Huawei began helping customers implement Ascend AI accelerators to win the market from Nvidia

Huawei is looking to expand its presence in the AI ​​accelerator market in China, which is currently dominated by Nvidia. The Chinese manufacturer began to help local companies implement its products for inference – launching already trained AI systems, writes the Financial Times.

Image source: Huawei

Nvidia’s products remain vital in the development, or training, of AI systems, while Huawei has positioned its Ascend accelerators as the hardware of choice for Chinese customers to run those systems – the computation needed to generate answers to queries. The Chinese tech giant is betting that the pace of training AI models will slow, but demand for inference hardware will rise as AI applications, including chatbots, are deployed more frequently. Nvidia and Huawei accelerators run on different software, and the Chinese company helps its customers ensure compatibility between the two ecosystems; Chinese officials are calling on local tech giants to move away from Nvidia and actively purchase Huawei chips.

Due to US sanctions, Chinese companies can only officially purchase Nvidia H20 accelerators with artificially low performance, and they are in high demand because they perform better than local alternatives. Huawei Ascend chips are not yet considered as a complete replacement for Nvidia products: the manufacturer has not been able to solve problems with the interaction of chips in large clusters when training large models. The Chinese company is also trying to convince developers to stop using Nvidia CUDA software – they are easy to use and significantly speed up data processing. It is expected that some of the problems will be resolved with the release of new Huawei Ascend 910C accelerators.

Image source: NVIDIA

At the end of 2024, sales of AI equipment in China brought Nvidia $12 billion – the company supplied 1 million of its H20 chips to the country, twice as much as Huawei Ascend 910B was sold. However, in China there are competitors for both Huawei and Nvidia. The duo, which included the Chinese Internet giant Baidu and chip developer Cambricon, have achieved some success in developing chips for AI; In the US, Amazon and Microsoft, in particular, are actively switching to their own chips to run AI. Another obstacle to the growth of Huawei Ascend’s share is the limited resources of the manufacturer – the company is not always able to satisfy the demand for its products. This problem is also associated with American sanctions against the Chinese semiconductor industry.

The development models of the American and Chinese AI markets are sharply different: Meta✴, xAI and OpenAI are building huge clusters of Nvidia accelerators, while in China more attention is paid to solutions in the field of running AI systems – here you can achieve “large efficiency gains even with less powerful chips” and ultimately reduce the cost of launching applications with AI, experts say. On the other hand, training the powerful Chinese model DeepSeek V3 was much cheaper than its American counterparts. DeepSeek said that Huawei successfully adapted the V3 launch to Ascend accelerators and provided developers with detailed instructions on how to use the chips.