Stepping into the intelligent computing center of Broad Data, located in Qianhai, Shenzhen, one cannot help but notice a striking blue and white rectangular building that stands out prominentlyWithin this intelligent computing center lies a room filled with rows of neatly arranged intelligent computing cabinets, where servers operate at high speeds, engaged in complex data analysis and computation tasks.
“As an integrated computing power service provider, we have deployed the DeepSeek-R1 671B complete version on our own computing power cluster for our clientsBy directly providing AI models to users, we create demand for computing power in the market, which in turn boosts the sales of our high-density cabinets and computing power services,” said Bai Xu, the vice president of Broad Data, during an interview with Securities Times.
The recent surge in popularity of the domestic large model DeepSeek has led to numerous companies across various industries announcing their deployment of this model, resulting in a massive spike in demand for computing power at the application endIndustry experts analyze that DeepSeek has lowered the barrier to AI applications, shifting the demand for computing resources from a “training-dominated” mode to a “inference-dominated” modeIn this rapidly changing industry landscape, intelligent computing centers must accelerate their transformation from simple computing power providers to specialized service providers that offer professional services covering the entire lifecycle of computing power, thus improving the efficiency of computing resource utilization.
Recently, enthusiasm among enterprises for deploying DeepSeek remains robustReporters from Securities Times have noted a flurry of announcements from listed companies, indicating that they have completed related deployments of DeepSeek’s large modelsFor instance, on February 18, Hai Kan Co., Ltd. announced on an interactive platform that it had integrated the DeepSeek model into its self-developed AI intelligent review platform and other vertical models, fine-tuning and optimizing its own platform
Advertisements
On the same day, ZKX Information also reported that it had completed the localized deployment of the DeepSeek series of large models, focusing on creating a large model for the bulk commodity industry based on the DeepSeek-R1 model through distillation technology.
<p“An important aspect of DeepSeek is that it educates and promotes AI technology to the general publicPeople from various industries are becoming familiar with this large model and are eager to try using itThis concentrated access and use have sparked an explosion in computing power demand,” Gu Licheng, a solutions architect at Zhonghao Xinying, remarked in an interview with Securities TimesSince the beginning of this year, many clients have inquired about the compatibility of the company’s intelligent computing center with DeepSeek. “Currently, our intelligent computing center can run various distilled versions of the DeepSeek model, all showing excellent performance,” Gu noted.
As an open-source large model, DeepSeek has significantly reduced the barriers to AI application with its high performance and low cost, becoming the preferred choice for many downstream AI application companies and end manufacturers to deploy AI modelsFor instance, Wisdom Bud, a technology innovation information service provider that recently integrated the DeepSeek-R1 large model, reported that models developed like DeepSeek indeed help companies lower their initial development investment, particularly in saving expensive infrastructure trial and error costs.
Interviews revealed that enterprises not only seek localized deployment of DeepSeek’s large models but many companies are also looking to utilize DeepSeek’s large model as a foundation for customized development based on their industry data to train vertical models or build specialized applications.
"We were particularly impressed by a law firm that wanted assistance in customizing DeepSeek using their own industry data to build a specialized intelligent application
Advertisements
We are currently collaborating with partners to meet the client's needs," Bai Xu mentionedHe expressed that DeepSeek aids in accelerating the digital transformation of Chinese enterprises, where clients in sectors such as government, finance, and manufacturing are likely to see substantial increases in their requirements for high-density cabinets and edge computing devices due to the expanded AI application scenarios, like intelligent approvals and digital twins.
The transition from “training-dominated” to “inference-dominated” has been beneficial for domestic computing chip manufacturersOne of DeepSeek’s core advantages is its low training costWhile DeepSeek has not publicly disclosed the exact training costs for its R1 model, the previously released paper for the V3 model indicated its training cost was merely $5.576 million, about one-tenth of the training cost of Meta’s open-source large model.
According to several interviews, large models primarily consist of two stages: training and inferenceTraining entails utilizing massive datasets to train a large model, usually requiring substantial computational capacity and storage resources; inference involves applying a trained model to real-world tasks, like answering questions, generating text, and recognizing images and videos.
"Traditional large model training is like a black hole for computing power, where single training costs can easily run into tens of millions of dollars, creating technological barriers that only tech giants can overcomeThe arms race in computing power escalates the global competition for GPU resources, resulting in a highly centralized development of fundamental model research," remarked Shen Jiaqing, the deputy director of Shanghai Jingyi Industrial Research Institute, during an interviewHe pointed out that DeepSeek disrupts this established order, leading to a restructuring of computing power demand; the previously highly centralized demand for training computing power is expected to diffuse towards the application end’s inference computing power demand.
