CBN News 2026-11-09 09:36:53
Editor: Qi Xinyi
Previously, there were reports that Microsoft decided not to renew the older-generation GPU computing power lease agreement signed with Oracle in 2023. Yicai's "One Rating" believes that this move is not due to a peak in the overall demand for AI computing power, but rather a systematic upgrade in the procurement standards of leading cloud providers. It marks that the global AI computing power market is bidding farewell to a period of general price increases and entering a K-shaped structural differentiation phase characterized by scarcity and premium prices for "new specification clusters" and a downward diversion of demand for older-generation computing power.
"Yi Rating" points out that as the technology iteration cycle of NVIDIA GPU has been shortened to around 2 years, the requirements of top customers for cluster-level full-stack delivery capabilities have significantly increased. The older generation of chips have been phased out due to their inability to match high-speed interconnection and training efficiency, resulting in a substantial reduction in their economic lifespan. However, the demand from small and medium-sized enterprises (SMEs) and vertical industries, which are sensitive to price, can still effectively utilize the physical lifespan of the older generation of GPU chips, supporting their residual value models.
This means that the profit model of computing power rental providers is undergoing a fundamental shift. Manufacturers that rely on high leverage for expansion and have a single customer base will face the risk of asset revaluation; those with a diversified customer base and comprehensive cloud service capabilities are expected to maintain stable returns through tiered operations. In this context, the focus of valuation in the industry chain will shift from mere volume expansion to a deeper consideration of asset quality and generational compatibility, and the certainty of high-end chips and supporting infrastructure is expected to be further strengthened.
The demand for computing power has changed: from focusing on peak values to calculating “effective Token”.
The competition in future computing power is not just about the peak performance of a single chip, but also about the overall system capabilities.
AI Evolution Express | Apple Officially Announces Apple Intelligent Integration with Alibaba Qianwen
Apple officially announces Apple intelligent integration with Alibaba Qianwen; The world's first integrated space computing power cloud platform is launched in Beijing......
High-value data is still in short supply; what's holding back the "effective use" of data at the last mile?
The development of intelligent agents is also forcing the reconstruction of underlying infrastructure.
AI Evolution Express | Hainan Computing Power Joins Strategic Partnership with Tencent Cloud
AI Evolution Express | Hainan Computing Power Joins Strategic Partnership with Tencent Cloud
ByteDance expands recruitment of computing power talents; Big Companies Enter the Stage of Hardware-Software Collaboration Competition
Manufacturers need to transform computing power into higher training efficiency, lower inference costs, and a sustainable product delivery capability.
