Russian businesses are rapidly shifting toward Chinese artificial intelligence models, with token usage surging 11-fold in just the first half of 2026 compared to all of 2025. This trend reveals how geopolitical tensions are reshaping the global AI landscape, pushing companies in sanctioned regions toward alternatives that offer both cost savings and the ability to keep sensitive data on domestic servers.

What's Driving Russia's Turn to Chinese AI?

The surge in Chinese model adoption stems from two practical pressures: economics and security. Western sanctions have made U.S.-developed AI tools increasingly difficult to access, while Chinese models offer substantially lower costs and can run entirely on Russian infrastructure. This matters enormously for companies handling sensitive information, as it means data never leaves the country.

The numbers tell a striking story. MTS Web Services Cloud, one of Russia's largest cloud platforms, reported that customers generated 400 billion tokens using Chinese models in the first half of 2026 alone. That's 11 times the total from all of 2025. On Yandex AI Studio, another major platform, consumption of Chinese models increased more than fivefold over the past year.

Which Chinese Models Are Winning in Russia?

Three Chinese models dominate Russian cloud platforms, each capturing significant market share:

• Qwen: The clear leader, accounting for 261.1 billion tokens on MTS Web Services and holding a 21.2% share of token consumption on Yandex AI Studio
• GLM: Generated 91.2 billion tokens on MTS Web Services, making it the third most-used Chinese model across major platforms
• Kimi: Produced 81.4 billion tokens on MTS Web Services, while DeepSeek ranked second on Yandex with a 13.5% share

For comparison, U.S.-developed GPT-OSS models accounted for only 7.5% of token consumption on Yandex AI Studio, highlighting how dramatically the competitive landscape has shifted.

How Are Russian Companies Deploying These Models?

The deployment strategy matters as much as the model choice itself. Rather than accessing Chinese AI through cloud services hosted abroad, Russian companies are installing these models directly on their own servers. This approach, called on-premises deployment, keeps all data within company systems and eliminates reliance on external infrastructure.

"The main format is deploying the model on Russian infrastructure. In this format, the data remain within the company's systems, and the choice of model is purely technical and does not depend on security requirements," said Alexander Tugov, director of Selectel's AI division.

Alexander Tugov, Director of AI Division at Selectel

This arrangement is particularly valuable for enterprises handling proprietary information, financial data, or other sensitive materials. Dmitry Cheklov, founder of the Hybrid advertising technology ecosystem, noted that many businesses were guided by operating costs, where Chinese models offered a clear advantage.

What Does This Mean for Russia's "Sovereign AI" Ambitions?

Russia has been pursuing what officials call "sovereign AI," a domestic industry built on models developed entirely within the country. Under proposed rules, Russian-made models would be assessed for compliance with the country's officially defined "traditional spiritual and moral values." Only models passing this review would qualify for government contracts and access to state data systems.

Under

However, experts are skeptical about Russia's ability to compete. Natalya Kaspersky, head of cybersecurity company InfoWatch, has emphasized the scale challenge involved in building large foundation models from scratch.

"When you are walking along the same tracks on which trains are traveling, you have essentially no chance of catching up with that train. I do not think we have any realistic chance of creating something groundbreaking in large foundation models," said Natalya Kaspersky.

Natalya Kaspersky, Head of InfoWatch

Building competitive large language models requires vast financial resources, computing power, and energy infrastructure that Russia currently lacks. The irony is stark: while Moscow pursues sovereign AI, Russian businesses are increasingly dependent on Chinese models to fill the gap.

How Is DeepSeek Expanding Beyond Models?

Meanwhile, Chinese AI developers are moving beyond just offering cheaper models. DeepSeek launched DeepSeek Harness v0.1, an open-source agent framework that competes directly with Anthropic's Claude Code and OpenAI's Codex. This represents a significant strategic shift, as DeepSeek is now offering not just models but the entire software infrastructure developers need to build AI agents.

DeepSeek Harness is built on a modular architecture where practically every component can be replaced or customized. This includes the model itself, which is treated as just another plugin rather than the fixed center of a vertically integrated system. The framework launched under the MIT open-source license, making it freely available for developers to download and modify.

The capabilities are substantial. Developers using Harness can inspect code repositories, edit files, execute shell commands, search files and the web, maintain plans, invoke skills, delegate work to subagents, and enforce approval policies. These are the core features that make modern coding agents useful for real development work.

What's Changing in DeepSeek's Pricing?

DeepSeek is also restructuring its API pricing model. Beginning August 16, 2026, the company is abandoning flat-rate pricing in favor of peak and off-peak rates. Even the discounted off-peak prices will be substantially higher than current rates, signaling that the company is moving toward a more sophisticated pricing strategy as it expands its product offerings.

The official version of DeepSeek-V4-Pro, released alongside Harness, now features significantly enhanced agent capabilities and support for OpenAI's Responses API and Codex integration. The model is available across DeepSeek's web interface, mobile app, and API, with developers able to control reasoning effort across three levels: Non-think for fast tasks, Think High for complex problem-solving, and Think Max for difficult problems requiring substantial reasoning.

Together, these developments show Chinese AI companies moving aggressively into the developer tooling layer, not just competing on model quality or price. For Russian businesses facing Western sanctions and for developers worldwide seeking alternatives to U.S.-dominated AI infrastructure, these moves represent a meaningful shift in how AI systems are built and deployed.