Executive from Frost & Sullivan: DeepSeek Leads the Rise of Domestic Computing Power, with Utilization Rates of AI Centers Gradually Recovering

With the advent of the domestic large model DeepSeek-R1, which has achieved outstanding performance with less computing power and at a cost less than one-tenth of the industry's standard, the market has undergone a profound reflection on the traditional logic of 'training models by piling up computing power.' After DeepSeek was launched, the market share of computing giants such as NVIDIA experienced fluctuations and stock prices plummeted. However, overall, the emergence of DeepSeek has not significantly reduced China's demand for high-end computing chips; instead, it has driven structural growth in computing demand through its advantages in efficient reasoning and low-cost training. This can be seen from the recovery trend of the stock prices of computing giants like NVIDIA.

Thanks to the innovative collaborative optimization of algorithms, frameworks, and hardware, DeepSeek-R1 can achieve high-performance results with less computing power, significantly improving computational efficiency. Although DeepSeek has reduced the computing power requirements for training a single model through technical optimization, the overall computing power demand is still rising due to the rapid growth in user volume and the continuous expansion of application scenarios. This phenomenon can be explained by the 'Jevons Paradox': although technological progress has improved resource utilization efficiency, the resulting demand growth may instead lead to an increase in total resource consumption.

The open-source strategy of DeepSeek effectively alleviates the pressure brought about by the shortage of high-end computing chips. By reducing training costs and improving inference efficiency, it has promoted the rapid penetration of domestic chips in the market. Especially in training and inference scenarios, domestic chips have gradually replaced imported high-end products. Against the backdrop of growing private deployment of enterprises and edge computing needs, DeepSeek's low-cost solutions help enterprises reduce their dependence on high-end computing power while increasing the substitution rate of domestic chips and further promoting the popularization of core application technologies with domestic solutions.

However, despite the technological progress of DeepSeek that has reduced the demand for traditional high-performance chips to some extent, NVIDIA's high-performance computing power remains an important component that cannot be ignored in the market. Although NVIDIA's market share in high-end training chips (such as H100 and A100) has declined, due to its strong performance advantages in handling large-scale deep learning training and inference, it still holds a dominant position in the field of high-end computing power. Especially in industries with extremely high computing power requirements, NVIDIA's chips remain the preferred choice for many enterprises and research institutions.

After the Spring Festival, driven by DeepSeek's low-cost and efficient inference capabilities, there have been significant changes in the vacancy rate of AI centers and the price of computing cards. The emergence of DeepSeek has not only spurred a surge in the inference market but also helped some previously vacant domestic AI centers gradually resume use. Nevertheless, the overall utilization rate of AI centers still faces considerable challenges, especially in some areas where the actual usage rate remains low.

The main reasons for the low actual utilization rate of intelligent computing centers can be attributed to several aspects. Firstly, an overly optimistic prediction of future computing power demand was a significant factor. Due to the industry's anticipation of rapid growth in future computing power needs, intelligent computing centers were built on a scale that was too large and at a speed that was too fast. However, the actual market demand did not keep up in time, resulting in idle computing resources. Secondly, there is a lag in the supporting construction of the computing power technology ecosystem. While the hardware construction of intelligent computing centers is underway, the related application software, management platforms, and other ecosystems of computing power technologies have not been perfected simultaneously, which has led to ineffective matching and full utilization of hardware resources. Many intelligent computing centers already possess a high level of capability in terms of hardware equipment, but due to the imperfection of the ecosystem, the utilization efficiency of computing power remains low.

In the future, with the rise of domestic technologies such as DeepSeek and the expansion of application scenarios, the utilization rate of intelligent computing centers is expected to gradually increase as technologies continue to improve and the application ecosystem becomes increasingly rich. At the same time, with further clarification of market demand and gradual maturity of the technology ecosystem, the role and value of intelligent computing centers will also become more apparent in the future.