The emergence of DeepSeek has sent ripple effects throughout the domestic AI chip industry, positioning it at the forefront of technological innovation in ChinaAs of February 13, a total of 17 notable AI chip enterprises such as Huawei's Ascend, MuXi, Tensun Intelligent Chip, Moore Threads, and Haiguang Information have announced their commitment to rapidly deploy and train models based on DeepSeekThis encompasses a vast array of applications, spanning from training sessions to inferential tasks, thus creating a full ecosystem that supports various operational scenarios.

The homegrown AI chip sector has thrived over the past two years, experiencing exponential growth fueled by the global surge in artificial intelligence advancementsAccording to market research conducted by the China Commercial Industry Research Institute, the market size for AI chips in China is projected to reach an astonishing 120.6 billion yuan (approximately $18 billion) in 2023, marking a year-on-year growth of 41.9%. Further estimates predict this figure will escalate to around 153 billion yuan (roughly $22 billion) by 2025, boasting an impressive compound annual growth rate (CAGR) of over 30%.

DeepSeek's introduction is set to accelerate this momentum significantlyResearch insights published by CITIC Securities underline that DeepSeek's cost-effective nature has considerably lowered the barrier for companies aiming to enter the AI sectorThis increase in demand for inferential capabilities is anticipated to boost the deployment of domestic AI chips in real-world applications, thus mitigating China's dependency on foreign hardwareMoreover, Galaxy Securities has expressed optimism regarding the explosive demand for local computational power, alongside the structural uplift of the semiconductor manufacturing supply chain.

The H-share market has responded enthusiastically, with domestic semiconductor firms garnering heightened attention from investorsOn February 12, the only domestic semiconductor Exchange Traded Fund (ETF) tracking the CSI Semiconductor Index surged by 3.39%, exceeding 2 billion yuan in transaction volume, with total trades amounting to approximately 2.01 billion yuan.

In discussing the profound implications of DeepSeek, Luo Guozhao, Director of the Chip Global Testing Center's Chinese laboratory, highlighted that the model enables greater opportunities for the development of domestic AI chips, thereby invigorating the nation's AI ecosystem nurtured by large models.

One of the prominent drivers encouraging many domestic AI chip manufacturers to rally behind DeepSeek is its revolutionary capability to dismantle the high-performance computation prerequisites that have typically encumbered leading models like ChatGPT

Advertisements

Luo elaborated on how the inception of ChatGPT necessitated high-caliber computation chips like NVIDIA's A100 and H100, largely due to NVIDIA's overall edge in terms of computational power, ecosystem support, optimization, and scalabilitySince 2012, the demand for computational power in AI model training has doubled approximately every three to four months, leading to an unprecedented annual growth rate of up to tenfoldNVIDIA's powerful single-card performance and parallel computation abilities adequately satisfy these escalating needs.

However, OpenAI's dependency on NVIDIA's GPUs has inadvertently entrenched a rigid technological and supply chain alignment across the industry, with NVIDIA holding an excess of 90% share in the global AI training market by 2023. In contrast, DeepSeek aims to change this competitive landscape by introducing its innovative modelsFollowing the launch of DeepSeek V3 in December 2024, they introduced a suite of models boasting parameters ranging between 15 billion and 671 billion in early 2025, including the noteworthy R1 and R1 Distill series.

Notably, these advanced models claim to offer cost-efficiency ratios that are 90% to 95% better than their counterpartsThis pricing structure suggests that the latest iterations of DeepSeek can rival the performance of OpenAI’s elite models with only a tenth of the cost associated with GPT-4, and the API usage fees are a mere third of what OpenAI chargesThese versions exhibit notable advantages in inferential capacity, thanks to a series of innovative training methods reinforced through process enhancements.

Analyst Pu Junyi of Dongfang Securities pointed out that while computational demands were previously concentrated in the training phase, where NVIDIA chips excel due to their robust performance and high bandwidth, the inferential phase exhibits different requirementsIn the context of private deployment of AI computational resources, the expectations for computing performance and cluster communication speeds are comparatively relaxed

Advertisements

DeepSeek’s models, with their lower power consumption and less demanding storage requirements post-distillation, align favorably with the properties of domestic AI chips.

This alignment indicates that performance gaps for local computational needs may experience relief during the inferential stage, with attractive pricing strategies likely to stimulate heightened demand for domestic computational chips and server providers based on locally produced hardware.

Moreover, the collaboration with DeepSeek significantly enhances the synergy between hardware and software for domestic AI chip manufacturersLuo emphasized that the industry’s reliance on NVIDIA was not solely because of superior power metrics, but also due to the software 'moat' built around its CUDA architectureThis hardware-software integration is pivotal in optimizing algorithms and bolstering AI execution capabilitiesDeepSeek, through collaborative developments, aims to minimize computational costs, paving the way for local chips to gain a foothold in the industry.

Historically, the fierce competition for local large models led to fragmentation in the technology framework, with various architectures and toolchains for different modelsLuo noted that the introduction of DeepSeek provides a potential solution, generating a streamlined path for development.

Guanxin Securities analyst Hu Jian pointed out that successful integration with national large models would validate and enhance the capacity of domestic GPUs to handle complex AI tasks, ultimately driving growth across the entire AI industry chain in China.

A case in point is the domestic GPU manufacturer Moore Threads, which announced on February 4 its capability to effectively deploy inferencing services for the DeepSeek R1 Distill model within its GPU intelligent computing clusterThe relevant spokesperson explained that the DeepSeek R1 Distill series models have demonstrated a remarkable capacity to extract inferencing functionalities from larger models, achieving impressive results in various benchmark tests

Advertisements

These models have already been deployed successfully on the KUAE cluster, further confirming their practicality and scalability.

The implications for the future are substantial, with DeepSeek allowing the transition of capabilities from large models to smaller, more efficient variants, potentially enabling high-performance inferencing on domestic GPUs.

Another significant trend is the growing support for DeepSeek among major cloud service providers in ChinaCompanies such as Huawei Cloud, Tencent Cloud, and Alibaba Cloud, alongside at least 12 independent cloud and intelligent computation firms, have affirmed their support for DeepSeekThis collective endorsement is widely regarded as a pivotal step towards constructing a cohesive ecosystem for the Chinese AI industry, spanning across hardware and software.

The establishment of a robust ecosystem is essential for the holistic advancement of China's AI industryZheng Weimin, an academician at the Chinese Academy of Engineering and a computer science professor at Tsinghua University, recently remarked that while domestic AI chip manufacturers have made significant strides over the past few years, the immaturity of the software ecosystem remains a key hurdle to the growth of the industry.

Liu Xiaofeng, AI product director at Guokai Microelectronics, stated that the emergence of DeepSeek is beneficial in expediting technological advancements, bringing to a close the current landscape of “chaotic model competition.” This shift will enable domestic AI chip firms to converge their technological paths, granting participants clearer direction and cohesion for product ecosystem development.

Statistics from the domestic AI sector reveal that within just 20 days of its launch, DeepSeek’s application already boasted daily active users exceeding 21.2 millionIndustry expectations posit that DeepSeek might herald a watershed moment for Chinese AI applications by 2025. According to consulting firm Omdia, the market size for generative AI in China is anticipated to grow 5.5 times over the next five years, potentially reaching $9.8 billion by 2029.

CITIC Securities' analysis indicates that DeepSeek, through its technological breakthroughs and open ecosystem, is poised to catalyze rapid advancements across the entire domestic AI industry

Advertisements

Advertisements