On March 21, the Star Daily reported (reporters Huang Xinyi and Mao Mingjiang) that since the beginning of the year, the open-source large model DeepSeek has exploded in popularity, like a giant stone thrown into a lake, creating ripples and directly igniting a frenzy of intensive product launches for large model integrated machines. In no time, players from server manufacturers, Cloud Computing Service providers, industry application developers, and large model technology suppliers have entered the scene. According to IDC Statistics, nearly a hundred manufacturers have swiftly launched AI integrated machine products in the market. However, from its initial explosive popularity to truly being usable and effective in enterprises, the DeepSeek integrated machine still has a long way to go.

▍Surge in Demand: An increase of over 200%!

Recent visits by the Star Daily to Cloudwalk Technology, Cloud Tianlfei, Yoke Technology, and Volcano Engine revealed that customer demand for the DeepSeek integrated machine is extremely hot. "The demand for integrated machines is very high now, hospitals, schools, governments, and financial sectors are all looking for us," revealed a business staff member from Huawei's computing division.

Luo Yi, Vice President of Cloud Tianlfei, stated, "The demand for Hardware related to intelligent computing power is indeed quite substantial. Compared to the customer density brought by ChatGPT in the previous two years, the demand from the government and central state-owned enterprises has more than doubled. One point that government customers recognize about the integrated machine is that it does not require a large amount of preliminary preparation work for deployment, which can lower the AI application threshold. In fields with a good data governance foundation, such as government affairs and finance, various applications can be developed quickly."

Yang Hua, Secretary of the Board of Directors of Cloudwalk Technology, also pointed out that customer demand is showing a significant growth trend, especially among enterprises with high demands for data security and privacy protection, who show a strong interest in localized deployment of integrated machines. "According to our recent business leads and sales situation, clients in fields like finance and government affairs have particularly strong demand for AI integrated machines. For instance, we are working on a project with a Bank, which involves using the integrated machine to integrate multi-source data (financial phase three platform, employee credit reporting, compliance systems, etc.), constructing a risk data mart and knowledge graph to achieve automation of risk identification and standardization of accountability."

A relevant person from Volcano Engine stated that after the launch of their AI integrated machine, they have already received a large number of inquiries and orders from enterprise customers, with demand exceeding expectations. Li Tianpeng, Chief Architect of the Emerging Industry Division of Yoke Technology, introduced that the integrated machine primarily targets government, central state-owned enterprises, Medical, Finance, Education, and other application scenarios requiring high data security and industries sensitive to latency.

In terms of pricing, the DeepSeek integrated machine's one-time delivery cost ranges from hundreds of thousands to millions depending on the models it carries and the Software it includes. Cheng Yin, a research manager at IDC China, believes that DeepSeek has driven optimistic development expectations for AI applications in the entire Chinese market; if it can drive more enterprises (especially small and medium-sized long-tail enterprises) to deploy AI, the integrated machine market will welcome further development.

However, under this wave, some industry insiders have expressed concerns. An IT industry insider stated: "It is necessary to be cautious as the current market is a bit overheated. Almost all government departments are raising demands, and everyone is following suit. Customers need to consider their actual situations. First, it is essential to clearly recognize that although the capabilities of current large models have improved, there is still a gap between depth and business integration, which depends on the user’s data governance and business sorting capabilities."

▍Ideal vs. Reality: Long Order Cycle

Although there is a booming demand for consulting on the DeepSeek integrated machine, the actual ordering cycle is long. Multiple server manufacturers stated to the Star Daily reporter that there are many inquiries about the DeepSeek integrated machine, but they are mostly in the research phase. Customers who directly place orders are mostly those who had already set a related budget and chose to replace their previously determined large model orders with DeepSeek.

"Many customers are in the testing phase; those willing to directly place an order to try one unit usually have a budget already. This has a significant impact on established large model companies. Those who initially planned to purchase closed-source large models are now switching their budgets to deploy open-source models with the arrival of DeepSeek." said a representative from a manufacturer.

