As one steps into the Smart Computing Center of Bodadata located in Qianhai, Shenzhen, the striking blue and white square architecture captures immediate attentionWithin this center, rows of intelligent computing cabinets stand orderly, with high-speed servers diligently engaged in complex data analysis and computation tasks, emphasizing the cutting-edge technology of the digital age.

The Vice President of Bodadata, Bai Xu, elaborated on their operations: "As a converged computing power service provider, we deploy the full version of DeepSeek-R1671B on our own computing clusters for client useBy directly providing AI models to users, we generate market demand for computing power, which in turn boosts sales of our high-density cabinets and computing power services.” This statement underlines a pivotal moment—where artificial intelligence (AI) merges seamlessly into multiple industries, shaping the work environments of countless enterprises.

With the rising popularity of the domestic large model, DeepSeek, a wave of companies across various sectors has announced their deployments of this powerful AI tool, leading to an unprecedented surge in demand for computational powerExperts in the industry have analyzed that DeepSeek is significantly lowering the barriers to AI application, shifting the demand from a "training-dominated" to a "reasoning-dominated" modelAs the landscape of the industry evolves, smart computing centers are urged to transition from merely supplying computing resources to offering specialized services that cover the entire lifecycle of computational power and improve resource utilization efficiency.

A notable surge in DeepSeek deployments continuesCompanies have shown interest in integrating DeepSeek into their operational frameworks, reflecting a robust enthusiasm for harnessing the capabilities of this AI modelFor instance, on February 18, Haikan Co., a company involved in artificial intelligence, announced that it had integrated DeepSeek into its self-developed AI review platform

Advertisements

Similarly, Zhaochuang Information reported that it completed localized deployment of DeepSeek's series of large models, focusing on developing specialized models for bulk commodities using distillation technology.

This enthusiasm surrounding DeepSeek is pivotalGu Licheng, a solutions architect at Zhonghao Chip, highlighted the impactful role of DeepSeek in democratizing AI, allowing individuals across various industries to access and utilize this modelHe noted an uptick in inquiries regarding the compatibility of their smart computing center with DeepSeek since the recent Chinese New Year. "Our smart computing center can now support various distilled versions of the DeepSeek model, and the performance has been excellent," Gu stated, showcasing the model's efficacy.

DeepSeek, as an open-source large model, plays a transformative role in AI applicationWith its high performance and low-cost structure, it makes deploying AI models a feasible option for numerous downstream companies and end manufacturersAn example is Wisdom Bud, a technology innovation service provider that recently integrated DeepSeek-R1. According to the company’s technical lead, the open-source models like DeepSeek genuinely assist enterprises in significantly reducing upfront development costs, particularly in mitigating the steep infrastructure investments usually required during the trial-and-error phase of development.

Interviews revealed that companies not only seek to adopt DeepSeek's large models locally, but there is a strong trend towards creating tailored applications based on specific industry data. "A law firm stands out; they requested assistance in customizing DeepSeek using their proprietary data for developing a unique intelligent application," Bai Xu revealedHe further emphasized that DeepSeek accelerates the digital transformation of many domestic companies, particularly in industries such as government, finance, and manufacturing, where demands for high-density cabinets and edge computing devices are expected to rise significantly due to expanding AI application scenarios (like intelligent approvals and digital twins).

The shift from “training-dominated” to “reasoning-dominated” models marks a new chapter for domestic computing chip manufacturers

Advertisements

One significant advantage of DeepSeek is its affordable training costs, setting it apart from other modelsAlthough the exact costs of training the R1 model have not been disclosed, previous publications indicated that the training cost of the V3 model was only $5.576 million, about one-tenth of the cost for similar models by Meta.

In terms of operational phases, large models are divided into training and reasoning stagesTraining requires extensive data to develop a robust model, typically necessitating substantial computational power and storageIn contrast, reasoning applies the trained model to execute real-world tasks, such as generating text or recognizing images and videos.

Traditional model training resembles a black hole for computational resources, requiring exorbitant investment—often amounting to tens of millions of dollars—creating barriers that only technology giants can crossThe competition for GPU resources globally has heightened the centralization of foundational model developmentYet, as Vice Dean Shen Jiaqing of Shanghai Jinyi Industrial Intelligence Research Institute pointed out, DeepSeek disrupts this status quo by redistributing the demand for computing resourcesThe once concentrated training demands will experience a shift toward the application end, leading to a more diverse allocation of computing power.

