Regarding capital expenditure, Tencent executives said at the performance conference that the company's current capital expenditure is mainly used for AI infrastructure construction. This type of investment is not a one-time consumption, but rather forms infrastructure that can be used sustainably. It can not only be used for next-generation model training, but also continuously supports token generation. As the infrastructure built in 2026 and 2027 continues to function in subsequent years, it is expected that the operating pressure brought about by related investment will gradually ease by 2028. At the same time, investment costs for AI infrastructure are also expected to change positively. The purchase price of equipment is relatively high in 2026. It is expected that by 2028, the purchase price of similar equipment will drop significantly. More importantly, the computational efficiency of model training and inference is rapidly improving, which means that under the same infrastructure conditions, model training can be completed with greater efficiency in the future and more tokens can be processed. Executives believe that infrastructure reuse, reduced hardware procurement costs, and improved computational efficiency will jointly improve the input and output efficiency of the AI business, and have a more optimistic judgment on 2028.

Zhitongcaijing · 1d ago
Regarding capital expenditure, Tencent executives said at the performance conference that the company's current capital expenditure is mainly used for AI infrastructure construction. This type of investment is not a one-time consumption, but rather forms infrastructure that can be used sustainably. It can not only be used for next-generation model training, but also continuously supports token generation. As the infrastructure built in 2026 and 2027 continues to function in subsequent years, it is expected that the operating pressure brought about by related investment will gradually ease by 2028. At the same time, investment costs for AI infrastructure are also expected to change positively. The purchase price of equipment is relatively high in 2026. It is expected that by 2028, the purchase price of similar equipment will drop significantly. More importantly, the computational efficiency of model training and inference is rapidly improving, which means that under the same infrastructure conditions, model training can be completed with greater efficiency in the future and more tokens can be processed. Executives believe that infrastructure reuse, reduced hardware procurement costs, and improved computational efficiency will jointly improve the input and output efficiency of the AI business, and have a more optimistic judgment on 2028.