Li Tianpeng pointed out that the order volume ratio for the DeepSeek integrated machine is lower compared to mature Cloud Computing Services products. "Although customers' demand willingness is very strong, integrated products generally need to be tested and evaluated on-site to confirm effectiveness before signing contracts. Currently, due to resource constraints, hundreds of users are in the testing scheduling phase, resulting in temporary lower order volumes compared to Cloud Computing Services products."

"Many users need to deeply consider the integration of DeepSeek with their own business scenarios. After consulting with integrated machine manufacturers, assessments have to be made on various aspects such as business, procurement costs, and access methods. The ways DeepSeek can be integrated are also very diverse, with some customers turning to API or cloud resource delivery methods. When consulting, users mainly focus on the implementation of DeepSeek in their application scenarios and hope the manufacturers provide suggestions based on their experience. The hardware configurations of integrated machine products in the market are highly homogenized, with little performance difference, so users pay more attention to application scenarios, value-added features, and subsequent services." Li Tianpeng further explained.

Regarding the implementation of integrated machines, Luo Yi suggested, "Using Cloud Computing Services, first validate and test through simulation data. Then confirm that this business can be closed-loop and embedded into the business flow, and then choose a suitable computing architecture to integrate smart computing into the IT architecture of enterprises and governments. The integrated machine is a way to experience it within an acceptable cost."

Currently, the procurement cost of the full version of the DeepSeek integrated machine is over one million, which is a significant investment. However, considering the actual implementation of business, most practitioners recommend users to directly deploy the full version. "The 32B large model can run on NVIDIA's 4090 chip. Customers using 32B and 70B models are mostly for testing and fine-tuning; the real business must be implemented using the full version." said a server sales representative.

Luo Yi also suggested using a full-powered model for business exploration. "This does not mean that these small-sized models are useless, but when exploring a new business, it is essential to use the strongest model to run the entire business through a closed loop before considering using low-cost models to fill in different business segments."

▍Challenges of Implementation: The Gap Between Large Models and Business Scenarios

There is still a significant gap between large models and business implementation. Luo Yi believes that, "The relevant technical talent reserves in the government and enterprises are severely insufficient. On one hand, this requires knowledge dissemination; on the other hand, the integration of large models with business is a gradual process that requires the organization of business flows and even the orchestration of those flows, relying on deep participation from business experts."

Li Tianpeng from UCloud also pointed out that talent is a major challenge. After companies purchase privatized large model integration machines, software and hardware operations and system secondary development require professional talent support, facing long-term challenges in talent accumulation and technological precipitation. In terms of business scenarios, integrated machines are not just hardware businesses; as the application of large models matures, they will become core components for clients, requiring a complete software ecosystem, such as addressing issues of security, permission management, and optimization for integration with business.

A representative from Volcano Engine stated that challenges exist in technical complexity, cost, data quality, system integration, talent shortage, security, business adaptability, operation and maintenance, and ROI. "For example, AI integrated machines involve the integration of hardware, software, and AI models, which has a high technical threshold and requires regular updates (for instance, supporting future upgraded models of DeepSeek), necessitating manufacturers to provide comprehensive maintenance services. When the performance of AI large models does not meet expectations in actual business scenarios, companies not only need to equip specialized algorithm engineers, but also require AI integrated machine manufacturers to provide software-level tools for model compression, Quant, and distillation to help optimize model performance; automated tuning tools to assist companies in quickly finding optimal parameters; and professional AI consulting services to help companies solve performance bottlenecks."

Yang Hua, Secretary of the Board from CloudWalk Technology, summarized that there are still two major mountains to conquer for the implementation of large model integration machines: the first is the deep water area of scenarios. Clients do not want a "universal model," but rather solutions that penetrate deeply into the business's capillaries; for instance, certain industrial quality inspection recognition precisions must reach 99.9%, which requires deep integration of industry know-how with AI; the second is ecological fragmentation. Currently, the multiple hardware combinations of integrated machines, especially in terms of domestic products, have non-unified hardware standards and high hardware adaptation costs, making it feel like assembling "LEGO" for companies when they buy integrated machines, leading to significant post-maintenance headaches.

