More than 1,000 deployed systems would give Chinese buyers a domestic alternative at real scale; the 45% utilization claim is the number rivals and customers will try to verify.
Huawei has put a deployment number on its biggest bet against Nvidia in China. The company says more than 1,000 sets of its Ascend SuperPoD systems are now in use, according to remarks by Zhang Dixuan, president of Huawei’s Ascend computing product line, reported by Guangming Online. He spoke on November 2 in Wuzhen, Zhejiang province, where the Ascend 950 SuperPoD won one of 20 Pioneering Science and Technology Awards at the 2026 World Internet Conference. The conference’s Wuzhen Summit runs from November 2 to 5 with 27 sub-forums.
Huawei says more than 1,000 Ascend SuperPoD systems have been deployed across internet companies, telecom carriers, finance, education and energy. Zhang also gave performance figures for the Ascend 950 system: in a 1,024-chip configuration, it uses up to 45% of its theoretical compute during model training, which he put at 1.8 times the industry level, and inference latency falls as low as 5 milliseconds.
The figures are Huawei’s own, presented at an awards ceremony, and the reports cited no independent benchmark. They are still the most specific public account so far of how widely China’s main domestic alternative to Nvidia’s rack-scale systems has spread. Zhang framed the system as a response to demand, saying China leads the world in large-model calls and token consumption and that SuperPoDs have become the industry’s consensus design for AI infrastructure, according to Guangming Online.
Huawei describes the 950 as the largest commercial system of its kind
A SuperPoD links a large number of AI chips so they can behave like one computer, which matters for training and serving very large models. According to Zhang, a single Ascend 950 SuperPoD combines 1,024 neural processing units (NPUs), Huawei’s AI chips, with more than 1,300 interconnect chips and more than 4,000 optical modules. Huawei calls it the industry’s largest commercial SuperPoD.
The system is built on Huawei’s own Lingqu (灵衢) interconnect protocol. CNMO reported that one protocol connects every component in the system as equals, which Huawei says breaks the communication bottlenecks of traditional master-slave designs. The system comes in 128-chip and 1,024-chip configurations and uses unified memory addressing, which CNMO said lets the system operate like a single computer for models with trillions or tens of trillions of parameters.
Zhang also said more than 40 large models have completed native pretraining on Ascend chips. Pretraining is the most compute-hungry stage of building a model, when it learns from large datasets before it is put to work answering requests.
Utilization is the claim that matters most to buyers
For buyers comparing Huawei with Nvidia, raw chip counts matter less than how much of a system’s capacity a model can actually use. A 45% utilization rate on a 1,024-chip system, if it holds up across customers, would narrow one of the main practical gaps that have kept Chinese developers reliant on Nvidia hardware for training.
Demand for the hardware has been visible in large customer plans. EastFrontier reported in September that DeepSeek plans a giant Huawei chip cluster for inference. Zhang said Ascend would keep developing around a "SuperPoD plus super-cluster" architecture and would stay committed to open source and openness.
The award placed Huawei among a mixed field. According to the list published by IT Home, other winners in the key technology group included the Jiuzhang 4.0 photonic quantum computing prototype, Intel’s third-generation Core Ultra processors built on its 18A process, and a general-purpose robot manipulation foundation model. Engineering winners included AMD’s Ryzen AI Max+ processor, Microsoft’s Copilot Cowork and the Qoder agent platform.
Customer disclosures will test the deployment figure
Huawei did not break the 1,000-plus figure down by chip count, generation, customer or location. It did not say how many of the deployed systems are the newer 950 rather than earlier SuperPoD designs.
The clearest confirmation would come from customers. Chinese cloud providers and model developers that describe their training runs, particularly any that report full pretraining of a large model on Ascend 950 systems with utilization figures of their own, would show whether Huawei’s numbers hold up outside its own presentations.