Zhipu said during a conference call on September 16 that big model companies need firm scaling. With an investment of 30 billion yuan of computing power, about 100,000 P can be built, 40% reserved for training and R&D, and 60% for business reasoning. Thanks to GLM-5.3's high inference cost ratio, theoretical inference gross profit can reach up to 80%, and 60% of inference computing power can earn 40 billion yuan a year, covering computing power investment and model training and development expenses. In reality, it is affected by fluctuations in computing power utilization, business discounts, and computing power prices.

Zhipu said during a conference call on September 16 that big model companies need firm scaling. With an investment of 30 billion yuan of computing power, about 100,000 P can be built, 40% reserved for training and R&D, and 60% for business reasoning. Thanks to GLM-5.3's high inference cost ratio, theoretical inference gross profit can reach up to 80%, and 60% of inference computing power can earn 40 billion yuan a year, covering computing power investment and model training and development expenses. In reality, it is affected by fluctuations in computing power utilization, business discounts, and computing power prices.