IT holding T1 conducted tests in 2026 to evaluate the performance of various GPU manufacturers, including Chinese alternatives to Nvidia. The study involved testing over 20 AI models and performing more than 1,000 runs to assess the economic viability of diversifying hardware infrastructure.
The findings indicate that using Chinese accelerators for tasks like AI document assistants and knowledge base searches can reduce inference costs by 2 to 2.5 times while maintaining comparable output quality. For computer vision, Chinese solutions proved price-competitive, although they lagged behind Nvidia in processing speed.
T1 suggests a hybrid infrastructure strategy: utilizing Nvidia hardware for training and inference of large LLMs, while offloading secondary services and specific inference tasks to alternative chips. Replacing an underutilized Nvidia H100 with a Chinese C550 model in a cloud environment can save over 2 ₽ million annually.
CEO Dmitry Kharitonov emphasized that AI readiness should be measured by tangible economic results achieved within 6 to 9 months. He also noted that total cost of ownership (TCO) must account for licensing, maintenance, and staff training, especially for critical information infrastructure (KII) requiring secure, domestic hardware.
