AI + Supply Chain: How Large Models and Intelligent Decision-Making Reconstruct the Entire Supply Chain | 2026 AI Partner · Beijing Yizhuang AI + Industry Conference
How exactly does AI create value for industries? Witness the crucial leap of AI transforming from a "hot topic" to "productivity."
Today, SF Express uses AI to dispatch one-third of China's air cargo. From the collaboration of millions of people and the dispatch of hundreds of thousands of vehicles to the dynamic routing of 50 million parcels per day—the complexity of the logistics supply chain is the perfect testing ground for AI implementation. How does an ultra-large-scale network use large models, Agents, and operational research algorithms to turn the "24-hour delivery" promise into an everyday reality and transfer this capability to more Chinese enterprises?
The following is the speech content of Tang Kai, the senior vice president of the Digital Intelligence Supply Chain Solutions Business Group of SF Express Group, sorted and edited by 36Kr:
Tang Kai | Senior Vice President of the Digital Intelligence Supply Chain Solutions Business Group of SF Express Group
Good afternoon, distinguished guests. I'm very glad to be here in Yizhuang to share with you the application of AI in the logistics supply chain.
Before we start, let me talk about the supply chain. From an intelligent or digital perspective, we need to abstract or stratify the supply chain. It's impossible to deal with the supply chain as a whole. We'll look at the entire supply chain collaboration, as well as the digital and intelligent progress of the supply chain, from three levels: elements, links, and networks. The basic elements of the supply chain are people, vehicles, goods, and places. All supply chain actions and executions are related to the connection of these elements. The mobilization of elements forms links, and the intersection of links forms networks. Just looking at SF Express, let me mention one point about its ultra-large scale in the supply chain field. Regarding people, vehicles, goods, and places, in the SF Express system, the scale of collaborative employees exceeds one million. The number of self-owned vehicles is over 200,000. Even more significant is China's air freight network. There are 240 all-cargo planes in China, and 110 of them are owned by SF Express. The denser auxiliary cargo capacity network, which also transports goods in the auxiliary cargo holds of passenger planes, has one-third of its capacity in the hands of SF Express. That means for every three planes flying over China, one is carrying SF Express cargo.
Given the high density of elements, we also need to consider the complexity of links. For express delivery, the links are collection, transfer, sorting, and delivery. For the supply chain, they are procurement supply chain, production supply chain, channel supply chain, consumption supply chain, and after-sales supply chain. The four links of collection, transfer, sorting, and delivery may sound simple, but there are 120 sub-links in the entire operation process, forming the entire link of collection, transfer, sorting, and delivery. The execution complexity of this link is very high. All these links intersect to form a large network. In this large network, what SF Express needs to do is to conduct intelligent dispatch and intelligent collaboration based on the digital network for all resources and links on it. Why can SF Express promise customers 24-hour or 48-hour delivery today? Because all link mobilization processes are based on autonomous capabilities, which we call intelligent and autonomous dynamic dispatch.
Regarding dynamic dispatch, when a parcel is sent from Beijing to Shenzhen today, the static route determines which plane it should take, which courier should pick it up and deliver it, and which transfer station it should pass through. However, planes may be delayed, couriers may get sick, and cars may break down on the road. All nodes may have problems. SF Express processes 50 million parcels a day, and 50% of them need to activate dynamic routing. For example, if a vehicle stops for an unusually long time at a place where it shouldn't stop on the route, the system will automatically identify whether it is in an abnormal state. After confirmation, such as if it breaks down on the highway, the entire system will automatically dispatch nearby available vehicles to save the delivery time. The entire link must ensure no problems within 24 hours. Based on the resource collaboration and resource dispatch of the large network, this must be accomplished by intelligent and digital means, not by manual work. Large models became popular after ChatGPT emerged at the end of 2022, but in 2016, we began to deploy on a large scale our intelligent capabilities based on operational research algorithms, traditional machine learning, and reinforcement learning. We were the only Asian company to enter the Franz Edelman Awards final in 2025. This award is the highest in the field of operational research, representing our capabilities in the entire field of operational research and traditional machine learning. In the era of large models, we have started to shift to large model capabilities on a large scale.
In such a complex network, what SF Express does is based on full-stack self-research of full-link, end-to-end, full-volume digitalization. From the underlying capabilities, we even have our own mapping department. Although we can use Gaode Navigation and Baidu Navigation, SF Express needs industrial-grade navigation, so we have our own mapping department. The precision of our navigation (far higher than ordinary navigation) is remarkable. With Baidu Navigation, you can only be guided to a residential area, but with Fengtu (SF Express Map), you can know how to get around inside the residential area, which companies and households are on each floor. All this information is automatically updated and collected by our couriers during their daily deliveries. This is what we call full-stack self-research capabilities.
Since 2023, we have been building large model networks and systems on a large scale. Currently, we have approximately two large vertical domain models. One is the Fengyu Large Model, which is a multi-modal large model mainly for internal related business use, including network dispatch and various detailed applications. The other is the Fengzhi Large Model, a vertical domain large model for the logistics industry launched to the market. Regarding the application of the entire model today, a topic that people are more aware of is Agent. From opening the underlying model, to the Agent platform, to the market, and to the self-construction capabilities, we have made all the transformations internally. Currently, there are approximately 5,000 Agents running in real-time within SF Express. Among them, more than 200 are official Agents, such as the Intelligent Network Dispatch Agent and the Operational Optimization Agent. There are 200 official Agents, and the remaining 4,800 are small Agents built by non-technical personnel from various business departments based on the Agent platform. They consume approximately 80 billion tokens per day.
