Ukraine is changing its approach to the implementation of artificial intelligence. During the war, the priority is shifting away from access to the latest models and towards the ability to exercise full control over their operation. This is a trend that may also influence the decisions of governments and companies in other countries.
The Ukrainian Ministry of Digital Transformation has announced that it will give preference to AI models running on its own infrastructure. This means opting for on-premises solutions rather than services operating exclusively on a provider’s cloud. The aim is to minimise the risk that access to key tools might be blocked or restricted by their owners in the future.
This approach was prompted by recent decisions to restrict access to Anthropic’s advanced models. According to Roman Kyslyi, Chief AI Officer at the Ukrainian Ministry of Digital Transformation, these decisions have shown that technological independence is becoming a matter of national security.
As Kyslyi emphasises, the key criterion is not the model’s country of origin, but the ability to run it under the user’s full control. “If the vendor ensures it can be run on our local infrastructure, there are no restrictions.”
Currently, the AI assistant in the government’s Diia app uses the Google Gemini model, which runs on servers within the European Union. Google has provided Ukraine with free tokens, meaning the state incurs no costs for using the service. Nevertheless, personal data is removed before queries are sent, as the model remains under the provider’s control. Kyslyi describes this solution as a temporary measure.
At the same time, Ukraine is developing its own AI model in collaboration with the telecoms operator Kyivstar. It is being built on the basis of Google’s open-source Gemma model and is due to be launched in the autumn. Ultimately, it will be used by public administration, businesses and the military. The Ministry also analysed the Mistral and GPT-OSS models, but opted for Gemma, which achieved comparable results in tests to models available exclusively as a cloud service.
