The advent of the new year in 2025 marked an era that could be aptly dubbed the "AI Year," as cities across the globe witnessed spectacular demonstrations by robots and drones showcasing advanced capabilitiesHowever, the technology that caught the most attention was none other than DeepSeek, a cutting-edge AI model that has rapidly garnered a global user base.

DeepSeek has emerged as a frontrunner in the AI landscape, particularly when compared to other modelsIn terms of token pricing, DeepSeek-R1 is about 50% cheaper than OpenAI's o3-mini, which certainly contributes to its growing popularityWithin a remarkably short time frame, DeepSeek achieved an impressive daily active user count of 22.15 million, which represents 41.6% of the 53.23 million users of ChatGPT, thereby establishing itself as the fastest-growing AI product in history.

Experts and analysts are confident about DeepSeek's trajectory, with many asserting that it has the potential to become the most widely-used AI tool in the worldUsers have started exploring various applications for the model, including language learning, relationship diagnostics, and even astrology, with some functionalities currently being restrictedNotably, DeepSeek seems to resonate particularly well with the Chinese audience, leading to a burgeoning dependency among its user base.

From the broader perspective of the AI industry's development, the implications of DeepSeek are becoming increasingly apparentOne significant factor is the reduction in computational costs associated with running AI applicationsA professional in the field remarked, “Many tasks that previously required powerful ‘H cards’ (NVIDIA’s H-series GPUs) can now be performed with a standard 4090 GPU.”

Moreover, compared to smaller open-source models, DeepSeek is revealing its considerable commercial potentialHowever, the widespread adoption of DeepSeek has not yet disrupted the market dominance of major AI corporations like NVIDIA, suggesting that the future for domestic GPU products remains a challenging journey.

Wu Jiajin, a professor and doctoral advisor at the Electronic Engineering Department of Xi'an University of Electronic Science and Technology, believes that the emergence of DeepSeek is creating issues for the monopoly established by major American technology firms

Advertisements

Historically, companies such as OpenAI invested heavily in research and development, yet DeepSeek has managed to achieve comparable results with minimal training costsThis indicates that significant investment is not the sole pathway to technological breakthroughs.

A former senior executive at Xiaomi emphasized that the standout feature of DeepSeek lies in its algorithmFrom a technical standpoint, DeepSeek has innovated around the Mixture of Experts (MoE) framework, expanding from 160 experts in its version 2 to 256 in version 3. This development allows the model to efficiently operate with 671 billion parameters while only activating 37 billion parameters, thereby overcoming previous limitations associated with training MoE models.

The direct impact of DeepSeek is evident in its ability to diminish the market competitiveness of American closed-source modelsThe vast number of daily active users, alongside multiple failed attempts to restrict DeepSeek, has ignited a surge in AI talentFollowing a sanctions announcement directed at DeepSeek, it was reported that within 72 hours, the DeepSeek Silicon Valley Research Institute received 327 job applications from leading American tech companies, including 17 chief engineers from Meta.

Wang Yuyuan, the Marketing Director of Qingyun Technology, opined that while ChatGPT was the catalyst for the industry's initial leap, DeepSeek represents another critical milestoneIn her observation, even Meta's Llama has not generated the same level of excitement within the open-source domainImportantly, the low-cost, high-efficiency models introduced by DeepSeek are poised to ignite a new wave of AI developments: “This undoubtedly presents a substantial advantage for AI application developers and individual creators,” she elaborated.

One AI application developer shared with DigitalPlanet his experience running a private deployment of the DeepSeek 7B model on a laptop with only 8GB of video memoryWhile he acknowledged that the performance was not yet optimal, he expressed excitement about the model's potential, stating that the ability to run it on lower-end hardware could alleviate the burden of expensive GPU-related costs.

As he illustrated, he was in the process of developing software for automatic exams and problem-solving that requires real-time computational power to generate test questions

Advertisements

He believes the project demands relatively modest computational resources, and if those requirements can be met, it could become a highly sought-after AI application.

Wang Yuyuan argued that the decline in AI computational costs will lead to an influx of new applicationsShe noted that across various fields, individuals have begun experimenting with generative AI, which is evident in the creative endeavors people have undertaken during the festive season.

Moreover, with the significance of domestic products and private deployment increasing, previously conservative ToB clients in the AI space are becoming more enthusiasticSeveral industry insiders reported that on the first working day of the new year, their companies initiated new AI product procurement schemes.

In conclusion, the explosive popularity of DeepSeek can be viewed as a boon for the AI industry.

Stimulating Development in the Computational Industry

DeepSeek's impact extends well into the computational supply sector, showcasing an impressive “catalyst effect.”

A professional in the computational industry remarked, "Previously, renting a computational cabinet meant a commitment of at least five years." They estimated that at that time, a more affordable computational cabinet might cost around 3000 yuan monthly, leading to a total expenditure of about 180,000 yuan over five yearsOn the other hand, high-end cabinets located in core areas of first-tier cities could cost as much as 12,000 yuan monthly or more, resulting in total costs that might exceed 720,000 yuan in the same timeframe.

Currently, users can opt for billing based on tokens or computational usage, with options to choose specific hardware and rental durationsThis increased accessibility for AI developers, combined with declining operational costs and performance optimization, could potentially lead to the renewed utilization of previously idle computational resources.

One former IDC industry professional disclosed, “The explosive demand for DeepSeek could enhance the operational models of existing intelligent computing centers.” Despite the proliferation of these centers, he noted that many operate below expectations. “I estimate that 80% of domestic computational cards are underutilized, with most rented computational services relying on NVIDIA products,” he explained.

This underutilization is significant, with some centers reportedly employing 50% domestic cards, which were previously ineffective for training and were only capable of handling some inference demands

Advertisements

However, with new operational methods for using computational power, domestic cards could see increased deployment.

Wu Jiajin posited that the reluctance of major model producers to adopt domestic cards stems from a desire to minimize complicated adjustmentsHe argues that model developers typically lean towards stable solutions, allowing them to concentrate on research rather than compatibility issues with GPU manufacturers.

So how was DeepSeek developed? Experts note that it employed PTX, which provides lower-level control compared to CUDA, NVIDIA’s deep learning ecosystemPTX enables developers to finely tune GPU operations, optimizing task allocation and maximizing resource efficiencyIn large-scale distributed training scenarios, it sidesteps the abstraction layers of the CUDA framework, thereby diminishing potential efficiency losses; data indicates a 16% improvement in theoretical computational capability through PTX optimization.

The implication of utilizing PTX amidst the current backdrop of domestic GPU development is that it may facilitate future adaptations of DeepSeek to domestic hardwareDevelopers can follow NVIDIA's programming interface by understanding the core functions provided by domestic hardware, paving the way for software compatibility breakthroughs.

According to Wu Jiajin, the reason DeepSeek opted for PTX technology rather than relying solely on more abundant cards was due to a scarcity of GPU resources: “If there were enough cards, it wouldn't even be a consideration,” he remarked.

From this perspective, the greatest contribution of DeepSeek is that it empowers manufacturers to run models effectively on non-NVIDIA hardware.

Of course, achieving a foothold in NVIDIA’s market will require continuous efforts to enhance the ecosystemThanks to DeepSeek’s influence, users sensitive to costs or looking for autonomous solutions may be more inclined to turn to domestic GPU offerings.

In summary, it is clear that the arrival of DeepSeek can, to some extent, alleviate the fallout of high-end GPU embargoes.

Reimagining AI Hardware with Edge Models

The long-term effects of DeepSeek will be felt within AI hardware development

Advertisements

Advertisements