▍Future Trends: Public Cloud and Privatization Deployments Will Run in Parallel

Currently, nearly a hundred manufacturers have quickly launched AI integrated machine products in the market. Facing fierce market competition, Li Tianpeng, Chief Architect of UCloud's Emerging Industry Division, believes that different manufacturers have different advantages; hardware manufacturers excel in hardware construction costs, hardware adaptation optimization, and localization; cloud service and computing power service providers excel in end-to-end delivery, initial client interfacing and testing, model management, as well as computational scheduling, and the subsequent in-depth integration of large models with user business, thus providing clients with a better user experience.

Yang Hua, the secretary of the board at Yuncong Technology, stated that in terms of trend determination, with the emergence of Deepseek, costs will decrease, and the capabilities of large models and problem-solving abilities will improve, leading to more ecosystems brought by open source. Considering the demand for domestic safety and localization, the need for private deployment will be fully satisfied, and the comprehensive digital transformation of state-owned enterprises and government governance will accelerate, releasing demand and facilitating implementation. In addition, the demand for digital employees and corporate Agents will see explosive growth. This transition presents tremendous opportunities and growth space for technology commercialization companies that focus on vertical industry fields and understand both AI and business.

Yang Hua expects that in the public cloud sector, inference costs are expected to decrease by tenfold each year, which accelerates the adoption of public cloud services by small and medium enterprises and consumer-grade applications. From a market share perspective, public cloud primarily focuses on standardized services, likely holding a high share in areas such as the Internet, Education, and general retail. In terms of private deployment, high-sensitivity industries such as government enterprises, finance, manufacturing, and energy will remain dominated by integrated machines, focusing on vertical fields such as government enterprises, Medical, and Industrial. In terms of technology accessibility, open-source models and low-cost chips will further accelerate the penetration of large model applications.

"In the next three years, the market trend for large models in the public cloud and private deployment sectors will show a pattern of dual-track parallel and coordinated development. Public cloud deployment will continue to grow rapidly, especially in industries such as the Internet and e-commerce that have high demands for elasticity and rapid iteration. Private cloud deployment will also maintain rapid growth, quickly popularizing in industries like government enterprises, finance, and Medical where there are high requirements for data privacy and security," said a relevant person in charge of Huoshan Engine.

The mixed expert system MOE used by DeepSeek reduces the amount of large model parameters that need to be activated in each instance, significantly lowering inference costs and also bringing benefits to domestic chips.

"DeepSeek presents a huge opportunity for domestic intelligent computing chips and manufacturers. Of course, its form is not limited to integrated machines. The future will inevitably be a mixed computing architecture, and the trend of application scenarios being marginalized is very clear, considering issues such as private data security and low latency. Besides the government and large enterprises, the future implementation of intelligent computing at the edge will surely extend to small and medium enterprises, and even towards households and individuals," stated Luo Yi. "Our chips use a computing power building block architecture, which, despite being influenced by the process, can be flexibly assembled and freely expanded like building blocks through innovative architecture."

Yang Hua indicated that the deployment costs of traditional large models (like GPT-4) are high, while open-source models such as DeepSeek will reduce training costs to one-tenth, with future inference costs expected to drop by several magnitudes. The DeepSeek integrated machine supports real-time inference on single machines, with low marginal costs, suitable for flexible application by small and medium government units. "Subsequently, as we collaborate with ecological partners such as Huawei, Haiguang, and Cambrian to create scale effects, it will help further reduce hardware and deployment costs."

It should be noted that domestic computing power still has a very long way to go. "Currently, costs are still somewhat high, with vast room for optimization and improvement. Adapting domestic computing power to advanced models and enhancing cost-performance ratio requires a process," Luo Yi admitted.

Li Tianpeng also mentioned that optimization for computing power adaptation, especially regarding domestic computing power, currently has room for improvement in adapting effects and performance output when comparing domestic GPUs with NVIDIA GPUs; the GPU ecosystem also needs further optimization.