GuShiio.com Article source: Titanium Media
Image source: Generated by AI
At the beginning of the new year, various companies were competing to connect to DeepSeek. Start-up teams called APIs overnight on the public cloud to develop lightweight applications such as intelligent customer service. Financial institutions were eager to buy DeepSeek all-in-one machines, and local governments quietly deployed DeepSeek on the government cloud… Companies from different tracks are rushing into this arms race for smart upgrades with very different attitudes.
For CTO, DeepSeek’s amazing model effect will undoubtedly excite technology practitioners. But more CTOs are also beginning to be anxious: After enterprises use large models on a large scale, how can they accelerate the intelligent upgrade of their own enterprises through large models? How to plan? How to invest? In what direction? What business scenarios should we focus on? How to update the technology stack…
How can companies make good use of large models, including DeepSeek, in their business processes?
In Jingdong Cloud’s view, the enterprise application model will go through four stages, and the answer is also among them:
- At the beginning, companies chose to let employees use the large model first regardless of whether they could find application value. Currently, most domestic companies are at this stage and regard the deployment of DeepSeek as their top priority.
- As the use of universal models deepens, companies will begin to try to reduce weight, and the first choice is agents. Some companies have begun to explore this stage.
- After the application of agents is deepened, enterprises will find that the universal model has bottlenecks in solving the professional problems of the enterprise. Large and medium-sized enterprises will use their own data to distil the vertical model. This will enter the third stage, replacing the universal model with the vertical model.
- Finally, the vertical model strengthens the capabilities of agents, and some agents will grow into super agents. Multi-agents will collaborate to create new business formats and enter the ultimate scene intelligence.
Lightweight agents break the game:
From “usable” to “easy to use”
Agents are the first key to unlocking the value of large models. The technology industry has reached a consensus that agents have become the core driving force for the digital and intelligent transformation of enterprises. As a mainstream application form, agents can not only improve operational efficiency, but also lower the threshold for AI use.
According to Deloitte’s “Technology Trends 2025” forecast, agents will become an important part of future enterprises.
These scenarios have exerted great business value within Jingdong. For example, in terms of professional assistants, JoyCoder Intelligent Coding Assistant can help developers predict code, generate comments, and conduct intelligent reviews. The adoption rate of the generated code exceeds 40%, and the development cycle is shortened by 30%. In terms of marketing, AIGC technology can quickly generate high-quality marketing copy and product scene maps, improving production efficiency by up to 90%. For example, Yanxi digital person can automatically generate live broadcast scripts, which has covered 7500 live broadcast rooms, and the total number of GMV products exceeds 10 billion.
The relevant person in charge of Jingdong Cloud said: “The five scenarios we have summarized can already cover 80% of the general needs of enterprises to use agents and improve business efficiency. At this stage, during the implementation of agents, the core scenarios should be penetrated and penetrated., far more important than blindly pursuing the number of agents.”
At present, Jingdong Cloud has accumulated these agent capabilities on the Yanxi agent platform to externally support more companies in exploring agents.
For example, upstream and downstream roles in the food industry are scattered, and meetings are crucial for decision-making management. In Weiquan Food, Jingdong Cloud has built a meeting management system that can automatically generate meeting text records and classify them, effectively improving the company’s operating efficiency; in aircraft leasing in China, the company’s professional knowledge is complex, and manual inquiries affect decision-making efficiency. Jingdong Cloud has built an intelligent Q & A system to provide efficient and accurate Q & A interaction, and continuously optimizes performance based on feedback, greatly improving employee office efficiency.
After enterprises have achieved results on agents, for many medium-sized and large-scale enterprises, considering that AI applications will become more and more extensive in the future, self-built smart computing infrastructure will become a must-answer question for decision-making.
Build a smart computing base:
The “Re-investment moment” for technology decision-makers
When the AI computing power density exceeds 40kW/cabinet and model parameters move towards trillions, enterprises building smart computing infrastructure has transcended the scope of technological upgrades and has become a strategic choice to reshape the industry’s voice. Those companies that took the lead in completing intelligent transformation are just like the pioneers of electrification innovation a century ago, carving new coordinates on the map of the digital age.
Upgrading from general computing to smart computing, infrastructure faces triple challenges: power supply/cooling of high-density smart computing, microsecond interconnection of computing power clusters, and dynamic scheduling of heterogeneous computing power. For example, with the increase in server power consumption, the power of stand-alone cabinets is increasing significantly. Currently, the average data center is 4-6kW. In order to meet the operation of GPU servers, the power requirement of a new smart computing center stand-alone cabinet has reached 20 – 40kW. According to the newly released GPU server parameters, the power requirement of a stand-alone cabinet in the smart computing center will reach 40 – 120kW. At the same time, GPU servers now require that the network networking distance be as close as possible, and there are clear transmission distance requirements. This requirement directly causes the power density of the computer room to increase at the same time.
Based on Jingdong Group’s complex scenario practice, Jingdong Cloud has built a one-stop large model product matrix, from the underlying smart computing infrastructure, to the middle-level model services and tools, to the upper layer Agent application development, supporting enterprises to quickly deploy large models and applications.
In terms of power supply/cooling for high-density smart computingThe power density of smart computing centers is five times that of traditional data centers, and it has strict requirements for power supply and heat dissipation. The Alpha Intelligent Computing Module independently developed by Jingdong Cloud supports integrated delivery of liquid cooling systems. The PUE is controlled within 1.15, and the cost of the same scale is directly reduced by 15%. The new generation of liquid cooling servers has a 50% increase in heat dissipation efficiency, and a stand-alone cabinet is 20kW annual power saving 8500 degrees.
