China’s total computing power reached 962 exaflops (EFLOPS) in 2025, ranking second globally and accounting for approximately 21% of worldwide AI computing capacity, according to the 2025 China Server Industry Development Report released at the Smart Computing Foundation event in late July. As Pandaily reported, the report also confirmed a structural inflection point in China’s server market: for the first time, AI server sales surpassed general-purpose server sales in a single year, marking a fundamental shift in the country’s computing infrastructure priorities that has significant implications for the global AI hardware race.
The AI Server Inflection Point and What It Signals
The crossing of this threshold is significant beyond the headline numbers. General-purpose servers, the workhorses of enterprise IT, cloud computing, and web services, have historically dominated the server market in volume and value. The fact that AI servers, which are far more expensive and specialized, have overtaken them in sales reflects the sheer scale of investment that Chinese technology companies, cloud providers, and government entities are pouring into AI infrastructure.
This shift is consistent with the broader pattern of China’s dark AI computing power being far higher than public benchmarks suggest. Much of China’s AI compute is deployed in environments that do not report to international benchmarking bodies, meaning that the 962 EFLOPS figure, while significant, likely understates the true scale of China’s AI computing capacity. The report’s authors acknowledge that the figure covers publicly reported infrastructure and that actual deployed capacity is higher.
The inflection point also reflects a change in the nature of demand. A year ago, the primary driver of Chinese AI server purchases was large language model training by frontier AI companies. Today, the demand is increasingly driven by inference, the deployment of trained models at scale for commercial applications. This shift from training to inference is a sign of maturation in China’s AI industry: companies are no longer just building models, they are deploying them at scale and generating revenue from them.
Domestic Substitution as the Structural Growth Driver
A key structural driver of this growth is the Xinchuang (信创) domestic substitution policy, which mandates that government agencies, state-owned enterprises, and entities in critical sectors such as finance, energy, military replace foreign technology with domestically produced alternatives. This policy provides a reliable, policy-backed demand floor for Chinese server manufacturers, insulating them from the cyclical volatility that affects commercial markets.
The sectors accelerating domestic server adoption are precisely those with the largest and most stable IT budgets: government agencies, major banks, state energy companies, and defense-related entities. As China’s domestic AI chips enter the training stack and the efficiency gap with international alternatives narrows, the case for Xinchuang compliance becomes commercially as well as politically compelling. Companies that previously resisted switching from Nvidia hardware on performance grounds are finding that the gap has narrowed enough to make domestic alternatives acceptable for most workloads.
The Gap That Remains Between China and the United States
Despite the impressive headline figures, China’s 21% share of global AI computing capacity compares to the United States’ position as the clear global leader. The Wall Street Journal has reported that China’s AI computing power stood at around 14% of the US in 2025, a figure that, while contested, reflects the reality that the US lead in frontier AI training infrastructure remains substantial.
The 962 EFLOPS figure includes a wide range of computing hardware, from high-end AI training clusters to lower-performance inference servers, and the quality-adjusted gap between Chinese and American AI computing infrastructure is likely larger than the raw numbers suggest.
Nevertheless, the trajectory is clear. China is building AI infrastructure at a pace that few other countries can match, driven by a combination of state policy, commercial investment, and the domestic substitution mandate. The crossing of the AI server threshold in 2025 is a milestone that reflects not just the scale of China’s AI ambitions, but the degree to which those ambitions are now embedded in the country’s industrial and financial infrastructure, making a reversal of the trend unlikely regardless of what happens to US-China relations.
The Global Context: Where China’s Compute Fits in the AI Race
China’s 962 EFLOPS figure does not exist in a vacuum. The US remains the global leader in AI computing capacity by a significant margin, and the quality-adjusted gap, accounting for the performance differential between Chinese and American AI chips, is larger than the raw numbers suggest. However, China’s trajectory is what matters for long-term competitive dynamics. The country that builds the most AI infrastructure today is building the foundation for the AI applications and capabilities of tomorrow, and China’s infrastructure investment is now running at a pace that few other countries can match.
The AI server inflection point also has implications for the global semiconductor supply chain. As Chinese AI server manufacturers scale up production, they are creating demand for memory chips, power management ICs, networking components, and cooling systems that flows through the global supply chain, even as the compute chips themselves are increasingly sourced domestically. The ripple effects of China’s AI infrastructure build-out extend well beyond the country’s borders, and the 2025 inflection point is likely to be remembered as the moment when China’s AI ambitions became structurally self-reinforcing.