GuShiio.com Wen| Semiconductor industry, author| fengning
Recently, a new top talent has emerged in the technology circle, called the DeepSeek All-in-One Machine.
How popular is it?
The circle of friends keeps brushing the screen, and some people tease: I don’t have the DeepSeek all-in-one machine, and I’m embarrassed to say that I understand technology.” rdquo;
Why can a device become popular?
DeepSeek all-in-one machine, why did it become popular?
The all-in-one machine is an integrated computing device specially designed for the application and deployment of large artificial intelligence models. It usually includes a central processing unit (CPU), graphics processing unit (GPU), memory, operating system, AI platform software and various model algorithms. Soft and hard components.
Simply put, the all-in-one machine can be regarded as a machine like a computer or mobile phone. When purchased by enterprises/individuals, they can use it directly out of the box to help enterprises and individuals meet their AI application needs.
The popularity of the DeepSeek all-in-one machine is like a counterattack script in the entertainment industry.“” In the past few years, many manufacturers have launched all-in-one machines, but the market response has been relatively flat. So, why did the DeepSeek all-in-one machine quickly become popular as soon as it was unveiled?
The fundamental point is that it meets current needs and is a high-quality product that conforms to the times.
The early launch of all-in-one machines was mainly to meet the model needs of large model factories. Quoting the analogy of an AI engineer in the industry: This is like the activity launched by traditional mobile phone operators to buy phone bills and send mobile phones. The all-in-one machine is equivalent to the mobile phone in our metaphor, and the big model is the analogy to phone bills.In this way, consumers seem to be buying all-in-one machines, but in fact they are buying the services subsequently provided by the large model factory.
Analyze from the perspective of business models, during the period when the model is closed, the commercial closed-loop model deployed in the private domain of large model factories is:Make profits by selling services, and then invest the earned money in model updates, thereby enhancing users ‘stickiness to products, and then continuing to sell services, forming a closed business loop.However, this model has obvious drawbacks, and a simple summary is expensive in one word. Moreover, there are still many opaqueness in the entire model, such as performance opaqueness and price opaqueness. Nowadays, models are moving towards open source, and business models have undergone tremendous changes.
DeepSeek means that not only has an open source + easy-to-use model, but the business model of the all-in-one machine has also undergone tremendous changes. Under this new business model, hardware factories are responsible for selling hardware, and users can DIY their own all-in-one machines and then install a set of universal software. Based on the low-level software that is easy to use, an ecosystem is gradually built. After the hardware factory makes a profit, it will continue to invest in research and development to improve hardware performance; after the software factory makes a profit, it will work hard to improve the ease of use of the software.
With this topic, let’s talk about what the DeepSeek all-in-one machine brings?
The core selling point of the DeepSeek all-in-one machine is its powerful AI computing power.It is equipped with the latest AI chips and can efficiently process complex tasks such as images, voice, and data. ldquo; Performance overpowers traditional equipment, which is the essential difference between DeepSeek all-in-one machines and traditional all-in-one machines. For users who need high-performance computing, the DeepSeek all-in-one machine is no longer a useless thing, but an artifact.
Secondly, the all-in-one machine also has many application advantages such as efficient energy consumption management, strong scalability, intelligent operation and maintenance management, and extensive industry applications.
So, who likes such a hot product?
Central state-owned enterprises are important customer groups
Behind the intensive release of all-in-one machines is the surge in demand from party and government agencies and customers in key industries to introduce DeepSeek.
Party and government agenciesGovernments in Guangzhou, Shenzhen, Hohhot, Wuxi and other places have officially announced the completion of the deployment of the DeepSeek series of large models to help optimize service scenarios in many popular government areas such as government office, urban governance, and people’s livelihood services.
For example, in Futian District, Shenzhen, AI digital intelligence employees based on DeepSeek have been launched. In addition to meeting DeepSeek’s general capabilities, they can also combine the actual business processes of each department and unit to tailor personalized agents. The first batch can meet the needs of 240 business scenarios, greatly improving the correctness and efficiency of government processing.
Central and state-owned enterprises, all three major operators have connected to DeepSeek. Sinopec has completed the first batch of localized deployments of DeepSeek by central enterprises, significantly improving the efficiency of the company’s petrochemical-related work. In order to protect sensitive data, party and government agencies and users of central and state-owned enterprises prefer privatization deployment.
At the same time, relevant data shows that more than 20 central enterprises have connected to DeepSeek, covering energy, communications, automobiles, finance, construction and other fields.
The emergence of the DeepSeek all-in-one machine is a perfect fit with the needs of party and government agencies/central state-owned enterprises. On the one hand, the DeepSeek all-in-one machine can lower the threshold for computing power use; on the other hand, government agencies and central state-owned enterprises often involve citizen information, government data, national security, etc. A large amount of sensitive data information, the all-in-one machine can provide reliable and worry-free solutions for it.
Today, the number of central state-owned enterprises has reached 460,000. At present, 23% of central enterprises have deployed large models, and the penetration rate of large models is expected to increase further in the future. It is assumed that the number of central state-owned enterprises using the large model will reach 35%, 50%, and 70% in the next three years. In the future, it is estimated that 80% of the central state-owned enterprise model use scenarios will need to be privatized and deployed.
