China’s true artificial intelligence computing capacity has long been obscured by a combination of geopolitical caution and differing measurement standards. However, a recent disclosure from the Ministry of Industry and Information Technology (MIIT ) has provided a rare glimpse into the scale of the country’s “dark” compute power. According to the South China Morning Post, China has achieved an astonishing 1,882 exaflops of AI computing power, a figure that is approximately 6,000 times higher than the performance of its fastest publicly acknowledged system on the global Top500 list.
This revelation highlights a significant divergence between China’s actual infrastructure capabilities and the metrics tracked by international benchmarks. The Top500 list, a Germany-based ranking of the world’s fastest supercomputers, currently lists the United States’ El Capitan system at the Lawrence Livermore National Laboratory as the global leader, with a performance of roughly 1.8 exaflops. In contrast, China’s fastest known system on the November 2025 Top500 release registered below 0.1 exaflops. The MIIT’s announcement suggests that China’s aggregate AI computing power dwarfs these public figures, fundamentally altering the perceived balance of technological power.
(Related: China Sets Its Sights on 80 Percent Chip Self-Sufficiency by 2030)
The Measurement Disconnect and Geopolitical Silence
The massive discrepancy between the MIIT’s 1,882 exaflops figure and the Top500 rankings is partly due to differing measurement methodologies. The Top500 list traditionally measures high-precision, double-precision floating-point operations (FP64), which are essential for complex scientific simulations. The MIIT figure, however, utilizes an AI-specific format that counts simpler calculations, such as half-precision (FP16) or lower, which are optimized for training and running neural networks.
Even when adjusted to align with the Top500 standard, the implications remain staggering. Experts estimate that the 1,882 exaflops of AI compute would translate to roughly 120 to 230 exaflops of traditional supercomputing power. This adjusted figure is still orders of magnitude higher than China’s public benchmarks and far exceeds the combined capacity of the top systems currently listed globally. This “dark” compute power represents a vast, unlisted reservoir of processing capability dedicated specifically to advancing artificial intelligence.
The absence of China’s most powerful systems from the Top500 list is a deliberate strategic choice. Jack Dongarra, a Turing Award winner and co-founder of the Top500, noted in 2023 that China had stopped submitting details of its advanced machines amid escalating geopolitical tensions. By withholding this data, Beijing avoids drawing further attention from US policymakers, who have increasingly targeted China’s access to advanced semiconductors and computing equipment through export controls.
Building a Nationwide Computing Grid
The MIIT’s disclosure is not merely a boast of raw power; it is tied to a broader national strategy to democratize access to AI infrastructure. Zhang Yunming, the MIIT vice-minister, stated that China is actively building a nationwide, multilayered computing grid. The goal of this initiative is to make computing power widely available and affordable, particularly for small and medium-sized enterprises that lack the capital to build their own data centers.
This grid approach treats computing power as a fundamental utility, akin to electricity or water. By pooling resources across regions and facilities, the government hopes to optimize utilization rates and lower the barriers to AI innovation. This strategy aligns with projections from IDC and Inspur, which last year forecast a 46% annual growth rate for AI computing in China between 2023 and 2028. The rapid expansion of this infrastructure is critical for supporting the proliferation of domestic large language models and AI applications.
(Related: China’s NDRC Directs Ultra-Long Treasury Bond Proceeds Toward AI Infrastructure)
The scale of China’s hidden compute power also provides context for the rapid advancement of its domestic AI models. A recent report from the Stanford Institute for Human-Centered Artificial Intelligence (HAI) concluded that the performance gap between leading AI models in China and the US has largely closed. The MIIT’s 1,882 exaflops figure suggests that Chinese developers have access to the massive computational resources necessary to train these frontier models, despite the constraints imposed by US export controls. As China continues to build out its nationwide grid, this vast reservoir of “dark” compute will likely serve as the engine for its next generation of AI breakthroughs.