Ukraine to Favor AI Models Free From Provider Control

Ukraine has announced that from now on, when choosing artificial intelligence tools for public administration, business and the military, it will favor models that can be run on its own servers, independent of any foreign provider. As Reuters reported, the decision was confirmed by Roman Kyslyi, chief AI specialist at Ukraine's Ministry of Digital Transformation.

Currently, the AI assistant built into the Diia app, used by more than 23 million Ukrainians, runs on Google's Gemini model, accessed remotely via servers located in the European Union. Ukrainian authorities strip personal data from queries before they reach Google, since, as Kyslyi noted, they have no control over the model itself.

A spark from Washington

The direct trigger for the shift was the US administration's decision ordering Anthropic to cut off access to its two most powerful models, Mythos 5 and Claude Fable 5, for all users outside the United States, citing national security concerns. Anthropic had to fully disable both models for customers outside the US in order to comply with the order.

For Kyiv, this was a warning sign that even a key ally could cut off access overnight to technology underpinning public services and military systems. Ukraine, fighting a defensive war and heavily dependent on the continuity of digital administration, considered this an argument too important to ignore.

This confirms that AI sovereignty is not just a defense slogan, it's a necessity - Roman Kyslyi, chief AI specialist, Ukraine's Ministry of Digital Transformation

Gemini as a stopgap solution

Kyslyi described the current use of Gemini in Diia as a transitional solution, not a permanent one. Google provides Ukraine with free tokens for the model, which lowers costs, but the lack of control over the underlying infrastructure remains a problem in a country that treats the continuity of digital services as a matter of national security.

The intended long-term solution is Ukraine's own model, built together with telecom operator Kyivstar on top of the open version of Google Gemma. The launch is expected in autumn, with the model set to reach public administration, private companies and military applications simultaneously. Before settling on Gemma, Ukrainian specialists also tested Mistral's models and OpenAI's open GPT-OSS, but Google's solution offered the best balance of performance and local deployment capability.

A model as commodity, not service

Kyslyi stressed that from Kyiv's perspective, the language model itself is, in practice, an interchangeable technological component rather than a service the country needs to depend on. If a provider agrees to make its model available for deployment on infrastructure controlled by Ukraine, there is no obstacle to using it further, regardless of whether it comes from Anthropic, OpenAI or Google.

This approach differs from the practice followed by most governments, which use public cloud models without major reservations. Ukraine, given its state of war, treats dependence on a foreign provider as a real operational risk, comparable to dependence on energy or weapons supplies.

Implications for Europe

Kyiv's decision fits into a broader European debate about technological sovereignty in AI, also present in discussions around the EU's AI Act and national projects such as Poland's Bielik. For Polish public institutions and companies considering AI deployments, it is another signal that relying solely on models remotely controlled by a single foreign provider carries political risk, not just technical or financial risk.

The Ukrainian model based on Gemma is still under development, and its actual performance compared with commercial closed models will only be known after the autumn launch. Kyslyi did not disclose details on the budget allocated to the project or the number of people working on it with Kyivstar.

Sources: Ukraine to Pick AI Models Operated Without Provider Control, Official Says (usnews.com), Ukraine to pick AI models operated without provider control, official says (finance.yahoo.com), US export ban on Anthropic's AI models further strains alliances (aljazeera.com)