The rise of DeepSeek has garnered significant attention globally, not just for its innovative technology but also for the opportunities it has unlocked in the artificial intelligence domain. Since its launch less than a month ago, DeepSeek has already emerged as the fastest-growing AI application worldwide, with a sharp increase in daily active users. According to statistics from AI product rankings, as of January 31, 2023, the global daily active users of DeepSeek have surpassed 20 million, surpassing ByteDance's Doubao product, while accounting for about 41.6% of ChatGPT's user base.
Despite this rapid ascent, users have often faced issues when trying to engage in frequent and deep conversations with DeepSeek. Instances of server delays are common, frequently resulting in notifications stating, "The server is busy, please try again later." This has led some users to humorously speculate that DeepSeek's model being dubbed R1 is tied to its ability to "run" only once each day, underscoring the challenges of scaling such a promising technology.
On February 6, DeepSeek officially announced the temporary suspension of API service top-ups due to limited server resources. As of the time of writing, this recharge service remains unavailable. AI professionals have noted that their teams had developed AI search functionalities based on the DeepSeek model, but following its explosive popularity, they experienced significant API service delays and response time-outs, which in turn stymied the generation of search results. During the Spring Festival, they resorted to working overtime to shift services to a backup GPT-4o model.
In a larger context, the emergence of DeepSeek marks a "breakthrough" moment for AI technologies, presenting a multitude of promising opportunities for both upstream and downstream industries. Cloud providers and chip manufacturers, armed with substantial computational capabilities, have rapidly mobilized in response.
Leading the charge were major cloud providers, both domestic and international. Notably, early in the Lunar New Year celebrations, tech giants like Microsoft and Amazon integrated the DeepSeek-R1 model into their cloud platforms. Following suit, major Chinese cloud firms such as Huawei Cloud, Alibaba Cloud, Baidu Intelligent Cloud, ByteDance's Volcano Engine, and Tencent Cloud announced on February 1 the availability of the DeepSeek model for developers and corporate clients.
Simultaneously, a number of domestic chip manufacturers have proclaimed their successful adaptation and deployment of the DeepSeek model. Companies like Mutex, Timesys Intelligence, Moer Thread, and Birun Technology have publicly stated that they have completed the necessary alignments with the DeepSeek model. These chip manufacturers either utilized their own computational platforms or formed alliances with downstream AI infrastructure platforms to facilitate the model's deployment.
An industry insider explained that the swift action taken by these cloud providers can largely be attributed to the relatively low costs associated with integrating the DeepSeek model. The model has been trained using NVIDIA GPUs, which many cloud providers have in abundance, allowing for direct and efficient deployment. In contrast, domestic chip manufacturers, given their unique hardware instruction sets, face a more complex task of adaptation and migration, resulting in higher associated costs.
Whether they are cloud providers or chip manufacturers, there is a shared desire to capitalize on the buzz surrounding DeepSeek. As DeepSeek's official API service faced instability, many companies sought to redirect users to their platforms, leveraging their existing computational resources to offer DeepSeek model services. Users who explored these alternative platforms indicated that in many cases, the pricing and response times were adequate, prompting them to consider developing AI applications based on DeepSeek-R1 through third-party services.
Social media platforms also witnessed a surge of promotional materials from various third-party services, claiming to provide seamless usage experiences that circumvented the congestion on DeepSeek's official site. Some platforms even boasted of their integration of "domestic chips + local large models." For instance, Silicon Valley Flow collaborated with Huawei Cloud, deploying the DeepSeek model on their large model cloud service platform. Huawei even embedded DeepSeek-R1 into the purely domestic Harmony version of its Xiao Yi assistant app.
Yuan Jinhui, the founder and CEO of Silicon Valley Flow, disclosed on social media that prior to the release of the DeepSeek-V3 model, founder Liang Wenfeng suggested deploying the model on their platform using least 20 NVIDIA H800 servers. Citing cost concerns, they opted not to pursue that route initially.
Only after witnessing DeepSeek's meteoric success did the Silicon Valley Flow team decide to proceed with the adaptation of domestic chips. This led them to partner with Huawei, and throughout the Spring Festival holiday, they worked diligently, discussing issues at all hours, even holding late-night meetings. Ultimately, on February 1, they successfully launched their DeepSeek model service utilizing domestic chips.
The intersection of DeepSeek and domestic chips presents an intriguing narrative. Differentiating between the stages of training and inference of a large model is critical to understanding how these two elements interplay. During the training phase, a large model is constantly evolving, requiring extensive datasets and ongoing adjustments to internal parameters to recognize patterns. Inference, by contrast, represents the practical application of the model once it has completed its training.
A former AI engineer from a prominent tech firm elaborated that while the training phase demands higher computational power and bandwidth, it typically involves experimenting with various model architectures and operations, usually resulting in a preference for NVIDIA's hardware and its development toolkit, CUDA. The inference phase, having less stringent hardware and software requirements, provides a fertile ground for many domestic chips aiming to optimize and accommodate well-trained models.
