Sberbank: Russia Needs Sovereign AI, but It Can't Do Without Controlled Foreign Data

Sberbank believes Russia needs sovereign artificial intelligence so it does not depend critically on foreign models and services. At the same time, a fully isolated strategy looks too expensive and unrealistic in 2026: without limited access to foreign data, it is hard to build a quality model.

Why control matters

Sberbank's logic is simple: if a key model, API or infrastructure component is controlled from outside, the rules of access can change at any moment. For banks, the public sector and large corporations, this is not an abstract risk but a direct operational question. A model is available today; tomorrow licenses, export restrictions, pricing or terms of use change, and entire processes start to depend on someone else's decisions. Against this uncertainty, the idea of sovereign AI turns from a political slogan into a matter of technological resilience.

Sovereignty here does not mean simply launching a local chatbot. It means control over the whole chain: compute, data, fine-tuning, security and deployment rules. This matters especially for companies in regulated industries, because they handle sensitive information and cannot build long-term products on services whose availability cannot be guaranteed. Betting on in-house models and platforms therefore works as insurance against an external shutoff and imposed restrictions.

The price of an in-house model

But this course comes at a high price. Training a strong foundation model from scratch in 2026 is no longer a story about one research team and a few lucky experiments. It takes large computing capacity, stable hardware supply, long tuning cycles, teams of engineers, and months, sometimes years, of iteration. Even if the money is there, the path is slow: quality does not arrive right away, and mistakes at early stages are very costly.

In practice, this means several heavy investments at once:

• buying and running GPU clusters
• collecting and cleaning large datasets
• teams of ML, data and infrastructure engineers
• legal and security control of data

That is why Sberbank's thesis sounds pragmatic rather than maximalist. The point is not that isolation at any cost will deliver the best result, but that full technological sovereignty is too expensive today if it is built apart from the world's knowledge base. You can create your own perimeter and your own model, but that does not remove the fundamental problem: competitive quality requires scale, time and access to diverse training material.

The data compromise

Here the second part of the position appears: in practice, it is impossible to give up foreign data entirely. Modern models learn from huge volumes of text, code, scientific papers, technical documentation and multilingual content, a significant share of which is created outside Russia. If this layer is cut off artificially, the model quickly runs into gaps: it will understand international context worse, work less well with code, and lose accuracy in science, finance and engineering tasks.

Sberbank calls a mixed approach the optimal strategy: reliance on domestic development plus limited, controlled use of foreign datasets. The key word is control. That means not uncontrolled connection to external services, but clear rules for selecting data, local storage, filtering, checking usage rights, and the ability to keep working inside its own perimeter at any moment. This approach reduces dependence without cutting quality where a global data corpus is still indispensable.

What this means

For the Russian AI market, this signals that the debate is shifting from the slogan "ours or theirs" to a more practical formula. The winners will be those who can build local infrastructure and a model stack while carefully using the world's data without direct dependence on foreign platforms. Otherwise, either quality will be weak or the risk of a cutoff will remain too high.