Since the start of the year, some Chinese cloud computing and AI service providers have adjusted prices for computing power products. According to Economic Daily, prices for some services have risen by more than 30%, though the actual increases vary by provider, product type and billing model.

Changes in computing power prices affect not only cloud providers and compute leasing companies, but also ripple through the industry chain to GPUs, storage, advanced packaging, data centers and downstream AI applications.

Demand Growth and Supply Constraints Drive Prices

Li Mingyan, director of the Institute of Intelligent Economy at the CCID Industry Research Institute, said the latest rise in computing power prices has been driven mainly by growing demand and supply constraints. The continued expansion of generative AI and agent applications has increased enterprise demand for high-performance chips, AI data centers and inference services.

Rising Token usage is one sign of that shift in demand. Citing relevant data, the report said China’s average daily Token usage had exceeded 140 trillion as of March 2026, up from 100 billion in early 2024 and 100 trillion at the end of 2025.

Data from the National Data Administration shows that China used 199.48 exabytes of data for AI training and inference in 2025. Of that total, inference accounted for 101.34 EB, surpassing training data for the first time.

This indicates that computing power demand is expanding from centralized model training to continuously running inference and application services. However, the future balance between inference and training demand will still depend on model architectures, application types and computing efficiency.

Supply is constrained as well. Zhu Keli, founding president of the China Institute of New Economy, said the availability of hardware such as high-end AI chips and high-bandwidth memory affects the cost of procuring servers and building computing power clusters. AI data centers also require supporting power, networking and cooling systems, and typically take time to move from construction to operation.

Price Changes Ripple Through the Industry Chain

Upstream investment in computing power infrastructure is driving demand for hardware including GPUs, storage, advanced packaging and servers. Li said some upstream segments could benefit from rising computing power demand, though supply-demand dynamics, pricing and profitability continue to vary across products.

Compute leasing companies, meanwhile, face both stronger demand and higher costs. Rising market demand could improve equipment utilization and pricing power for computing power service providers, but higher costs for servers, electricity and data center facilities could also squeeze margins.

According to the report, some industry participants estimate that computing power may account for 70% to 80% of operating expenses at certain compute leasing companies. The ratio does not apply to every company and must be assessed against each provider’s business model and financial data.

For downstream foundation-model and AI application companies, higher computing power prices could increase the cost of model training, inference and product operations. Companies with fewer financial resources may need to control spending by using external computing power, reducing model size or optimizing algorithms.

DeepSeek’s API price adjustment took effect on August 17. The report cited industry views that model providers may adjust API prices in response to a range of factors, including inference demand, hardware costs and service operations, and that price changes alone do not point to any single cause.

Companies Seek to Reduce Computing Power Costs per Unit

Zhang Linshan, a researcher at the Academy of Macroeconomic Research under the National Development and Reform Commission, said companies can reduce their consumption of computing resources through techniques such as model distillation and quantization. They can also use hybrid cloud and heterogeneous computing power scheduling to allocate resources flexibly based on the type of workload.

For large companies, building in-house computing power infrastructure offers greater control over resources but requires substantial upfront investment and ongoing operating costs. Small and medium-sized businesses can instead use cloud services or compute leasing to reduce their initial hardware spending.

Zhu said computing power providers need to improve resource scheduling efficiency and explore service models such as usage-based billing and flexible leasing. Beyond hardware supply, algorithm optimization and industry-specific solutions also affect the actual cost of computing power services.

Industry Shifts Focus to Token Efficiency

As computing power resources span different chips, workloads and regions, unified scheduling has become an important way to improve utilization.

Tencent Cloud, for example, uses a unified scheduling layer to manage different types of computing chips and computing power resources, allocating capacity according to business needs across training and inference, cloud and edge environments, and online and offline workloads.

Li Wei, deputy director of the Cloud Computing and Digitalization Research Institute at the China Academy of Information and Communications Technology, said the industry’s competitive metrics are expanding beyond floating-point performance to include Token efficiency. Cloud providers need to reduce the cost per Token through model architecture, inference engines and resource scheduling.

The report also noted that in April 2026, the computing power network was incorporated into the construction of China’s six national networks, with more than 70 major computing power corridors already built around the country’s computing power hub nodes.

Experts recommended further improving the nationwide integrated computing power scheduling system to facilitate the flow of computing power resources among regions and market participants, while strengthening research and development in advanced process nodes, advanced packaging, foundational software and heterogeneous computing.

The outlook for computing power prices will continue to depend on chip supply, data center construction, electricity costs, model efficiency and market competition. Pricing trends may diverge across different types of computing power products, and industry forecasts will need to be assessed against actual market data.