Here’s what’s on our plate today:

• 🧪 Ukraine’s push to build a sovereign AI using Google’s Gemma framework.

• 🗞️ TikTok’s AI hate surge, Amazon’s dub backlash, YouTube kids slop.

• 🧰 Weekend To-Do: AtlasAI, Castor, RunPod for hands-on sovereign stacks.

• 🗳️ Poll: Is AI sovereignty real without control of data/compute/infra?

Let’s dive in. No floaties needed…

The Laboratory

Why Ukraine is building its own AI system with Google

Modern nation-states depend on some foundational systems that keep the country running smoothly. Systems that control the supply of power, water, communications, transportation, finance, and public health, due to their importance to the citizens collectively, make up the critical infrastructure of a country.

To ensure these systems continue operating without any hiccups, nations spend a large amount of their time and energy updating and maintaining them.

In 2025, technological advancements both for military and civilian use have become crucial to the running of a country, making them part of the critical infrastructure. And since AI systems are increasingly being deployed on battlefields for troop coordination and enemy monitoring, countries are now looking to ensure they have control over both the physical and algorithmic sides of the technology.

While the physical side that includes chips, data entry, and electricity grids takes a long time to develop, countries are looking to partner with big tech companies to ensure they have control over the algorithms that will eventually be deployed on these data centers.

In line with this thought process, Ukraine is reportedly working on its own large language model built on Google’s open-weight Gemma technology.

The country’s digital ministry and mobile operator Kyivstar shared that the idea is to develop an independent AI system that can serve both defense and everyday civilian uses, as interest in the technology surges.

To start the process, training will start on Google’s cloud hardware, but the system will eventually run entirely on Ukrainian infrastructure, so the country keeps full ownership of a tool used daily by 23 million citizens.

Why Google’s Gemma?

Ukraine is already using AI in areas like drone coordination, aerial surveillance, and satellite imagery analysis. The country is reportedly relying on the U.S. company Palantir to supply tools that help track Russian attacks and identify disinformation campaigns.

However, this means that it is reliant on a foreign commercial entity for its defense capabilities. One that fails to meet its demands.

According to Oleksandr Bornyakov, the Deputy Minister of Digital Transformation of Ukraine, many systems fail to understand the mixed Ukrainian, Russian, and Bulgarian dialect spoken in his hometown of Bolhrad in Odessa Oblast.

Kyivstar’s chief product officer, Misha Nestor, who is leading the effort, added that existing models often mistranslate legal texts and produce unreliable hallucinations.

The new LLM Ukraine wants to develop will begin training on Google’s infrastructure and later shift to Ukrainian servers, ensuring the country maintains full control over an AI system.

Google’s Gemma framework was chosen after reviewing other options, including Meta’s Llama, France’s Mistral AI, and Chinese models such as DeepSeek and Qwen. These were overlooked as they did not meet the country’s needs.

Part of the reason behind choosing Gemma, as shared by the country's digital ministry, is that the model is one of the most efficient open models by size and quality, with multilingual support, including Ukrainian, already within its capabilities. Additionally, it also has an extended tokenizer supporting up to 128,000 tokens with multimodal text-and-image processing.

It is important to note that language is an important aspect of ensuring modern AI systems can be fine-tuned for military purposes.

However, it is not just Ukraine that is trying to find reliable solutions to these problems.

The global push for sovereign AI

Ukraine’s project fits into a much broader global trend. A CNBC panel in July 2024 noted that open-source AI models are one of the most effective ways for countries to build sovereign AI systems, because they allow nations to innovate independently while still benefiting the wider population.

And countries around the world have taken note. Saudi Arabia has created HUMAIN, a state-backed effort to build Arabic multimodal models.

The UAE is working with Microsoft and aims to become the world’s first fully AI-native government by 2027. Similarly, Singapore, India, and Japan are open-sourcing their local language models, showing that sovereignty does not require doing it alone, but rather giving nations more control and choice.

However, gaining access to models is just part of the story.

According to the World Economic Forum, true AI sovereignty depends on six pillars: strong digital infrastructure, sufficient computing power, solid data governance, ongoing research and development, clear regulations, and a skilled workforce.

So, while countries are rushing to build up the infrastructure, train competent staff, and acquire the technological know-how to build up their AI capabilities, Ukraine, it seems, is looking to save millions by avoiding paying licensing fees to foreign companies, while simultaneously protecting military applications from potential vulnerabilities in external platforms.

The debate over what sovereignty really means

Ukraine’s pitch for a partnership with Google seems like a good move; however, some critics argue that many so-called sovereign AI programs create only the appearance of national control. Governments partner with foreign tech giants for cloud services, chips, or model training, which means the underlying infrastructure remains dependent on external vendors.

Without control over the full AI stack, which includes training data, compute, and deployment, countries risk lock-in, limited transparency, and political vulnerability. In this view, reliance on NVIDIA hardware or Chinese open-source models still amounts to a diluted form of sovereignty.

Others highlight the security risks of open-source models, which are often promoted as the foundation of true sovereign AI. Opening powerful systems to the public can allow malicious actors to remove safety features or customize models for harmful uses.

This has already pushed leading AI companies, including OpenAI and France’s Mistral, to pull back from full openness. Ukraine expects its own model will be attacked immediately after launch, underscoring how open-access systems must grapple with threats like prompt-injection and other adversarial manipulation.

The limits of going in alone

Additionally, there is concern around the resource gap. Training state-of-the-art models requires vast computing power and capital that many nations simply do not possess. While smaller models can be built more cheaply, competing with frontier systems demands massive investment and technical capacity.

This risks widening global inequality as wealthier nations and major tech firms capture most AI-driven gains. And even when countries build models domestically, using datasets influenced by foreign cultural or political norms raises new questions about whether the output can ever be truly sovereign.

Ukraine may be pitching for sovereignty in the AI space with the help of open-source models and Google. However, without the three key elements of data, compute, and connectivity, all of which are crucial for AI services to run smoothly, it runs the risk of turning into what analysts call "managed interdependence” rather than full AI self-sufficiency.

TL;DR

• Ukraine is building its own LLM with Google’s Gemma.

• The goal: full algorithmic and infrastructure control for defense and civilian use.

• Sovereign AI is trending globally, but true autonomy is rare.

• Experts warn of “managed interdependence” instead of real sovereignty.

Headlines You Actually Need

• TikTok’s AI Sludge Problem: Anti-immigrant content generated by AI is racking up billions of views on TikTok, raising major questions about moderation and platform accountability.

• Amazon Quietly Pulls AI-Dubbed Anime: After online backlash, Amazon removed AI-generated English dubs from an anime series, but left viewers wondering why they ever tried it in the first place.

• AI Slop Floods YouTube Kids: Junk educational content created with generative AI is surging on YouTube, exploiting parents, children, and recommendation algorithms alike.

Friday Poll

Weekend To-Do

• AtlasAI: Build and deploy custom AI models using local data.

• Castor: Open-source data cataloging tool to control and audit your datasets.

• RunPod: Spin up private cloud GPUs to train models on your terms.

Rate This Edition