IT holding T1 tested more than 20 artificial intelligence models on graphics accelerators alternative to Nvidia. The test results showed that switching to Chinese chips in certain scenarios allows reducing inference costs (executing queries on a trained AI model) by 2–2.5 times while maintaining comparable response quality.
T1 CEO Dmitry Kharitonov presented the study findings at Kazan Digital Week 2026. To evaluate the hardware, the company established a dedicated laboratory where it conducted over 1,000 test runs on accelerators from various manufacturers.
Alternative GPUs demonstrated the most noticeable cost savings when running AI assistants handling documents and corporate knowledge bases. In computer vision tasks, Chinese solutions were also cost-competitive, but lagged behind Nvidia in data processing speed.
For training and fine-tuning large language models, T1 suggests combining hardware: using Nvidia for the most resource-intensive processes, while migrating auxiliary services and part of the inference to other platforms.
The company also cited an example of savings: replacing an underutilized Nvidia H100 accelerator with a Chinese C550 cuts costs by more than 2 million rubles a year when deployed in the cloud.
According to Kharitonov, a business’s readiness for AI adoption is determined not by the number of pilot projects, but by measurable economic results that must be achieved within 6–9 months. When evaluating costs, companies should consider not only hardware, but also licenses, maintenance, and employee training.