Совкомбанк предупредил: закрытие китайских ИИ-моделей ударит по российской отрасли

Analysts at Sovkombank have warned that if Chinese developers close access to their AI models, the Russian industry will face rising costs and significant slowdown in development pace. This is reported by Forbes citing the bank's analytical research. Creating a "sovereign AI of world-class" from scratch will require a strategic overhaul — difficult and expensive.

Why This Matters Right Now

Russian companies in the artificial intelligence sector actively use open Chinese models — primarily DeepSeek and Qwen — as the foundation for their own products. This approach allows savings on training and computational resources, accelerating solution launches to market and dramatically lowering the barrier to entry for startups. Sanctions restrictions closed Russian businesses' direct access to leading American models through commercial APIs, so open Chinese solutions became de facto infrastructure for a significant portion of the AI market.

Over the past year, DeepSeek transformed the competitive landscape: the company released models comparable in quality to GPT-4 while requiring significantly fewer computational resources — and opened weight files for free download. This enabled hundreds of Russian teams to quickly create competitive products on powerful architecture without needing to train base models independently. The closure scenario looks realistic: as Chinese developers strengthen their commercial positions, business logic may override principles of openness.

A similar path was already taken by OpenAI — initially a non-profit organization with open code, today managing strictly proprietary models accessible only through paid APIs.

What Russian Industry Will Lose

Analysts identify several key risks for Russian AI companies:

• Sharp increase in costs for training models from scratch — running into tens and hundreds of millions of dollars
• Development slowdown due to lack of ready weights, architectures, and public research
• GPU infrastructure deficit critical for training large language models
• Accumulating technological gap with global leaders with each quarter of delay
• Need for large-scale hiring of ML specialists — a scarce category in the Russian labor market

According to Sovkombank's assessment, a complete "strategic overhaul" will require not only significant private investment but also systemic solutions at the state level — in education, computational infrastructure, and industry regulation. Without this, Russian companies risk ending up in a situation where neither Chinese nor American models are accessible, and their own base is too weak to compete at global level.

What Options Exist

The Russian AI industry is not standing still. There are homegrown developments: Sber's GigaChat, YandexGPT, Kandinsky for image generation, and a number of specialized models from startups. However, none of them yet reach the level of top Chinese or American competitors in terms of capability-to-cost ratio, speed, and inference cost per token. Analysts see several possible paths. First — accelerated technological partnership with friendly countries where open developments are preserved and scaled. Second — concentration on narrow vertical niches where Russian models can be competitive without matching GPT-5 or Qwen-Max: legal texts, medical documentation, government sector. Third and most ambitious — long-term state investments in computational infrastructure and personnel training with a planning horizon of five to seven years.

"Developing sovereign AI of world-class quality from scratch is a difficult and expensive strategic overhaul," the

Sovkombank materials state.

What This Means

Dependence on open Chinese models is a convenience that may prove temporary. The Russian AI industry should now be budgeting and building competencies for a scenario in which DeepSeek becomes a closed commercial service rather than a free tool. Delay in this step will result not just in additional expenses — it risks locking in technological lag for years to come.