China’s chipmakers are ahead in the fight for less expensive AI thanks to DeepSeek.
The emergence of DeepSeek’s artificial intelligence (AI) models is perceived as offering certain Chinese chip manufacturers, including Huawei, an enhanced opportunity to compete in the home market against more formidable U.S. processors.
For years, Huawei and its Chinese counterparts have endeavored to rival Nvidia in the development of high-performance chips capable of competing with the U.S. company’s products for model training, a process in which data is inputted into algorithms to enhance their decision-making accuracy.
DeepSeek’s algorithms prioritize “inference,” wherein an AI model generates conclusions, optimizing computing efficiency instead of depending exclusively on raw processing capacity.
Analysts assert that this is one reason the model is anticipated to partially bridge the disparity between the capabilities of Chinese-made AI processors and their more powerful U.S. equivalents.
Recently, Huawei and other Chinese AI chip manufacturers, including Hygon, Tencent-backed EnFlame, Tsingmicro, and Moore Threads, have announced that their products will support DeepSeek models, but limited specifics have been provided.
Huawei refrained from providing a statement. Moore Threads, Hygon EnFlame, and Tsingmicro did not reply to Reuters’ inquiries for additional commentary.
Industry officials currently anticipate that DeepSeek’s open-source characteristics and minimal fees may enhance AI acceptance and facilitate the creation of practical applications for the technology, aiding Chinese companies in circumventing U.S. export restrictions on their most advanced semiconductors.
Prior to DeepSeek’s emergence in the news this year, products like Huawei’s Ascend 910B were regarded by clients such as ByteDance as more appropriate for less computationally demanding “inference” tasks, which occur post-training and involve trained AI models generating predictions or executing tasks, such as in chatbots.
In China, numerous industries, ranging from automotive manufacturers to telecommunications providers, have declared intentions to incorporate DeepSeek’s models into their goods and operations.
Lian Jye Su, a lead analyst at tech research firm Omdia, stated, “This development is closely aligned with the capabilities of Chinese AI chipset vendors.”
“Chinese AI chipsets face challenges in competing with Nvidia’s GPU for AI training; however, AI inference workloads are considerably more accommodating and necessitate extensive local and industry-specific knowledge,” he stated.
NVIDIA REMAINS PREEMINENT
Bernstein analyst Lin Qingyuan stated that although Chinese AI processors are cost-competitive for inference, this advantage is exclusive to the Chinese market, as Nvidia chips remain superior even for inference jobs.
Although U.S. export laws prohibit Nvidia’s most advanced AI training chips from being sold in China, the company is permitted to sell less powerful training chips suitable for inference jobs to Chinese customers.
Nvidia released a blog post on Thursday discussing the increasing inference time as a new scaling law and asserted that its chips will be essential for enhancing the utility of DeepSeek and other “reasoning” models.
Nvidia’s CUDA, a parallel computing platform enabling software developers to utilize Nvidia GPUs for general-purpose computation beyond AI and graphics, has become an essential element of its supremacy, alongside computing power.
Historically, numerous Chinese AI chip manufacturers refrained from explicitly confronting Nvidia by urging customers to forsake CUDA; instead, they asserted that their chips were compatible with CUDA.
Huawei has aggressively pursued the development of an alternative to Nvidia’s CUDA, termed Compute Architecture for Neural Networks (CANN); yet, analysts indicate that it has challenges in convincing developers to forsake CUDA.
The software performance of Chinese AI chip companies is now inadequate. According to Su from Omdia, CUDA possesses an extensive library and a wide array of software functionalities, necessitating considerable long-term investment.