The deployment of large language models in Russia has significantly increased in cost over the past year. According to MWS Cloud, the average cost of a minimum configuration on Nvidia accelerators for launching one LLM grew from 13.7 million rubles in 2025 to 38.8 million rubles in 2026 — a 2.8-fold increase.
The study compared seven models from 2026 with six solutions introduced a year earlier. Calculations took into account the minimum number of GPUs required to handle a request of the maximum declared volume, as well as servers and network equipment.
The most expensive model to deploy in 2025 was Kimi K2 — about 35 million rubles for eight H200s. In 2026, Kimi K3 became the leader: eight B300 accelerators will cost approximately 80 million rubles. Qwen3.5, DeepSeek-V4-Pro, and GLM-5.2 each require eight H200s, with the cost of such a configuration estimated at 55 million rubles.
The increase in costs is associated with the growing size and capabilities of new models. To process large volumes of data, they require more video memory and powerful accelerators, which are also becoming more expensive due to high global demand.
MWS Cloud also studied the use of Chinese Huawei Ascend 910B. The average number of such accelerators for supported models increased from 10.7 to 13.6 over the year. However, there is no precise official price for them, and the estimated cost is 50–67% of comparable Nvidia solutions.