China has taken a giant step in its race for technological autonomy. According to industry sources, the Asian country has managed to bring together more than 100,000 domestically manufactured artificial intelligence accelerators in a single computing supercluster. The milestone is impressive, but it poses a huge problem: getting all those chips to work together efficiently.
The cluster, whose code name would be “Jiaolong” (dragon of the deep), would be deployed at a facility in Guangdong province and would use exclusively domestic accelerators, mainly the Huawei Ascend 910B and later versions, along with chips from other Chinese companies such as Cambricon and Biren Technology. The figure of 100,000 chips places it on par with the largest AI superclusters in the world, dominated until now by NVIDIA and its H100/B200 GPUs.
The problem of software and interconnect
The great challenge is not manufacturing the chips – China already produces tens of thousands of accelerators – but getting them to work in a coordinated way. In NVIDIA clusters, the combination of CUDA software, NVLink interconnect and unified memory allows thousands of GPUs to work as if they were a single one. China lacks an equivalent ecosystem.
To solve this, the Chinese government has launched a program that forces chip manufacturers to standardize their communication interfaces. The goal is to create a common software layer that allows accelerators from different manufacturers (Huawei, Cambricon, Biren) to coordinate within the same network. This is what engineers call “heterogeneous orchestration” and it is one of the most complex problems in distributed computing today.
Real vs. theoretical performance
In theory, 100,000 Ascend 910B accelerators offer a colossal raw computing power, on the order of exaflops in reduced precision (FP8/FP16). However, the first internal tests suggest that the cluster’s real efficiency hovers around only 30-40% of its theoretical capacity, far from the 70-80% that equivalent systems with NVIDIA achieve.
The discrepancy is mainly due to two factors: the latency in communication between chips when different manufacturers are mixed, and the lack of an ecosystem of optimized libraries such as cuDNN or TensorRT. Chinese engineers are working against the clock to develop domestic alternatives to these tools.
A race with geopolitical implications
This project is not just a technical matter. United States sanctions have cut off China’s access to NVIDIA’s most advanced chips (H100, B200 and Blackwell). Beijing responds with a large-scale import substitution strategy: if they cannot buy the best chips in the world, they will make their own and make them work as a pack.
The question experts ask is whether this strategy will work. On one hand, China has unlimited resources and political will. On the other, building a software ecosystem competitive with CUDA is not achieved in a few years – it is the result of almost two decades of continuous NVIDIA investment.
What’s coming
The “Jiaolong” cluster is expected to enter full production by the end of 2026, and its first uses will be the training of large-scale domestic language models (LLMs) – alternatives to GPT-4 and Claude – and computer vision applications for defense and surveillance.
The outcome of this experiment will largely define the future of the AI industry: either China proves that it can do without NVIDIA and build technological sovereignty in AI, or it confirms that the CUDA ecosystem is a moat impossible to cross. Whatever the outcome, the world will be watching.
