In recent months, the landscape of artificial intelligence (AI) has undergone remarkable changes, particularly with the advent of DeepSeekAs this transformative force drives down costs across various models—both open-source and proprietary—the gap among AI chip manufacturers is narrowing dramaticallyThis shift has sparked a competitive race among Chinese AI chip companies to adapt to different models offered by DeepSeek, a dynamic that is reshaping the industry.

Beginning around February 1, leading AI chip manufacturers in China began announcing their collaborative efforts to adapt to various models under the DeepSeek umbrellaAccording to incomplete statistics, at least 20 Chinese firms are now actively involved in this adaptation processSuch widespread engagement highlights a significant movement within the Chinese tech sector, marking a concerted effort to keep pace with global innovations.

The AI chip market encompasses various types of chips, including CPUs, GPUs, ASICs, and FPGAsAs the demand for large-scale parallel computing surges within the AI domain, the demand for GPU chips has seen a rapid increase, propelling companies like Nvidia to unprecedented levels of performance and stock pricesHowever, DeepSeek's emergence symbolizes a transformative shift toward reducing costs tied to AI inference, prompting the emergence of broader application markets.

This trend suggests a broadening of opportunities for chips beyond just GPUsChips like ASICs and FPGAs, which possess specialized advantages in AI inference, are also poised for substantial growthMany industry insiders believe that Chinese chip manufacturers have a unique capacity to solidify their foothold in the AI inference sector, potentially allowing them to capture some market share from Nvidia.

Nonetheless, a key question lingers: how will Chinese chips adapt in a space where Nvidia GPUs and its CUDA ecosystem have long dominated? Will this adaptation catalyze pressure on Nvidia's market stronghold? Such queries have become focal points of industry discussions.

Since the beginning of February, a flurry of activity among Chinese AI chip manufacturers has unfolded, with various companies announcing successful adaptations to different specifications within the DeepSeek ecosystem.

For instance, on February 2, Gitee AI announced the rollout of four variants of the DeepSeek R1 model—1.5B, 7B, 14B, and 32B—deployed on its Muxiyiyun GPU cloud platform

Advertisements

Just a few days later, on February 5, Gitee AI noted that the full-fledged DeepSeek V3 model (671B) successfully ran on the Muxiyun training and inference integrated GPU, making this version publicly available on their platform.

Similarly, on February 4, TianShu Intelligent Chip indicated that, in collaboration with Gitee AI, they completed adapting the DeepSeek R1 model in just a day, enabling services for the 1.5B, 7B, and 14B large model specificationsBy February 9, they announced that several different model specifications, including DeepSeek R1-Distill-Qwen and DeepSeek R1-Distill-Llama models, had also been made available across major platforms.

On February 6, SuiYuan Technology declared successful adaptation of the entire range of DeepSeek models, incorporating the native DeepSeek R1/V3 with 671B parameters along with various distilled modelsAcross these developments, the emphasis on 'adaptation speed' has emerged as a pivotal metricDistilled models, often featuring fewer parameters, were prioritized for adaptation, while more complex MoE models evidently take additional time.

This rapid adaptation informs observers about the ambition of Chinese AI chip manufacturers to validate their capabilities and responsiveness within the AI ecosystem.

Comparatively, Nvidia's GPUs have dominated the global market, exhibiting monopolistic characteristicsThis dominance is underpinned by three significant protective barriers: hardware GPU chips, the software CUDA ecosystem, and the NVLink connectionIf Chinese chips are to accelerate their development and market penetration within the GPU realm, creating a robust ecosystem is essentialThe extent to which this ecosystem is developed will heavily influence the capacity of AI chips to be fully utilized and adopted within various applications.

Nonetheless, building such an ecosystem is a daunting feat, as the CUDA ecosystem has been maturing for over a decadeChinese manufacturers are taking varied approaches: some are opting for proprietary architectures to build ecosystems starting with vertical applications, while others are focusing on compatibility with the established CUDA ecosystem.

For instance, Haiguang Information has indicated that its DCU chips, which utilize a GPGPU general-purpose acceleration architecture, can directly run DeepSeek models without extensive adaptation, with the technical team's focus primarily on accuracy validation and performance optimization.

As stated by an industry expert, "The rapid adaptation of many Chinese AI chip manufacturers to DeepSeek's technology marks a significant step toward internationalizing Chinese chip development." DeepSeek's partnership offers tangible benefits to Chinese manufacturers, allowing for accelerated adaptation of deep learning frameworks and distributed training, ultimately pushing toward a self-contained ecosystem of "Chinese computing power + major Chinese models."

Historically, the chief challenge for China's AI chips has been Nvidia’s dominion over AI training chips with its CUDA ecosystem

Advertisements

Yet, DeepSeek's introduction has disrupted this paradigmBy utilizing model distillation techniques and optimizing algorithms efficiently, DeepSeek has significantly lowered the computational requirements for modelsThis innovation, along with features like expert mixture systems and core components such as multi-head potential attention mechanisms and RMSNorm, facilitates high-performance operations with lower computational costs.

With the momentum created by DeepSeek, other tech giants such as OpenAI, Doubao, and Baidu have corroborated the trend of declining inference costsThe substantial drop in DeepSeek's training expenses has shattered the traditional view linking high training costs with superior model product performanceConsequently, industry focus has shifted from the traditionally fixed lower limits of training processes to the newfound upper possibilities afforded by inference capabilitiesFor downstream industries, even players with medium computational power can enhance performance thanks to DeepSeek’s contributions.

Traditionally, Nvidia GPUs were predominantly employed for training large AI modelsAs we transition into the inference stage, application developers are increasingly eager to create their own specialized AI inference chips, often in the form of custom ASIC chips to suit their requirements.

Major cloud service providers, including Google, Meta, and Amazon, have made strides in recent financial reports by highlighting the advancements in their proprietary inference chipsFor example, Google's TPU Trillium series streamlines search engine optimization, while Meta's MTIA series bolsters social algorithms and ad distribution.

According to TrendForce analyst Gong Mingde, there is an expectation that DeepSeek will drive cloud service providers (CSPs) to invest more vigorously in low-cost custom ASIC solutions, shifting their focus from AI training to AI inference—a trend likely to rise to an expected 50% market share by 2028.

In this evolving context, there is potential for growth in the development of AI inference chips across various sectors in China, including automotive, e-commerce, and infrastructure-related industries

Advertisements

Advertisements

Advertisements