China’s Tech Titans Aren’t Buying Huawei Chips Due To Overheating Issues And NVIDIA’s Existing Ecosystem Lock-Ins Via CUDA

This is not investment advice. The author has no position in any of the stocks mentioned. Wccftech.com has a disclosure and ethics policy.

Huawei hoped to wean off China from NVIDIA's dominance via its Ascend 910C GPUs, but continues to encounter substantial inertia in this endeavor, courtesy of NVIDIA's ecosystem lock-ins via the CUDA software and Huawei's own shortcomings.

To wit, the Information is now reporting that China's tech titans, including TikTok's parent ByteDance, Alibaba, and Tencent, have yet to order Huawei's AI chips in large quantities.

Several factors have coalesced to create a sizable inertia around Huawei's 910C GPUs, which are currently being directed towards China's large State-Owned Entities (SOEs) and local governments in the absence of a tech-driven order book momentum.

First, many of China's tech titans have invested quite heavily in NVIDIA's CUDA ecosystem, and an eventual break from NVIDIA's proverbial shackles will likely entail a significant investment of time and resources. In fact, the Information reports that many of these tech companies expect Huawei to adapt to their platforms instead of the other way around.

This issue becomes all the more prominent when one considers the fact that Huawei's alternative for CUDA, Compute Architecture for Neural Networks or CANN, is not as feature-rich as NVIDIA's bespoke software.

Second, most of China's largest tech companies are competitors of Huawei and, as such, still feel reluctant to go all-in on their competitor's offering.

Third, Huawei's Ascend 910C chips suffer from periodic overheating issues, which affects their perception of reliability in China's tech circles.

Of course, if DeepSeek were to go all-in on Huawei's AI chips, it would encourage a veritable army of open-source developers to build on Huawei's ecosystem. This eventuality has yet to occur, however.

Fourth, many of China's largest tech companies stashed away a sizable repository of NVIDIA's GPUs over the years. This inventory has yet to run out, creating minimal inducement for these companies to make the costly switch, for now at least.

Fifth, the US Department of Commerce made Huawei's chips quite toxic in May when it issued sweeping guidance that any company that used those chips without prior authorization could be deemed to be in violation of US export controls. This guidance hit Chinese entities with a sizable overseas footprint quite hard.

As we noted in a previous post, Huawei's Ascend 910C combines two older 910B chips to reportedly deliver 800 TFLOP/s of computing power at FP16, replete with a memory bandwidth of up to 3.2 TB/s. The chip is largely considered on par with NVIDIA's H100 GPU.

Recently, to provide an alternative to NVIDIA's supercomputer that combines up to 72 Blackwell chips using its bespoke NVLink, Huawei unveiled its CloudMatrix 384, which bundles up to 384 Ascend chips to provide comparable computing power but lacks direct support for memory-optimal calculation formats such as the FP8. When AI models are trained on the FP8 format, they typically consume a lot less memory. Of course, Huawei has created a translation tool to induce artificial compatibility with FP8, but this solution remains suboptimal.

Meanwhile, NVIDIA appears to be doing quite well even without tailwinds from China. For instance, UBS recently highlighted NVIDIA's statement during its Q1'26 earnings call that it had visibility into "tens of gigawatts" of AI infrastructure projects. UBS then performed a theoretical exercise, one that assumed a 20GW pipeline and between $40 billion and $50 billion that NVIDIA has already stated it stands to gain for each GW of AI infrastructure. Assuming a realization timeline of between 2 and 3 years for these projects, this pipeline represents ~$400 billion per year in revenue for NVIDIA's AI-critical data center segment alone!

Follow Wccftech on Google to get more of our news coverage in your feeds.