Several domestic chip manufacturers asserted that while DeepSeek introduced minor innovations structurally, it remains a large language model. Adapting DeepSeek primarily occurs at the inference application stage, rendering it relatively straightforward and quick to implement.
The spotlight on DeepSeek’s cost-effectiveness resulted in dramatic fluctuations in the stock market, with NVIDIA's shares collectively dropping significantly, a record-breaking loss in market capitalization within a single day. One prevalent theory circulating claimed that during its model development, DeepSeek circumvented the CUDA framework created by NVIDIA, thereby reducing dependency on NVIDIA processors. This claim originated from the assertion in DeepSeek's V3 model technical report that they implemented tailored PTX (Parallel Thread Execution) instructions while automatically tuning the communication block sizes, significantly decreasing L2 cache usage and minimizing interference from other streaming multiprocessors (SMs).
Nonetheless, assertions that utilizing PTX programming language signifies a breach of NVIDIA's CUDA dominance are flawed. Industry analysts clarify that PTX is, in fact, a component of CUDA and does not evade it. They explain that CUDA is a comprehensive software suite encompassing higher-level programming languages, an extensive collection of API libraries, and compilation tools designed for GPU programming. PTX functions as an intermediate assembly language within CUDA, closer to hardware and not typically aimed at developers. Consequently, developing within CUDA generally limits control over the GPU's finer aspects. However, employing the PTX language allows for greater flexibility in managing underlying hardware and optimizing performance, which might explain why DeepSeek requires less computational power.
Although the DeepSeek model remains trained on NVIDIA GPUs, its demonstrated efficiency in utilizing computational resources, combined with the resulting trend of adapting domestic chips, signals significant benefit for the chip industry.
Prior to this, domestic large model companies utilized domestic chips for certain model inference or testing training tasks, albeit on a limited scale. However, spurred by the DeepSeek phenomenon, the utilization rate of domestic chips is set to rise dramatically.
Is the year for practical AI applications truly upon us? The waves generated in the upper and middle sectors will inevitably cascade downwards, and with the proliferation of the DeepSeek trend, numerous industries, including smart hardware, automotive, and finance, have begun eagerly integrating the DeepSeek model into their services. In the past week, the Reading Group announced that its writer assistance product, "Writer Assistant," has integrated the DeepSeek-R1 model, marking DeepSeek's initial foray into the online literature space. They noted that when utilizing the intelligent Q&A feature designed to assist writers with research and inspiration, DeepSeek showcases a superior understanding of intent, capable of interpreting subtext and implied meanings.
Furthermore, the extended reasoning capabilities exhibited by the R1 model serve as a noteworthy source of inspiration for online literature authors. "Online literature writers, particularly seasoned ones, frequently express dissatisfaction with the cliched and repetitive outputs of AI content. What they require are insights and a means to organize their thoughts," Reading Group stated. Following the integration of DeepSeek, when authors request AI-generated outlines incorporating trending elements from a particular website, the AI not only provides a generated answer but first identifies and clearly lists relevant elements, even offering popular titles in that genre to assist writers in accessing the necessary specialized content.
Amid the competitive pressure induced by DeepSeek, OpenAI recently announced the public release of its latest o3-mini series, which also features reasoning chains. However, their researchers noted that while these reasoning summaries closely approach the original chains, they do not replicate them exactly. Analysts have suggested that OpenAI's decision might stem from considerations regarding user experience, privacy protection, output quality, technical costs, and trade secrets, aiming to offer valuable insights without adverse consequences.
Back in May of the previous year, DeepSeek's low pricing sparked a "price war" among domestic AI large model providers. Many within the industry believe that decreasing the cost of large models will propel applications into reality. The pricing advantage associated with DeepSeek's recently unveiled models remains staggering: Although the promotional period for the V3 model's favorable pricing concluded on February 9, the cost for API calls still remains merely a fraction of GPT-4o. In contrast, the pricing of the inference model DeepSeek-R1 is as much as 27-55 times lower than its counterpart, the o1 formal version.
Silicon Valley Intelligence, focusing on AI digital avatars and Silicon-based smart screens, has expressed optimism regarding DeepSeek's open-source model and disclosed reasoning chains, viewing it as an opportunity to enhance their AI digital avatar capabilities and service offerings. During the Spring Festival, their team acted swiftly to incorporate the DeepSeek model, thereby augmenting their digital avatar product offerings with improved natural language understanding and emotional recognition capabilities.
On February 10, Silicon Valley Intelligence, in collaboration with the computational power firm Huakun Zhenyu, unveiled a new solution. This integrated Silicon Valley Intelligence's proprietary AI digital avatar engine with the Kunpeng and Ascend clusters as a foundation for high-performance domestic computational power, endowing the DeepSeek model with outstanding response speeds and stability under massive data processing.
The domestic AI industry chain is poised for a challenging period, especially through this Snake Year Spring Festival. However, the ripples set off by DeepSeek hint at potential waves yet to come, suggesting the growth of this sector may soon form a much larger tsunami of innovation and application.