breach that leaks data. Also, third parties often share data with other parties under similar legal contracts that carry the same issues and risks as the initial contract. It is therefore critical to understand exactly what data will be processed, the full list of third parties that will have access, and how that list can change through a decision by any involved party or compulsory action.

I think it is critical to have a serious conversation about sovereign AI and how sovereign the various options are, since decisions made without clear understanding can have unforeseen consequences. Today, many companies are making decisions about AI systems more on vibes and FOMO than on risk and potential benefits.