A supply and demand game sweeping the industrial chain
——The impact and response of the rising price of computing power ①
Since the beginning of this year, cloud computing manufacturers and artificial intelligence (AI) service providers such as Tencent Cloud, Alibaba Cloud, and Baidu Smart Cloud have intensively raised the prices of computing power products, with some services increasing by more than 30%. This round of computing power price increases is not only a price adjustment for companies, but also a supply and demand game sweeping the industry chain.
Computing power is the core production factor of the digital economy. Why did this round of price increases for computing power come about? How are price increases transmitted to the upstream and downstream of the industrial chain? What impact will it have on the accelerating development of the AI industry?
Three factors push up prices
"The fundamental reason for the rise in computing power prices is the imbalance between supply and demand. The demand side is growing rapidly, while the supply side is restricted by multiple bottlenecks, and the gap between the two continues to expand." said Li Mingyan, director of the Intelligent Economy Research Institute of Saizhi Industrial Research Institute.
Zhang Linshan, a researcher at the Macroeconomic Research Institute of the National Development and Reform Commission, believes that the explosive development of generative AI applications has created a demand for massive computing power. Technology giants and start-ups have joined in the "Battle of 100 Models", resulting in an exponential surge in demand for high-performance graphics processors (GPUs) and intelligent computing centers.
The rapid growth in the number of token calls is a direct reflection of the expansion of the demand side. As of March this year, my country's average daily Token calls have exceeded 140 trillion, an increase of more than 1,000 times compared with 100 billion at the beginning of 2024, and an increase of more than 40% in 3 months compared with 100 trillion at the end of 2025.
The change in demand structure reflects the new trend of the computing power market. Data from the National Data Administration shows that in 2025, the total amount of data used for artificial intelligence training and inference in my country will be 199.48 exabytes (EB), and the amount of inference data will exceed the amount of training data for the first time, reaching 101.34 EB. According to estimates, the ratio of inference computing power requirements to training computing power requirements may reach 3:1 or even higher in the future. Training determines the upper limit of the model's capabilities, while inference determines the commercial implementation of the model. This change changes computing power from a phased investment to a continuous operating expenditure.
Supply-side constraints are also very prominent. Zhu Keli, founding director of the Guoyan New Economic Research Institute, believes that the production capacity of core hardware such as high-end AI chips and high-bandwidth storage is tight, and the cost of upstream components has risen, directly pushing up the cost of server procurement. At the same time, the construction of intelligent computing centers must be matched with power and network facilities. The production cycle of new computing power clusters is long, and short-term supply cannot keep up with the pace of demand expansion.
Li Mingyan judged that this round of computing power price increases is not a short-term fluctuation, but a landmark node for the AI industry to move from "subsidized expansion" to "commercial sustainability." The structural migration from training to inference on the demand side has opened up room for long-term growth. The supply-side bottlenecks of hardware, packaging, and delivery are difficult to fundamentally alleviate in the short term, and the industry's own business logic has shifted from "burning money" to "profiting." The superposition of three forces has jointly pushed up the price of computing power.
The industrial chain is accelerating adjustment
On August 17, the DeepSeek API (Application Programming Interface) price adjustment officially came into effect. Industry insiders analyze that computing power cost pressure is the main reason for the price increase of large artificial intelligence models.
The rise in computing power prices is being transmitted upstream and downstream along the industrial chain. A game around costs, profits and market share also unfolded. Price increases are first transmitted upstream, driving the expansion of investment in computing infrastructure. As the demand for AI continues to grow, cloud vendors continue to increase capital expenditures, driving the expansion of demand for core hardware such as GPUs, storage, and advanced packaging. "Upstream hardware manufacturers are ushering in a boom cycle with both volume and price rising. Upstream links such as chips, storage, and advanced packaging are the biggest beneficiaries of price increases." Li Mingyan said.
Computing power leasing companies in the middle reaches of the industry chain are facing both opportunities and pressures. Industry insiders said that as the demand for computing power continues to grow, the computing power leasing market space will further expand, and the bargaining power of related companies is expected to increase. At the same time, computing power costs account for 70% to 80% of its operating expenses, and rising upstream prices will directly compress profit margins.
