DeepSek Gives China Chip Makers Advantage In Cheap AI Competition

JAKARTA The emergence of the DeepSek artificial intelligence (AI) model is expected to provide greater opportunities for Chinese chipmakers such as Huawei to compete in the domestic market against stronger US-made processors.

Over the years, Huawei and other Chinese chip companies have struggled to compete with Nvidia in creating high-end chips capable of training AI models, a process where data is used to help learning algorithms make accurate decisions.

However, DeepSek models focus on "inference", namely when AI models produce conclusions by optimizing computational efficiency rather than relying solely on raw processing power. This is the reason why this model is expected to narrow the gap between China's AI processor and its competitors from the US.

Huawei as well as other Chinese AI chipmakers such as Hygon, Tencent-backed EnFlame, Tsingmicro, and Moore Threads recently announced that their products will support DeepSek models, although not many details have been released.

Open-Source Model

Industry executives predict that the open-source nature of DeepSek and its low cost could increase AI adoption and development of real applications in China. This also helps Chinese companies overcome US export restrictions on high-power chips.

Even before DeepSeek was in the spotlight this year, products such as Huawei Ascend 910B were considered more suitable for inference tasks that did not burden computation too much. Companies such as ByteDance chose it to run AI models that have been trained to make predictions or perform tasks such as chatbots.

In China, dozens of companies ranging from automakers to telecommunications providers have announced plans to integrate DeepSek models into their products and operations.

"This development is in line with the capabilities of China's AI chipset vendors," said Lian Jye Su, chief analyst at technology research firm Omdia. "Chipset AI made in China has difficulty competing with GPU Nvidia in AI training, but the workload of inference is much more flexible and requires a deeper local and industrial specific understanding."

Nvidia Still Dominates

Although China's AI chips are more competitive at costs for inference, Bernstein analyst Lin Qingyuan believes that this advantage is still limited in the domestic market. Nvidia remains superior, even for inference tasks.

Although US export restrictions prohibit the delivery of China's most advanced AI training chip, the company is still allowed to sell underpowered training chips, which Chinese customers can still use for inference.

Nvidia also relies on CUDA, a parallel computing platform that allows software developers to leverage Nvidia GPU not only for AI or graphics, but also for general computing. This makes CUDA a key component of Nvidia's dominance.

Many Chinese AI chip companies do not directly challenge Nvidia by asking users to leave CUDA, but prefer to claim compatibility with the platform.

Huawei is the most aggressive in its efforts to break away from Nvidia by offering a CUDA alternative called Compute Architecture for Neural Networks (CANN). However, experts consider that Huawei faces major obstacles in convincing developers to switch from CUDA.

"The software performance of Chinese AI chip companies is still lacking at this time. CUDA has a rich library and a variety of software capabilities, which require significant long-term investments," said Su of Omdia.