Sberbank: Russia needs sovereign AI, but controlled foreign data remains necessary

Sberbank believes that Russia needs sovereign artificial intelligence to avoid critical dependence on foreign models and services. At the same time, a completely isolated strategy in 2026 appears too expensive and unrealistic: without limited access to foreign data, building a high-quality model is difficult.

Why Control is Needed

Sberbank's logic is simple: if a key model, API, or infrastructure component is controlled from outside, access rules can change at any moment. For banks, the public sector, and large corporations, this is not an abstract risk, but a direct operational issue. Today a model is available, tomorrow licenses, export restrictions, tariffs, or terms of use change, and entire processes become dependent on others' decisions.

Against the backdrop of such uncertainty, the idea of sovereign AI transforms from a political slogan into a task of technological resilience. By sovereignty in this context is meant not simply running a local chatbot, but control over the entire chain: computing, data, fine-tuning, security, and deployment rules. For companies in regulated industries, this is especially important because they work with sensitive information and cannot build long-term products on services whose access cannot be guaranteed.

Therefore, betting on proprietary models and platforms looks like insurance against external disconnection and against imposed restrictions.

Cost of Building Your Own Model

But such a course comes at a high price. Training a strong base model from scratch in 2026 is no longer a story about one research team and a couple of lucky experiments. You need large computational resources, stable equipment supplies, long tuning cycles, teams of engineers, and months—sometimes years—of iteration. Even if the money is there, it's still a slow path: quality doesn't come immediately, and mistakes in early stages are very expensive. In practice, this means several major investments at once:

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

This is precisely why Sberbank's thesis sounds pragmatic rather than maximalist. It's not about isolation at any cost delivering the best result, but about the fact that full technological sovereignty today is too expensive if built in isolation from the global knowledge base. You can create your own system and your own model, but this doesn't negate the fundamental problem: competitive quality requires scale, time, and access to diverse training material.

Compromise on Data

Here appears the second part of the position: completely rejecting foreign data is in practice impossible. Modern models learn on enormous arrays of texts, code, scientific publications, technical documentation, and multilingual content, a significant portion of which was created outside Russia. If this layer is artificially cut off, the model quickly hits gaps: it will understand international context worse, work more weakly with code, lose accuracy in science, finance, and engineering tasks.

Sberbank identifies the optimal strategy as a mixed approach: reliance on domestic developments plus limited and controlled use of foreign datasets. The key word here is control. That is, not uncontrolled connection to external services, but clear rules for data selection, local storage, filtering, verification of usage rights, and the ability to continue work within your own system at any moment.

This approach reduces dependence but doesn't cut quality where the global data corpus is still necessary.

What This Means

For the Russian AI market, this is a signal that the debate is shifting from the slogan "ours or theirs" to a more practical formula. Winners will be those who can assemble local infrastructure and a model stack but at the same time carefully use the global data array without direct dependence on foreign platforms. Otherwise, either quality will be weak or the disconnection risk will remain too high.