DeepSeek gives China's chipmakers leg up in race for cheaper AI

In China, dozens of companies from automakers to telecoms providers have announced plans to integrate DeepSeek's models with their products and operations.

The rise of DeepSeek's artificial intelligence (AI) models is seen providing some Chinese chipmakers such as Huawei a better chance to compete in the domestic market against more powerful U.S. processors.

Huawei and its Chinese peers have for years struggled to match Nvidia in building top-end chips that could compete with the U.S. firm's products for training models, a process where data is fed to algorithms to help them learn to make accurate decisions.

However, DeepSeek's models, which focus on "inference," or when an AI model produces conclusions, optimise computational efficiency rather than relying solely on raw processing power.

That is one reason why the model is expected to partly close the gap between what Chinese-made AI processors and their more powerful U.S. counterparts can do, analysts say.

Huawei, and other Chinese AI chipmakers such as Hygon, Tencent-backed EnFlame, Tsingmicro and Moore Threads have in recent weeks issued statements claiming products will support DeepSeek models, although few details have been released.

Huawei declined to comment. Moore Threads, Hygon EnFlame and Tsingmicro did not respond to Reuters queries seeking further comment.

Industry executives are now predicting that DeepSeek's open-source nature and its low fees could boost adoption of AI and the development of real-life applications for the technology, helping Chinese firms overcome U.S. export curbs on their most powerful chips.

Even before DeepSeek made headlines this year, products such as Huawei's Ascend 910B were seen by customers such as ByteDance as better suited for less computationally intensive "inference" tasks, the stage after training that involves trained AI models making predictions or performing tasks, such as through chatbots.

In China, dozens of companies from automakers to telecoms providers have announced plans to integrate DeepSeek's models with their products and operations.

"This development is very much aligned with the capability of Chinese AI chipset vendors," said Lian Jye Su, a chief analyst at tech research firm Omdia.

"Chinese AI chipsets struggle to compete with Nvidia's GPU (graphics processing unit) in AI training, but AI inference workloads are much more forgiving and require a lot more local and industry-specific understanding," he said.

NVIDIA STILL DOMINATES

However, Bernstein analyst Lin Qingyuan said while Chinese AI chips were cost-competitive for inferencing, this was limited to the Chinese market as Nvidia chips were still better even for inference tasks.

While U.S. export restrictions ban Nvidia's most advanced AI training chips from entering China, the company is still allowed to sell less powerful training chips that Chinese customers can use for inference tasks.

Nvidia published a blog post on Thursday about how inference time was rising as a new scaling law and argued that its chips will be necessary to make DeepSeek and other "reasoning" models more useful.

In addition to computing power, Nvidia's CUDA, a parallel computing platform that allows software developers to use Nvidia GPUs for general-purpose computing, not just AI or graphics, has become a crucial component of its dominance.

Previously, many Chinese AI chip companies did not directly challenge Nvidia by asking users to abandon CUDA but instead, claimed their chips were compatible with CUDA.

Huawei has been the most aggressive in its efforts to break away from Nvidia by offering a CUDA equivalent called Compute Architecture for Neural Networks (CANN), but experts said it faced obstacles in persuading developers to abandon CUDA.

"Software performance of Chinese AI chip firms is also lacking at this stage. CUDA has a rich library and a diverse range of software capability, which requires significant long-term investment," said Omdia's Su.