OpenAI, Qwen and GigaChat: Why Russian Businesses Are Finding It Harder to Pick an AI Model

The Russian AI market is entering a phase where choosing a model is no longer just a question of answer quality. For companies, it now combines three factors: the availability of Western services, data requirements and the cost of local infrastructure.

How the choice is narrowing

The author describes a situation in which Western models such as those from OpenAI and Anthropic are becoming less and less available to Russian businesses, not only technically but also legally. Country and IP blocks already operate at some vendors, and in regulated industries even formally permissible access through a proxy solves little. If a request to an external API contains a client's name, phone number or voice, it starts to look like a cross-border transfer of personal data and runs into the requirements of Federal Law 152-FZ. As a result, AI agents for support, sales and contact centers fall into a higher-risk zone: what passes through the model is not abstract text but real user data.

Against this backdrop, demand for "sovereign" solutions is growing inside Russia, but the word often covers not homegrown models but adapted versions of foreign open-source systems. This is where the main trade-off begins: the greater the formal independence, the heavier the economics.

Three working scenarios

Three approaches have effectively taken shape on the market. The first is to build a base model from scratch, as Sber does with its GigaChat family. The second is to take a strong open model, most often from the Qwen family, and fine-tune it on a Russian corpus and domain data, as Yandex, T-Bank and Avito do. The third is to keep using Western APIs through a gray zone, if the business is ready to accept the legal risk.

• GigaChat: maximum control and locality, but very expensive training and inference.
• Qwen after fine-tuning: noticeably cheaper and faster to launch, but sovereignty is nominal.
• OpenAI and Anthropic: strong quality and clear economics, but access is becoming less and less stable.
• Hybrid setups: a compromise for mid-sized businesses, starting in the cloud and later moving to their own environment.

The problem is that every path carries a cost that cannot be ignored. Training from scratch requires tens or even hundreds of millions of dollars, large volumes of data and scarce GPUs of the H100 or H200 class. Fine-tuning Qwen looks more realistic, but the underlying architecture and weights remain Chinese. From the standpoint of strict regulatory logic, this is not full independence but a carefully localized compromise.

Where the money goes

The most painful argument in the article is not model quality but the price of inference. According to the author's calculations on his own agent platform, a minute of work with a comparable OpenAI model costs less than 1 ruble, while a minute on GigaChat-Max costs about 80 rubles. For voice agents and contact centers, that is a difference not of percentages but of almost two orders of magnitude. With this cost structure, you can build a technically good product but not make it work economically.

"A fully Russian solution is insanely expensive"

Infrastructure delivers another blow. The author estimates a server able to handle about a thousand concurrent agent sessions at roughly 55 million rubles. Then comes another trap: to keep the cost per token relatively low, GPUs must run at 80-90% utilization. That is hard with small, uneven demand. The hardware sits idle, while electricity, maintenance and depreciation costs do not go away. That is why AI pays off first where human labor is expensive and the load is constant: support, contact centers and legal functions.

What this means

For product teams, the conclusion is fairly stark: building the whole architecture on a single provider is already dangerous. A company working with Russian-language LLMs needs a model-agnostic setup that can switch quickly between OpenAI, GigaChat, Qwen-like solutions and a local environment. Otherwise, any new round of blocks, price changes or data requirements quickly turns a technical choice into a business problem.