According to calculations by Zheshang Securities, as DeepSeek’s rapid deployment demand increases, the market demand for all-in-one machines is expected to see significant growth. It is estimated that the demand for all-in-one machines will reach 150,000 units, 390,000 units and 720,000 units respectively from 2025 to 2027. The market space of DeepSeek all-in-one machines in central state-owned enterprises is expected to reach 1236, 2937, and 520.8 billion yuan. The market is expanding rapidly and the prospects are very broad.
DeepSeek all-in-one machine, intensively updated
As the influence of the DeepSeek model in the industry continues to rise, many companies are turning their attention to relevant integrated solutions and embarking on layout.
As shown in the following table, companies involved in cloud computing, smart computing, data centers and other fields have successfully launched all-in-one products suitable for DeepSeek through independent research and development or close collaboration with partners.
Although many of the above-mentioned companies have launched all-in-one products, in-depth analysis shows that there are many differences in hardware selection, adapted industry scenarios and adaptation models.
First of all, from the perspective of hardware selectionShengteng chips have unique advantages in computing performance and energy efficiency ratio, and can efficiently provide parallel computing capabilities for deep learning tasks. It can be seen from the table that Huawei, the three major operators, Digital China and many other companies have relied on Shengteng chips to build a strong computing power foundation for the DeepSeek all-in-one machine.
It is reported that more than 80 companies have quickly adapted or launched DeepSeek series models based on Tengteng and provided services to the outside world. It is expected that more than 20 companies will complete the launch in the next two weeks. This means that 70% of domestic companies will move closer to DeepSeek based on Shengteng. Compared with imported GPU solutions, Shengteng’s localized services and teams have a significant impact on the effectiveness of deploying DeepSeek. Taking a thousand-card-sized data center as an example, the automatic parallelization function of the MindSpore tool chain reduces the amount of distributed training code by 70%.
With the intensive release of DeepSeek all-in-one machines, Shengteng is deeply benefiting.
In addition, companies such as Xinzhi Software, Zhongke Shuguang, and Lenovo Group have chosen domestic chips such as Haiguang and Muxi as the cornerstone of computing power. At the same time, some companies are also actively exploring cooperation with domestic chip manufacturers on the basis of adopting common international brand chips.
Judging from the adapted modelSome companies rely on their own technical strength and resource advantages to successfully adapt the full-blood version model in the all-in-one machine, bringing users a more powerful performance experience. Huawei is one of them. Its DS version of the FusionCube A3000 training/promotion Ultra (full blood version) of the ultra-fusion all-in-one machine can perfectly adapt to the DeepSeek – R1 and V3 671B models. Relatively speaking, the all-in-one machines released by some companies have not yet clearly stated that they support the full-blood version of the DeepSeek model.
From the application scenarios,Relying on their advantages in communication networks and data resources, the three major operators have launched the DeepSeek all-in-one machine that has outstanding performance in the government field. Changjiang Computing’s DeepSeek all-in-one machine is specially optimized for data processing and other scenarios in the financial industry.
How to choose a suitable DeepSeek all-in-one machine?
Want to buy a full-blooded version of the all-in-one machine? Fei also
For DeepSeek all-in-one machines, configuration and price are the focus of enterprises. How to choose a suitable DeepSeek all-in-one machine also starts from these two aspects.
From a configuration perspectiveThe full-blooded version of the DeepSeek all-in-one machine is based on 671B parameters (such as R1/V3 model), supports ultra-long context understanding, and functions cover complex reasoning, code generation (LeetCode problem pass rate of 92%), scientific paper framework generation, etc. In terms of hardware requirements, it requires professional servers (such as dual H100 GPUs +1TB memory or 8-card A100 clusters), which is suitable for enterprise-level deployments. In terms of security, it supports localized deployment, and data does not need to be leaked out, meeting high security needs in medical, government and other fields.
The distillation version has a parameter scale of 1.5B to 70B parameters, and its functions focus on basic tasks (such as Python script writing and literature abstract translation). Its performance is only about 1/10 of that of the full-blood version. In terms of hardware requirements, it can be run on a single-card RTX3090 or a home PC, and version 1.5B (such as the MNN framework) can also be deployed on mobile phones. In terms of applications, it is suitable for lightweight needs, such as personal learning assistants, content creation, customer service conversations, or low-cost AI integration for small and medium-sized enterprises.
From a price perspectiveThe price of purchasing an all-in-one machine depends on factors such as the quantity required, the type of AI chip, the scale of concurrency, and the degree of secondary development.
The full-blood version requires a high-performance GPU or a dedicated AI chip, and the hardware cost is naturally higher.The full-blood version is subdivided into native full-blood version (FP8 calculation accuracy), translated full-blood version (BF16 or FP16 calculation accuracy), and quantified full-blood version (INT8, INT4, Q4, Q2 calculation accuracy). The price gap between the three versions of the all-in-one machine is also large. For example, the quantitative version starts at a minimum of 98,000 yuan, and the native full-blood version has a maximum of 2 million yuan.
The distilled version has lower hardware requirements and naturally has lower hardware costs.
Therefore, the higher the configuration of the equipment, the better, but meeting its own needs is the key.
It is worth noting that even if the all-in-one market continues to rise, it still faces many challenges, such as software and hardware adaptation progress, large-scale model application engineering measures, specific scenario adaptability, and post-service support. What can be clear is that the popularity of DeepSeek has increased the popularity of the entire industry of large models. The hardware demand for deploying DeepSeek smart computing machines in various industries has increased significantly, which has undoubtedly created more opportunities for domestic inference chips.