"As early as last year, we predicted a shift in the structure of demand for computing resources in smart centers from training to reasoningOnce large models are developed, they must be useful, usable, and user-friendly, and DeepSeek has successfully proven this, sparking a surge in interest," Bai Xu explainedCurrently, collaborations between domestic chip manufacturers and original equipment manufacturers (OEMs) to launch integrated systems for DeepSeek are underway, while major cloud platforms have also begun offering DeepSeek model services, further reducing barriers to the practical application of large models.

In addition to driving demand for reasoning chips, DeepSeek also positively influences the technological landscape for domestic chip manufacturers

Advertisements

As Gu Licheng noted, before DeepSeek's appearance, the domestic market for large models was scattered. "As Zhonghao Xin Ying, a company specializing in AI chip design and domestic computing hardware solutions, we face many adaptation demandsWith the gradual focus on domestic large models, we can expedite the standardization and marketization of our computing hardware platforms and software models," he saidThis highlights the synergy between technological advancements and market needs, paving the way for more tailored solutions in the computing industry.

Since late January, efforts have surged from cloud infrastructure providers to domestic computing chip manufacturers in adapting their technologies to support DeepSeekMajor cloud providers, including Huawei Cloud, Tencent Cloud, Alibaba Cloud, and Baidu Cloud, have announced the availability of DeepSeek models for developersThe three state telecom giants—China Mobile, China Telecom, and China Unicom—have also fully integrated support for DeepSeek, reflecting a collective movement towards enhancing AI capabilities across the industryDomestic chip firms have also jumped in; recent reports indicated that over ten manufacturers, including Tiangong Zhixin and Moore Threads, have officially announced compatibility with DeepSeek model services.

The emergence of DeepSeek heralds an entrepreneurial transformation for smart centersThere is a burgeoning focus on the future demand for computational capabilities amidst the rise of this modelRecently, fluctuations have been observed within the secondary market, particularly in the sectors influencing computing powerShen Jiaqing argued that in the short term, the reduced training costs from DeepSeek may decrease the requirement for large-scale computational resources, guiding investment toward developing practical solutionsHowever, in the medium to long term, the demand for computing power is expected to escalate, driven by the increased need for reasoning and refinement, as well as by the accelerated iterations on foundational models.

There is a general consensus among industry experts that DeepSeek represents a long-term advantage for smart computing centers

Yet, before plunging into this "promised wealth," there are concerns regarding the proliferation of smart computing centers across various regions in ChinaFollowing the AI boom, a plethora of centers has sprung up across the nation, reflecting a staggering 74.1% year-on-year growth in the scale of intelligent computing power in 2024, showcasing China's commitment to deepening its investment in this field.

Nonetheless, some centers grapple with underutilization and mismatched supply-demand dynamicsShen Jiaqing pointed out that the market's rapid supply surge results from policies promoting infrastructure development and the "moderate foresight" principle among major enterprisesDespite the growing demand, the market has not seen skyrocketing growth, primarily due to gaps in product effectiveness and maturity of technology during the industry’s nascent stagesCompanies are often cautious, weighing various factors when considering product purchases or services.

The promise of excessive computing capacity has led many smart centers into challenges, but DeepSeek could provide a remedy. "Just like the surge in mobile internet usage, once large models become habitual tools, an enormous demand for computing power will inevitably arise," Gu Licheng reflectedThe strategic positioning of smart centers is critical; being proactive will enable them to manage the influx of demand when it finally arrives.

Regarding future pathways for smart computing centers, DeepSeek is not merely a mechanism for change but a groundbreaking opportunityBai Xu indicates that the product model of smart computing centers is evolving from simply providing high-density cabinets and computing services to directly deploying model applications for enterprises.

"Smart computing centers need to shift from primarily hardware construction toward a new phase centered on delivering specialized services," asserted Shen JiaqingThe transformation of computing power from a basic resource to a universally applied productive tool is rife with challenges, yet these hurdles also represent market opportunities

Advertisements

Advertisements