Author | Xiao En, Editor | De Xin
What level of configuration is a 10,000-GPU cluster in the L4 autonomous driving industry?
Currently, only three types of players in China can achieve a scale of 10,000 GPUs: internet giants, independent large model companies, and cloud providers and operators focused on computing power leasing. Their businesses vary, but they share a common prerequisite—the need to train foundation models with tens of billions or even hundreds of billions of parameters from scratch.
In other words, a 10,000-GPU cluster is an entry ticket for "building large models," not a pass for "using large models." Zelos, a leader in L4 unmanned logistics, has secured this entry ticket.
On September 17, the world's largest unmanned delivery vehicle operator held a technical exchange conference in Suzhou, announcing that its GPU computing power cluster has surpassed the 10,000-GPU scale—reaching nearly 15,000 GPUs. According to data disclosed by Zelos at the event, this is the first computing cluster of this magnitude in the L4 autonomous driving industry.
Zelos currently operates over 30,000 unmanned delivery vehicles, covering more than 300 cities across China and 20 countries, with a cumulative real-world L4 operational mileage reaching 270 million kilometers. For reference, Waymo accumulated a total mileage of approximately 300 million kilometers from 2009 to 2025. In terms of mileage magnitude, Zelos has approached Waymo's accumulation over the past 16 years in just 5 years.
Why would a company engaged in urban logistics and distribution build its own large models? And how do these large models form a closed loop with its commercial operations? At the conference, Zelos's CTO Zhuang Li and the technical team deconstructed the logic behind this system.
Behind the 10,000 GPUs: From Scenario Bottlenecks to Computing Power Expansion
Zelos's computing power cluster is distributed across data centers in multiple cities nationwide. Zhuang Li disclosed at the conference that the cluster comprises over 14,000 GPU cards, with 67% dedicated to model training, and the remaining computing power covering world reconstruction, data annotation, and online inference.
In terms of utilization, this computing power is hardly idle. The cluster's utilization rate has been maintained at around 90% year-round—model training involves scheduling intervals and does not run at full capacity continuously, so 90% already indicates a heavily packed queue. Zhuang Li mentioned that the algorithm team is "queuing up every day" for computing power.
There is a saying in the industry to describe the threshold of 10,000 GPUs: "1,000 GPUs can just do AI, while 10,000 GPUs are the foundation for building general large models and supporting large-scale full-domain simulations." This summary is basically accurate; a 1,000-GPU computing power is sufficient to train medium-scale specialized models for inference and fine-tuning. However, to train foundation models with tens of billions or even hundreds of billions of parameters from scratch, 10,000 GPUs are the starting line.
Then, why would an L4 company engaged in urban logistics and distribution cross this threshold?
This starts with the scenarios Zelos faces. Zelos focuses on express delivery, fresh cold chain, supermarket retail, and rural postal routes. Unlike Robotaxi or highway scenarios, last-mile logistics faces a highly disordered physical world. Dense illegal parking, temporary hand gestures, construction detours, ground debris, and GPS failures—these edge cases that can be handed over to human drivers as a fallback in L2 advanced driver-assistance systems have no retreat in L4 unmanned delivery; the system must handle them independently.
Zhuang Li made a classification based on speed at the press conference.
General roads above 30 km/h represent relatively regular scenarios. Cruising, car following, obstacle avoidance, and passing through intersections—traditional modular autonomous driving systems can cover most situations. The core idea is to write rules for each scenario: what to do when encountering an obstacle, how to respond when recognizing traffic lights, and how to handle lane markings. This approach is feasible in environments with simple road conditions and unified rules.
However, China's traffic environment has led rule-based algorithms into a dead end.
Zhuang Li said, "Every urban district and every city interprets traffic rules differently. In some places, you can turn right on a red light, while in others, you cannot. If everything is written in rules, it will ultimately become an unmaintainable massive project."
End-to-end models are the next stage of exploration. A unified end-to-end approach can string together all image information to learn "how humans drive." However, Zhuang Li pointed out its shortcoming: it learns the results, not the reasons. For example, when meeting an oncoming vehicle on a narrow road, if the other driver flashes their lights twice, is it signaling to yield, or warning to stop? If the training data has not seen this, the model will struggle to understand this meaning.
The VLA (Vision-Language-Action) model is one of the latest technological directions, aligning human knowledge and driving behavior by introducing language understanding. The system not only sees "there is a red light ahead" but also understands "I need to stop because of the red light." Tesla and several leading domestic automakers have already deployed such models on the vehicle side.
However, vehicle-side deployment has a hard constraint: the upper limit of computing power. When entering complex road conditions between 5 and 30 km/h, the contradiction of insufficient vehicle-side computing power is further amplified.
Zhuang Li gave an example encountered by Zelos in actual operations: during construction, workers pulled the central isolation barrier to the right from the yellow line position. The system needs to judge—after the barrier is pulled over, has a new non-motorized vehicle lane been separated on the right, or is it just a temporary occupation of the lane? Does the adjacent yellow line still count as a valid lane marking?
A similar situation is when a "Construction in Progress" sign is placed on the road, but visually the road appears to be already repaired—should the system trust the sign or its own perception?
"This requires combining deeper human knowledge and temporal information to make judgments," Zhuang Li said. "It is difficult with 1 to 2 billion parameters on the vehicle side; at this time, enhancement in the cloud is needed."
