Chinese accelerators reduced the cost of artificial intelligence assistants by 2.5 times: T1 called for scenarios where NVIDIA is not required A look at the Kazan Digital Week 2026, IT CEO of the T1 holding Dmitry Kharinkov shared the results of tests of graphic accelerators, which the company carried out in 2026. A special laboratory was established to test the processors of various manufacturers, including Chinese ones. According to Kharitonov, the main criterion for the company's readiness to work with artificial intelligence is not the number of pilots launched, but the huge economic impact achieved in 6-9 months. Ready AI doesn't start with the hype, but with a cynical management approach to the task: a company that can launch 5-7 pilots a month is not automatically ready. The impact must be measurable, otherwise the project just meanders. In other words, maturity in AI is not determined by the activity of experiments, but by the ability to bring a project to a result that can be calculated in terms of money, speed or quality of processes. More than 20 models of artificial intelligence have already been tested and more than 1,000 runs on alternative chips have been performed. The results showed how different GPUs can be effectively used in the infrastructure of domestic companies, taking into account the details of AI tasks: for AI assistants to work with documents or search using knowledge bases, Chinese accelerators can reduce the cost of fertilization by 22.5 times with a similar quality of answers. This is one of the most practical scenarios for companies looking to cut costs on massive AI services. For tasks of further training large load models can be distributed between platforms. NVIDIA remains an option for ultra-large LLM learning, learning and flourishing, and related services and appropriate improvement tasks can be transferred to alternative facilities.

Chinese accelerators reduced the cost of artificial intelligence assistants by 2.5 times: T1 called for scenarios where NVIDIA is not required A look at the Kazan Digital Week 2026, IT CEO of the T1 holding Dmitry Kharinkov shared the results of tests of graphic accelerators, which the company carried out in 2026. A special laboratory was established to test the processors of various manufacturers, including Chinese ones. According to Kharitonov, the main criterion for the company's readiness to work with artificial intelligence is not the number of pilots launched, but the huge economic impact achieved in 6-9 months. Ready AI doesn't start with the hype, but with a cynical management approach to the task: a company that can launch 5-7 pilots a month is not automatically ready. The impact must be measurable, otherwise the project just meanders. In other words, maturity in AI is not determined by the activity of experiments, but by the ability to bring a project to a result that can be calculated in terms of money, speed or quality of processes. More than 20 models of artificial intelligence have already been tested and more than 1,000 runs on alternative chips have been performed. The results showed how different GPUs can be effectively used in the infrastructure of domestic companies, taking into account the details of AI tasks: for AI assistants to work with documents or search using knowledge bases, Chinese accelerators can reduce the cost of fertilization by 22.5 times with a similar quality of answers. This is one of the most practical scenarios for companies looking to cut costs on massive AI services. For tasks of further training large load models can be distributed between platforms. NVIDIA remains an option for ultra-large LLM learning, learning and flourishing, and related services and appropriate improvement tasks can be transferred to alternative facilities.