Ukraine to Favor On-Premise AI Models the Government Can Actually Control
Ukraine will preferentially adopt AI systems it can run on its own servers, Chief AI Officer Roman Kyslyi told Reuters, citing the US export order that cut off Anthropic access as proof AI sovereignty is a wartime necessity.
Key Takeaways
Ukraine will preferentially adopt on-premise AI models the government can run on its own servers, Chief AI Officer Roman Kyslyi told Reuters, limiting remote-only offerings like those from Anthropic and OpenAI.
The policy hardened after a US June export-control order forced Anthropic to cut Ukraine off from Claude Fable 5 and Mythos 5; those restrictions were only recently lifted.
Ukraine's Diia app assistant currently runs on Google's remote-only Gemini via EU servers as an 'interim' solution, with Google providing free tokens and Ukraine stripping personal data from queries.
Ukraine is building its own LLM with telecom operator Kyivstar, adapted from Google's open-weight Gemma family, with a first beta due in autumn 2026 for government, industry and military use.
Kyslyi said the ordering principle is control, not geography: any vendor willing to provide models for Ukrainian on-premise infrastructure faces no restrictions, calling the model itself 'essentially a commodity.'
Kaan Tınmaz
Ukraine will preferentially adopt artificial intelligence systems it can run on its own servers, a senior official from the Ministry of Digital Transformation told Reuters on Tuesday, framing self-hosted models as a wartime necessity after the United States ordered Anthropic to cut Kyiv off from its most capable systems in June.
The Policy: On-Premise or Nothing
Roman Kyslyi, Chief AI Officer at Ukraine's Ministry of Digital Transformation, said the government will favor "on-premise" models it can deploy under Ukrainian control while limiting solutions that, by design, remain in a vendor's hands — a category that today includes the flagship offerings from Anthropic and OpenAI. "AI sovereignty isn't just a defensive talking point, it's a necessity," Kyslyi told Reuters.
Kyslyi said the ordering principle is not geography but control. "If the vendor will provide it to run on our on-premise infrastructure, there are no restrictions," he said, calling the model itself "essentially a commodity."
Gemini as an "Interim" Bridge
The AI assistant inside Diia, Ukraine's flagship government services app, currently runs on Google's remote-only Gemini through EU-based servers, with Google providing free tokens so no budget is spent. Ukraine strips personal data from queries before sending them to the model, Kyslyi said, describing Gemini as an "interim" solution while a truly sovereign stack is built.
Home-Grown Model With Kyivstar
Ukraine is developing its own large language model with Kyivstar, the country's largest telecom operator, adapted from Google's open-weight Gemma family. A first beta is due in autumn 2026 and is designed to serve government, private industry and the military. The ministry evaluated Mistral models and OpenAI's GPT-OSS before settling on Gemma, saying Gemma and Mistral "matched remote-only alternatives on many performance tests."
Why Now
The policy hardened after Washington ordered Anthropic to disable access to Claude Fable 5 and Mythos 5 in Ukraine under a June export-control action — restrictions that were only recently lifted. Kyivstar's LLM effort mirrors a broader European turn toward controllable stacks; the same day, the European Commission unveiled its own AI-and-cybersecurity action plan emphasizing sovereign access to advanced models. Reuters also reported Tuesday that Chinese authorities are weighing curbs on top open-weight AI models, adding fresh uncertainty to the supply of the very systems Ukraine now wants to run in-house.
Reporting based on coverage from Reuters via Yahoo Finance and Ukraine's Ministry of Digital Transformation.
After the US ordered Anthropic to disable access to Claude Fable 5 and Mythos 5 in Ukraine under a June export-control action, officials concluded that AI sovereignty is a wartime necessity. Chief AI Officer Roman Kyslyi said Ukraine will favor models it can deploy under its own control rather than systems that remain in a vendor's hands.
What AI currently powers Ukraine's Diia government app?
Diia's AI assistant runs on Google's remote-only Gemini through EU-based servers, with Google providing free tokens. Ukraine strips personal data from queries before sending them, and Kyslyi described Gemini as an interim solution while a sovereign stack is built.
What is Ukraine's home-grown AI model project?
Ukraine is developing a large language model with Kyivstar, its largest telecom operator, adapted from Google's open-weight Gemma family. A first beta is due in autumn 2026 and is designed to serve government, private industry and the military. Mistral and OpenAI's GPT-OSS were also evaluated before Gemma was chosen.
Does Ukraine's policy ban foreign AI vendors?
No. Kyslyi said the criterion is control, not geography: if a vendor lets Ukraine run its model on on-premise infrastructure, there are no restrictions. Remote-only offerings, currently including flagship models from Anthropic and OpenAI, are the ones being limited.
