TECH NEWS – Although they could easily back domestic hardware from Huawei, China’s leading tech firms continue to side with Nvidia for several reasons. Huawei had hoped its Ascend 910C GPUs would finally loosen Nvidia’s grip on the Chinese market, but so far it faces major obstacles: Nvidia’s closed CUDA software ecosystem and Huawei’s own technological shortcomings remain barriers.

According to The Information, Chinese giants such as ByteDance (TikTok’s parent), Alibaba, and Tencent have yet to place large orders for Huawei’s AI chips. Several factors have led to inertia around the Ascend 910C GPUs, which are mostly finding homes in big state-owned companies or local governments due to a lack of tech sector demand. Breaking free from Nvidia’s entrenched CUDA ecosystem would require massive investments of time and resources, and most companies expect Huawei to adapt to them, not the other way around.

This is even more apparent since Huawei’s own Compute Architecture for Neural Networks (CANN) simply can’t match the feature set of Nvidia’s bespoke software. On top of that, many of China’s tech titans are direct competitors to Huawei, so they’re naturally reluctant to embrace a rival’s hardware. Reliability is also an issue: the Ascend 910C chips are known to suffer from overheating, damaging their reputation. If, for example, DeepSeek were to adopt Huawei’s AI chips, it could kick-start open-source development around Huawei’s ecosystem, but that shift hasn’t happened yet. Meanwhile, most Chinese tech giants are still sitting on large Nvidia GPU reserves, leaving little incentive for a costly switch.

The US Department of Commerce dealt Huawei’s chips a major blow in May, issuing guidance that any company using them without prior approval could be violating export controls, hitting Chinese firms with overseas operations especially hard. The Ascend 910C combines two older 910B chips, boasting up to 800 TFLOP/s on FP16 and 3.2 TB/s of memory bandwidth, making it a direct rival to Nvidia’s H100. Huawei’s CloudMatrix 384, which links up to 384 Ascend chips, seeks to challenge Nvidia’s supercomputer lineup but doesn’t support memory-optimized formats like FP8 natively. While Huawei has built a translator tool for FP8 compatibility, this workaround is far from ideal.

Despite all this, Nvidia seems to be thriving even without a Chinese tailwind. UBS recently highlighted Nvidia’s claims that it has visibility into tens of gigawatts of AI infrastructure projects, and their estimates suggest the company could rake in as much as $400 billion in annual data center revenue if those projects come online within a few years.

Source: WCCFTech, The Information