The Russian high-performance computing market continues to adapt to new geopolitical realities. The focus is on the successful testing of a hardware-software system built on eight Chinese graphics accelerators. This step is not just a search for a replacement, but part of a systemic strategy to reduce infrastructure risks, which, by all appearances, is becoming the new standard for domestic technology giants.
Test details and initial results
As part of the pilot project, two artificial intelligence models from MWS AI were launched: the larger Cotype Pro 3 with 27 billion parameters and the lighter Cotype Light 3 with 9 billion parameters. Both models are designed for working with text and graphics, as well as for creating AI agents. Notably, the developer considered five different options for Chinese accelerators, but the specific chip used in the final configuration has not been disclosed.
Key performance metrics inspire optimism. With an input context volume of about 27,000 tokens, the system generated the first response in approximately 8 seconds, and subsequent tokens were produced at intervals from 111 ms, corresponding to speeds of up to 9 tokens per second. In certain scenarios, according to integrator Rubytech, performance was comparable to systems based on Nvidia H100. Moreover, optimization of drivers and the software environment made it possible to increase platform performance by 2–2.2 times.
Cautious optimism and ecosystem challenges
Despite the encouraging results, full-scale commercial deployment of such systems is not expected before 2027. And here, the key factor is not the hardware, but the software ecosystem. Transitioning from the Nvidia CUDA architecture to alternative platforms is a process that can take several months even for a single model, let alone for mass adoption.
My analysis shows that we are witnessing the formation of a multipolar accelerator market. Major players such as VTB and Sberbank are at different stages of mastering Chinese GPUs, but no one expects a quick abandonment of Nvidia. Tests show that claimed specifications are not always confirmed in practice: for example, the Tongxin chip from Shanghai Wutongshu High-Tech turned out to be slower than both the Nvidia A100 and A40 due to compatibility issues and immature software. At the same time, more advanced solutions, including Huawei Ascend, already achieve 60–80% of H100 performance.
The state is also joining the process: the Ministry of Industry and Trade is discussing localization requirements for AI chips, which could give preferences to domestic manufacturers in government procurement. However, in my assessment, it will take several more years before a mature and competitive market takes shape. For now, Russian companies are forced to balance between the need for diversification and the practical limitations of existing software.