The demand for computing power consists of two parts: training and reasoning. Assuming that a model is trained from 0, 800 Nvidia A100 cards are required to run for 240 days. The monthly rent for an A100 server is 50,000, and the rent for training this model is 30 million. The link that consumes the most computing power is operation. An Nvidia A100 server can currently handle 20-30 characters per second at an efficient rate. Assuming that there are 100 million daily visits and 10 conversations per person, assuming that the number of replies is 150 words, it will take 1.66 million hours to process it with only one server, so at least 69,400 servers are needed to handle this visit volume, and the rent needs to be 40 billion per year. Force supports 1.35 billion units. The latest data shows that the global supercomputing computing power scale in 21 years was 1.4 billion units. According to the data in 21 years, it can only support 100 million visits per day worldwide, while the domestic supercomputing calculation example is only 300 million units.