When the pressure of price increases is eventually transmitted to the downstream, the cost pressure on AI application companies will become more obvious, and industry differentiation may further intensify. Computing power costs are becoming an operational burden for large model companies and AI application companies. Leading model manufacturers have become important incremental customers in the computing power rental market other than cloud service providers. However, weaker small and medium-sized AI companies may be squeezed out of the market due to high computing power costs, accelerating industry reshuffle.
Zhang Linshan believes that faced with the challenge of rising computing power prices, companies should adopt strategies to reduce costs, increase efficiency and diversify their layout. Technically, we reduce computing power consumption through model distillation, quantification and other means, and explore alternatives such as domestic chips; in terms of resources, we adopt a hybrid cloud architecture to flexibly schedule computing power resources to smooth price fluctuations; strategically, large enterprises can appropriately deploy self-built computing power facilities, while small and medium-sized enterprises can use the computing power leasing model to reduce initial investment and maximize the value of unit computing power through refined management.
Zhu Keli said that computing power supply companies cannot simply rely on price increases to make profits. They must improve resource scheduling efficiency, revitalize idle computing power, and enrich multiple service models such as pay-per-use billing and flexible leasing. In the long run, the entire industry must move away from simply competing for hardware resources, extend to algorithm optimization and scenario solutions, and rely on technical efficiency to hedge against the pressure brought by rising hardware costs.
Improve the computing power supply system
How long will the rising price of computing power last? Many experts have judged that in the next two to three years, computing power prices will show a trend of stabilizing and rising at first, and then gradually stabilizing and diverging. In the short term, high-end training computing power will remain high. As global computing power production capacity continues to be invested and the domestic computing power ecosystem matures, the growth of general inference computing power will slow down; in the long term, prices will gradually fall.
Faced with the dual pressures of supply shortage and increased demand, in addition to expanding hardware supply, optimal utilization of existing and new computing resources through the scheduling capabilities of cloud services has become the choice of many enterprises.
Take Tencent Cloud as an example. Through a unified scheduling layer, Tencent Cloud integrates computing resources of different chips, different tasks, and different locations into the same platform management, and dynamically matches them according to actual business needs - elastic allocation of training and inference, collaborative responsibility between cloud and edge, and off-peak and online off-peak operation.
Li Wei, deputy director of the Cloud Computing and Digital Research Institute of the China Academy of Information and Communications Technology, believes that the industry will shift from competing for floating-point computing power to competing for Token performance. Cloud service providers need to systematically reduce token costs through model architecture optimization and other means, so as to build differentiated competitive advantages at the price level and achieve dual breakthroughs in technical cost reduction and commercial competitiveness.
In April this year, the computing power network was officially included in the construction of the national "six networks", and more than 70 major computing power channels have been built around the national computing power hub nodes. Li Mingyan believes that on this basis, we should accelerate the promotion of "one network" dispatch of computing power resources across the country, break down regional and entity barriers, and make computing power truly flow as freely as hydropower; continue to explore market-oriented cooperation models such as "computing power for green electricity" to make western green power truly become the low-cost energy support for eastern computing power; increase investment in research and development to break through key technological bottlenecks such as advanced processes and advanced packaging.
Building an efficient, inclusive, and secure computing power supply system requires coordinated promotion by multiple parties. Zhang Linshan suggested that at the technical level, we should continue to break through the key technologies of high-end chips and basic software, improve the heterogeneous computing power scheduling platform, and improve the utilization rate of computing power resources; lay out a national integrated computing power network, deepen the "Eastern Data and Western Calculation" project, promote cross-regional circulation and transactions of computing power, lower the threshold for computing power use, and promote inclusive computing power. At the industrial level, encourage the construction of green intelligent computing centers and develop liquid cooling and low-carbon computing power infrastructure.
(Economic Daily reporters Lai Qichun Huang Xin)