“As early as last year, we anticipated that the demand structure for intelligent computing centers would shift from ‘training-dominated’ to ‘inference-dominated.’ After large models are trained, they need to be useful, usable, and easy to use
Advertisements
DeepSeek has successfully achieved that, leading to its popularity,” Bai Xu addedCurrently, domestic computing chip manufacturers are collaborating with Chinese OEMs to launch integrated machines featuring DeepSeek, and major cloud platforms in China have begun offering DeepSeek model services, significantly lowering the entry barrier for large model applications.
Not only has DeepSeek spurred growth in inference chip demand, but it has also brought benefits to domestic computing chip manufacturers at the technology levelGu Licheng asserted that prior to DeepSeek’s introduction, the landscape for domestic large models was akin to a “hundred flowers blooming” scenario. “As an AI chip design and domestic computing hardware platform solutions provider, Zhonghao Xinying is involved in numerous adaptation demands for large modelsWith the gradual focus on domestic large models, this will accelerate our establishment and marketization of domestic solutions for computing hardware platforms and software models,” Gu stated.
Indeed, since the end of January, cloud infrastructure service providers and domestic computing chip manufacturers have been tirelessly integrating with the DeepSeek large modelMajor cloud service providers such as Huawei Cloud, Tencent Cloud, Alibaba Cloud, and Baidu Cloud have announced the listing of DeepSeek-related models available for developersChina’s three major telecom enterprises—China Mobile, China Telecom, and China Unicom—have also fully integrated DeepSeek into their servicesDomestic chip companies have been quick to respond as well; according to incomplete statistics from Securities Times, over a dozen domestic chip companies, including Tianguozhixin, Moore Threads, Haiguang Information, Yuntian Lifang, Birun Technology, Suiruan Technology, MXW, and Kunlunxin, have announced their integration with DeepSeek model services.
The rise of DeepSeek has triggered widespread attention to future computing power demands
In the secondary markets, computing power sectors and relevant stocks have recently experienced significant fluctuations.
Shen Jiaqing believes that in the short term, as DeepSeek lowers training costs, it will lead to a decrease in the demand for large-scale computing power during the training phase, allowing resources to be redirected towards the development of specific solutions. “However, in the medium to long term, DeepSeek will first enhance the demand for inference and fine-tuning computing power—utilizing some previously idle computing resources dedicated to training; second, it will stimulate the rapid iteration of foundational large models, possibly leading to an increase in computing power demand,” he indicated.
Industry experts commonly agree that DeepSeek presents long-term benefits for intelligent computing centersHowever, before DeepSeek can bring about a windfall, there exists a degree of skepticism regarding the surge of intelligent computing centers being built across various locationsIn recent years, driven by the vigorous development of the AI industry, there has been a nationwide rush toward planning and constructing intelligent computing centers, with many large and small centers springing up all over the countryA recent assessment report released by IDC in collaboration with Inspur Information indicates that China’s intelligent computing capacity is expected to grow by 74.1% year-on-year in 2024, highlighting significant investments in this domain.
Nonetheless, some intelligent computing centers face issues related to underutilization and mismatched supply and demandShen Jiaqing analyzed that from a supply-side perspective, in response to national and local industrial planning policy directions, and encouraged by government initiatives, state-owned enterprises and large private companies in telecommunications and IT hardware/software have been developing large-scale intelligent computing centers guided by the principle of “appropriate foresight,” which has led to a market oversupply
Conversely, on the demand side, although market demand has been increasing, there has not been a noticeable explosion in demand. “Firstly, domestic hardware and software still lag behind in performance and applicable fields, leading to less-than-expected application effects; secondly, during the initial stages of industrial development, technology remains immature—enterprises tend to be cautious in purchasing products or services, carefully considering multiple factors,” Shen explained.
Excess computing power has troubled many intelligent computing centers in China, but DeepSeek may help alleviate this issue. “Just like the earlier popularization of mobile internet, when large models truly become tools that people habitually use, it will spark an enormous demand for computing power,” Gu Licheng stated, stressing the importance of preparing intelligent computing centers adequately to handle surges in demand when needed.
For intelligent computing centers, DeepSeek also presents new avenues of development and opportunitiesAccording to Bai Xu, the product model of intelligent computing centers is no longer limited to high-density cabinets and providing computing power services; instead, they can directly deploy model applications for enterprises to use.
“Intelligent computing centers should transition from a focus primarily on hardware construction to a new development stage centered around offering professional services,” argued Shen JiaqingComputing power needs to evolve from being a basic resource to becoming a universally applied productivity tool, and there are still blocks in between—these represent market opportunities for intelligent computing centers. “By offering comprehensive professional services covering the entire lifecycle of computing power, and collaborating with various entities across the industrial chain to provide a holistic solution that includes scheduling, measuring, optimizing, fine-tuning, packaging, maintenance, and secondary development, we can effectively promote the application of computing services across various industries, thereby expanding overall demand for intelligent computing centers in the market,” Shen concluded.
Advertisements
Advertisements