AI sovereignty starts with the right to leave

Customising a model can be one of the fastest ways to become more dependent on its supplier. It can turn company data and engineering effort into an asset that works only with a particular base model, licence and serving environment. That is a poor definition of sovereignty.

The AI Daily Brief argues that recent Microsoft and Meta model announcements answer the enterprise questions of cost and sovereignty. That framing is commercially useful, provided sovereignty receives a harder test than private data handling or access to model weights. Cost concerns what the system consumes while it works; sovereignty concerns what the company can do when the relationship no longer works.

Microsoft says that for models sold by Azure, customer prompts, completions and training data are not available to other customers or model providers, and fine-tuned models are exclusively available to the customer. Those protections matter because they address confidentiality, access and permitted use. A customer can receive every one of those protections and remain unable to run its customised system elsewhere. Microsoft’s fine-tuning instructions require a supported base model and direct deployment through the Foundry portal.

Neither privacy nor exclusivity establishes portability. Portability asks whether the customer can take the artefact it funded, combine it lawfully with everything needed to make it function, and serve it on alternative infrastructure within an acceptable time and budget. The asset therefore includes more than a training dataset or an adapter file.

For parameter-efficient fine-tuning, Hugging Face’s documentation explains that loading the resulting model also requires the original base model. This dependence creates a commercial chain: the adapter depends on the base, the base depends on its licence, and practical operation depends on compatible serving software and hardware. If any link cannot move, ownership of the smallest file offers limited bargaining power.

Open weights can improve the position without settling it. Meta’s Llama 4 licence grants rights to modify and create derivative works, yet it also says a breaching licensee must delete and cease using the materials after termination. The same distinction applies to any model carrying contractual conditions: possessing weights does not erase the licence governing their use. A company evaluating an open-weight model should therefore assess the rights, dependencies and operating burden together.

This is where a sovereignty claim becomes a procurement question rather than a slogan. Suppose a company spends a year refining a model for quoting, claims assessment or technical support. If the base model is withdrawn, repriced, materially changed or restricted, can the company reproduce the service somewhere else without rebuilding the customised behaviour from scratch? The migration cost revealed by that exercise is a better measure of dependence than the location of the training data alone.

True independence may be uneconomic for a given workload. Buyers do not need to demand self-hosting in every contract, but they should know what they are surrendering in exchange for a managed service. First, define the exportable artefact before training begins. Require clarity on whether the customer receives full model weights, an adapter, checkpoints, configuration files, evaluation sets and enough documentation to reconstruct the system.

Second, record the exact base-model version and every material dependency used in production. A reproducible build has more exit value than an undocumented endpoint whose behaviour changes with the provider’s release cycle. Third, test the legal path as carefully as the technical path. Include rights to retain, modify and serve the customised artefact after the commercial relationship ends, subject to legitimate intellectual property and safety constraints.

Fourth, price an exit rehearsal into the project. A cheaper model can become expensive if migration requires fresh training, new evaluations and the redesign of surrounding workflows. Governance committees should treat that contingent liability like any other concentration risk.

If a provider will not support meaningful portability, the customer can still proceed with clear eyes and negotiate accordingly. The board should then understand that it is buying a service with switching costs, rather than building a sovereign intelligence asset. Vendors deserve credit for stronger privacy terms, fine-tuning options and accessible weights. That answer should influence procurement, but it should not end the inquiry.

Sovereignty is not a feeling produced by a private endpoint or a download button. It exists when the company has a credible, lawful and affordable way to leave.