In terms of computing power cluster interconnectionThe “golden triangle” of computing, storage, and networking is crucial to the efficient use of large models.
- Jingdong Cloud Super Intelligent Computing Integrated Computing Power Cluster, with single-cluster 100,000 card level cluster scheduling capabilities, relies on collaborative optimization of software and hardware, driving the computing power utilization (MFU) of the large model to jump to 75%.
- Yunhai AI storage, full stack support for hundreds of billions of large-scale models, 4K random write IOPS exceeds 10 million levels of speed response, memory and computing separation architecture achieves a double 30% breakthrough in performance and cost, and builds a universal storage base for ultra-large-scale AI training.
- Jingdong Cloud’s high-performance cloud network, the RDMA bandwidth has been increased to 3.2T, and the end-to-end communication delay has dropped to 2 microseconds, supporting a 100-billion-parameter model and lossless training, releasing the limit of AI computing power.
In terms of dynamic scheduling of heterogeneous computing power,, Jingdong Cloud vGPU AI computing power platform was created to meet the heterogeneous computing power needs of large models. Supports unified management of multiple computing power and refined operation and maintenance. By deeply optimizing the DeepSeek architecture, the inference performance of the full-blood version is improved by 50%. At the same time, AI full-scene dense computing and container isolation technology are used to ensure zero leakage of model participation in training data in dual domains, providing users with financial level security protection.
In industry practice, Jingdong Cloud successfully supported a leading new energy vehicle manufacturer and a global leading new energy technology enterprise to build an intelligent computing base covering the entire group and realize refined management of kilocalorial-level AI computing power clusters. Technically, on the one hand, it innovates multiple computing power scheduling to significantly improve GPU utilization; on the other hand, it creates a full life cycle AI development environment to achieve out-of-the-box use and greatly improves R & D efficiency. At present, the platform has supported more than 20 core scenarios such as enterprise intelligent driving research and development and humanoid robots, and has become the group’s digital intelligence engine. It is estimated that within one year, the training cycle of the two companies ‘large models will be shortened by 40%, and the annual savings in computing power costs are equivalent to building two new data centers.
Vertical model distillation:
A critical leap towards a comprehensive smart scene
When the General Model completes its technical enlightenment, a once-in-a-lifetime opportunity belongs to those companies with high-quality own data. By using their own data to complete model distillation, they took the lead in promoting industrial efficiency changes, moving AI applications to the real value deep water area, and opening up comprehensive intelligent scenarios.
Jingdong Cloud believes that the formation of comprehensive intelligent scenarios can be achieved through two ways: First, agents grow into super-agents and become a new business format. The second is intelligent upgrades to past digital scenarios.
Within Jingdong, an efficiency transformation aimed at the miniaturization of large models is quietly underway. By injecting complex, real, and dynamic domain data in Jingdong’s retail, finance, health and other systems into the general model, the Yanxi model has been used in multiple intelligent scenarios such as health consultation, financial marketing, code generation, and collaborative office. The accuracy in specific tasks is improved by more than 20% compared to the general model. More importantly, neuron pruning technology achieves slimming without reducing intelligence, saving model memory by 70%, and increasing reasoning speed by 1.5 times, allowing hundreds of billions and trillion-level parameter models to be implemented easily.
Behind the breakthrough in effect is Jingdong’sModel distillation, data governance and mixing, training recipe optimization, large and small model coordinationLarge model development computing technology accumulated in multiple fields. Recently, the research results of this model development computing self-developed by Jingdong Exploration Research Institute will also be published in Nature for the first time, and will open services to external companies through the Yanxi AI development computing platform.
The technical dividends of large model applications form a multi-point blooming application picture that grows outward:
- In the field of data elementsBig models are reshaping the enterprise data value chain. Faced with massive unstructured data, traditional manual processing only has an accuracy rate of 50-60%, and the enterprise data element service platform can achieve 99% accurate identification through large models;
- In the field of intelligent operation and maintenanceThe AI NOC intelligent operation and maintenance system compresses the time for locating major faults from hours to seconds. Behind it is an operation and maintenance knowledge base that has accumulated the experience of more than 10,000 R & D personnel;
- In the field of urban governanceThe Yimeitong system uses agents to automatically capture data, increasing grassroots work efficiency by 70%, and the ICOS city platform shortens the government application development cycle by 60%.
- In the field of smart marketing, more than 350,000 merchants use AIGC to generate product maps, 3 billion-word marketing grass copy is generated by AI, and 500,000 AI sales Short Video are being released in real time… The transformation efficiency of industrial data has undergone qualitative changes. The Financial Growth Cloud has also opened up 460 million Jingdong Financial users to operate Knohow, helping bank customers achieve AUM and MAU growth by more than 30% through three smart scenarios: interactive marketing, advertising, and marketing content generation.
These scenarios verify the two major value fulcrums of the vertical model-not only to go deep into business capillaries to obtain data nutrients, but also to adapt to real business scenarios through engineering transformation. Jingdong’s model distillation technology honed through its own complex business scenarios is lowering the AI entry threshold for more companies and easily getting tickets to the big model era.
When Jingdong Cloud opened its big model “toolbox” to the industry, the anxiety of CTOs had a new footnote: Technology deployment is only the starting point, and the real competition lies in how to transform AI into business evolution genes and reshape AI productivity. Those companies that accumulate scenario awareness in lightweight applications, reserve technical flexibility in the construction of smart computing bases, and build knowledge barriers in vertical distillation are quietly opening up.
Some industry insiders asserted: “What the big model eliminates is not technology, but organizational slowness.” When DeepSeek democratizes AI capabilities, CTO’s anxiety is essentially the collective pain of corporate digital transformation-anxiety may be the best sobering agent in this evolutionary race for survival.
