Huawei's AI chips struggle to compete with Nvidia because they run hotter and their development tools lag behind CUDA. Huawei once hoped its new Ascend 910C line could help China reduce its dependence on Nvidia's server GPUs, but it appears the company is still facing significant obstacles. The main causes are Nvidia's lasting advantage from its CUDA software ecosystem and Huawei's own internal limitations.
According to sources cited in a new report from The Information, major Chinese tech companies, including ByteDance (TikTok's parent company), Alibaba and Tencent, have yet to place large orders for Huawei's AI chips.
The Obstacles Holding Back the Ascend 910C AI Chip Against Nvidia
Several factors have contributed to the stagnation around Huawei's Ascend 910C GPU, which is currently being supplied mainly to large state-owned enterprises and local governments because orders driven by real technical demand are lacking.
First, many leading Chinese tech companies have invested heavily in Nvidia's CUDA ecosystem, and breaking away from Nvidia's control of the cloud server AI chip market would demand a major investment of both time and resources. The Information also reports that many of these companies expect Huawei to adapt to their platforms, rather than the other way around, with them changing how they develop and manage software to suit Ascend chips.
The problem becomes clearer when you consider that CANN (Compute Architecture for Neural Networks), Huawei's alternative to the CUDA API, lacks the full feature set of NVIDIA's proprietary software.
Second, most of China's biggest tech companies are competitors of Huawei itself, so they remain wary of betting fully on a rival's products.
Third, Huawei's Ascend 910C chip runs into overheating problems during operation, which hurts perceptions of its reliability among China's large tech groups. If DeepSeek decided to back Huawei's AI chips, it would encourage a large army of open-source developers to build on Huawei's ecosystem. That hasn't happened yet, however.
Fourth, many of China's largest tech companies pre-ordered and stockpiled a large number of Nvidia GPUs over the past few years. That inventory remains plentiful, leaving corporations and cloud services with little incentive to make a costly switch, at least for now.
Fifth, the US Department of Commerce made Huawei's chips a "sensitive" item in May, when regulators in Washington issued broad guidance that any company using these chips without prior approval could be seen as violating US export controls. This guidance has a major impact on Chinese entities that operate overseas.
The Technology Race Between Huawei and Nvidia in AI
As mentioned in an earlier post, Huawei's Ascend 910C combines two older 910B chips to deliver roughly 800 TFLOPs of compute in the FP16 floating-point format, with memory bandwidth of up to 3.2 TB/s. The chip is rated on par with NVIDIA's H100 GPU.
More recently, to offer an alternative to NVIDIA's supercomputer that combines up to 72 Blackwell chips over its proprietary NVLink interconnect, Huawei introduced the CloudMatrix 384, which integrates up to 384 Ascend chips to deliver comparable compute capacity but lacks direct support for memory-optimized compute formats such as FP8. AI models trained in FP8 typically use significantly less memory. Huawei did build a translation tool to emulate FP8 compatibility, but that solution is still far from optimal.
Meanwhile, Nvidia appears to be doing very well even without support from China. For example, UBS recently highlighted Nvidia's statement on its first-quarter 2026 earnings call that it can see AI infrastructure projects with a total operating capacity of up to tens of gigawatts worldwide. For every gigawatt of AI server compute capacity put into operation, Nvidia is estimated to earn 40 to 50 billion USD in revenue.
That means if countries around the world keep expanding data center construction over the next 2 to 3 years, Nvidia could bring in 400 billion USD a year from its cloud server business alone.
In short: Huawei's Ascend 910C AI chip still faces many challenges in competing with Nvidia. Although it can reach about 60% of the Nvidia H100's performance in some AI inference tasks, overheating and a software ecosystem less developed than CUDA are major obstacles. On top of that, big Chinese tech companies remain wary of switching to a competitor's chips, and US policy restrictions also make it harder for Huawei to expand its market. Meanwhile, Nvidia continues to reinforce its dominant position in the global AI chip market, with revenue projected to reach hundreds of billions of USD in the coming years.