Based on our understanding of the entire industry and our own digital and intelligent capabilities, we have ready-to-use templates for the complexity of elements, links, and networks. For most companies, the complexity of elements, links, and networks is lower than that of SF Express. We abstract these capabilities to form several major aspects of capabilities. The first is the vertical domain large model, which we call the Logistics Decision-Making Large Model, based on the Fengzhi product system. The second is the end-to-end logistics operating system, called Fengzhi Cloud. And there is also the LAAS Logistics Open Platform for partners. We hope to provide consulting, AI, and multi-link digital services to empower the entire industry and make the supply chain of every enterprise, industry, and sector smarter.
With Fengzhi, we combine all our traditional logistics software with Agent capabilities. If you understand the supply chain a little, you know that there is a warehouse management software in the supply chain. We hope to provide customers with software functions directly equipped with Agent capabilities. The difference can be understood in this way: if you buy a traditional car, you need to buy additional tools to use it. If you buy a self-driving car, you are directly buying a service. The need for human operation of tools is transformed into a service. A self-driving car directly provides the service of transporting people or goods without the need for capability transformation. In the vertical application field, we directly provide Agent capabilities. We help enterprises quickly build the internal capabilities of SF Express that I mentioned earlier through Fengzhi's AgentOS. There is also an Agent Studio, and the Business Semantic Center helps everyone build a knowledge graph model on all digital foundations. Previously, your enterprise data was used by the system, but through the semantic center, the large model automatically builds models, enabling all data to be directly called by Agents. Through Agent Studio, you can directly build Agent applications on your existing digital platform to support large-scale use. We have two underlying applications as support. The first is Fengzhi Code, which is used for large-scale customization of software, significantly reducing the marginal cost of customization. The other is Fengzhi Evolver, which enables the large model to generate small models. For supply chain diagnosis and consulting today, we don't start with methodology but with models. We hope to quickly conduct consulting verification and supply chain diagnosis through data and POC. This part involves a large amount of model generation. How to break through the production capacity bottleneck? Through Fengzhi Evolver, the large model automatically generates small models, generating a large number of customized small models and significantly reducing the model optimization time.
This part is the large model - the Logistics Decision-Making Large Model, which performs three tasks: supply chain prediction, supply chain optimization, and supply chain analysis.
For supply chain prediction, we use the modal generalization of the large model to solve several special scenarios that are difficult to handle in prediction. For example, in new product prediction and chip prediction, due to the lack of historical data, it was difficult for previous models to calculate accurately. Through the large model for automated consulting and modeling, previously we used single SQ modeling, and now we use cluster modeling of the QA channel type, enabling better data correlation and data supplementation. In the data verification for the world's largest coffee chain enterprise, the accuracy was improved by approximately 5 percentage points, which was impossible for previous models.
Regarding supply chain optimization, we rewrote the classic operational research models using the large model. For example, in path optimization and special optimization scenarios, after rewriting these models with the transformer architecture, the computing time is increased by about 10,000 times. Previously, it took minutes to perform a multi-point path optimization, but now it can be reduced to milliseconds, and the calculation can be completed in 0.05 seconds. The resource consumption is one-fiftieth of the previous level. Previously, 50 CPUs were needed for the calculation, but now one GPU can complete it.
As for the analysis ability, this part combines the large model with small models and uses the supply chain intelligent agent approach. When intelligent agents were very popular in the entire industry in March 2025, we realized in the 2B application field early on that to solve the problem of the model black box and the problem of model efficiency and accuracy, we must use the intelligent agent approach. In August 2024, we launched the comprehensive application ability of the supply chain intelligent agent in the industry.
Before intelligentization, the first step for all enterprises is digitalization. No matter how powerful the model is, without accurate, real-time, and reliable data, good results cannot be achieved. Before intelligentization, for end-to-end digitalization, we use the Fengzhi Cloud Chain model system to achieve full-link digitalization for customers based on elements and links.
Finally, let me share a case. The world's largest coffee chain enterprise has more than 7,000 stores in China. We started helping them build a smart store system in 2019. We helped them with store-level sales prediction and replenishment prediction based on the store system. We have been doing this for six years, and now we are using the large model to iterate on the traditional model.
We are also responsible for the full-channel sales prediction in China for an internationally renowned beauty giant. Why do they entrust us with the full-channel sales prediction in China? Because the channel complexity in China far exceeds that of any single country. Their chain enterprises' practical experience in China is more advanced than that in the United States. When I talked to the CIO in the United States before, I learned that the capital structure of the business in China has changed, and the Chinese entity has acquired its business in China. This aspect has temporarily stopped progressing.
The depth, breadth, and complexity of the Chinese supply chain are unparalleled in the world. When serving MNCs, our supply chain practices in China are the world's best supply chain practices, which have been verified by many multinational enterprise customers. Previously, the best practice in supply chain digitalization was "copy to China." Now, what we are doing is "copy from China." In the luxury goods industry, the world's first smart store is the LV store in Changsha, which we helped build. In the industrial manufacturing field, the supply chain center platform of Eaton was developed in China and then promoted to the Asia-Pacific region. We conducted a POC for Nestle's high-level production scheduling. It didn't work well in China, and we needed local Chinese capabilities for localization experiments. When we combine the digital supply chain with the entire technological change and the cross-border application of AI capabilities, we can see that China has a great opportunity window to create the world's best practices in the digital supply chain.
We have also put forward all our capabilities from the bottom to the top, including consulting diagnosis and top-level planning, then to the application ability of large model AI, and to the end-to-end digital ability. Combined with the refined ability of different solutions for each industry and the full-link services for the supply chain from procurement supply chain, channel supply chain, sales supply chain, to after-sales supply chain, today we hope to work with more Chinese enterprises to create more world's best supply chain practices.
Thank you all.