The last-mile scenarios between 0 and 5 km/h represent the most challenging interval. Unmanned delivery vehicles need to find customer-designated parking spots, but these spots might already be occupied, located deep in a parking lot, or have the ground piled with express packages and various irregularly shaped items. Customer preferences also change over time—parking at location A was suitable yesterday, but today location A is occupied, so it needs to switch to location B.
These are not only issues of perception and planning but also involve the continuous accumulation of operational experience and the learning of customer habits.
The challenges across the three speed intervals compound to point to the same direction: scenarios require larger models to understand the complex world, and the vehicle-side computing power cannot accommodate such models. Therefore, the only option is to train a sufficiently large foundation model in the cloud and then distill its capabilities for vehicle-side deployment. And for training larger foundation models, a 10,000-GPU computing power is the starting condition.
Apart from the scenarios themselves, Zelos has also observed another change during model training.
Zelos's technical lead shared that when the model parameter scale remains unchanged and only data is increased, the performance improvement curve of the model will rapidly change from a linear curve to a logarithmic curve: the more data added, the slower the effect grows. If the model size continues to expand, the performance improvement curve will approach linearity again.
Zelos views this as another validation of the Scaling Law in the autonomous driving field.
Its data flywheel keeps spinning, the fleet continues to expand, and millions of kilometers of new data pour in every day. However, the model's capacity has reached a bottleneck: continuing to feed data yields diminishing marginal returns. Making the model larger makes the curve steep again. According to data disclosed by Zelos at the event, the most recent increase in model scale brought a nearly 5-fold reduction in takeover rates and safety accident rates.
Behind Zelos's ability to support such a level of investment are two additional conditions.
One is commercial scale. With 30,000 vehicles operating in over 300 cities, millions of kilometers of new data are generated every day. The speed of data accumulation itself demands stronger computing power to process it.
The other is capital reserves. Zelos has accumulated over USD1 billion in financing, with investors including Ant Group, Meituan, Baidu Ventures, Blue Lake Capital, and others. Its valuation exceeds CNY 10 billion, making it the world's first decabillion-dollar unicorn in the unmanned truck field. In January 2026, Cainiao injected its unmanned vehicle assets and funds entirely into Zelos.
Zelos does not intend to become a large model company. However, the complexity of the scenarios requires larger models, larger models need more data, and massive data requires stronger computing power to process.
Thus, the construction of a 10,000-GPU cluster became a natural progression.
APEX and Multi-Agent Architecture
The 10,000-GPU cluster is the foundation, and running on top of it is Zelos's foundation world model—APEX. The current parameter scale is between 10B and 100B, with the goal of advancing to a scale of over 100B.
Unlike the "bottom-up" approach to model training adopted by most autonomous driving companies, Zelos has taken a "top-down" path. The more common practice in the industry is to collect data for specific operational scenarios, train a specialized model, and directly align the model's capabilities with business conditions. Zelos's approach is the opposite—first train a general large model that understands the entire world, and then distill small models for various specialized scenarios from it.
Zelos used a metaphor at the event: "We are not just letting the system learn the working conditions of logistics vehicles; instead, we first let it learn the knowledge of the entire world. After a person finishes college, they always need to go out and earn money, and they always need to choose a profession—it's the same logic."
The training data for APEX is divided into two parts.
One part is Zelos's proprietary 270 million kilometers of real-world L4 operational data. This is the core asset that distinguishes APEX from general large models, containing a vast amount of complex working conditions encountered only in last-mile delivery.
The other part is massive internet multimodal knowledge, including text, images, and videos, used to build the model's ability to understand human social rules.
The former gives APEX the spatial modeling capability of the physical world, knowing what is where and where it might appear next. The latter enables APEX to understand the symbolic systems and social rules invented by humans.
The technical lead gave an example to illustrate the difference between the two: "If a person waves their hand in the distance, the spatial model can predict that they are waving to let you pass. But can you know whether their meaning is to let you pass by the side, or to stop and wait?"
Traffic lights, lane markings, and construction signs are all symbols agreed upon by humans; their meanings cannot be deduced solely from the kinematics of the physical world. What APEX needs to do is to superimpose the understanding of human behavioral intentions and social rules on top of its spatial analytical capabilities.
Based on APEX, Zelos has built a multi-agent architecture. According to the needs of different scenarios, multiple specialized models are distilled from the foundation model, each taking on a different "role."
The vehicle-side VLA, with approximately 2B parameters, is responsible for real-time driving decisions on regular roads at speeds above 30 km/h. Given the limited computing power on the vehicle side, the core of this model lies not in how large the parameters are, but in how high the utilization efficiency of the parameters is.
Zelos adopted a MoE (Mixture of Experts) architecture for sparse processing. The total parameter count of the model can be large, but in each specific scenario, only a very small portion is activated. Zelos's technical lead said, "When a person passes through an intersection, their eyes only look at the traffic lights and the direction of oncoming traffic, and they might make a decision after looking at just one or two frames. When passing a narrow stone pier, they don't look at anything beyond 10 meters, focusing only on the distance on the left and right."
This selective activation allows Zelos to deploy larger models under limited computing power. He estimated that when the vehicle-side computing power increases to around 2000 TOPS and the vehicle-side model advances to 10B, it can cover the vast majority of working conditions.
The cloud-side VLA, with approximately 8B parameters, is deployed in the cloud. When the vehicle-side model encounters uncertainty in complex scenarios such as construction detours and blurred lane markings, the cloud-side VLA provides enhanced judgments for scenario understanding, making up for the shortfall in parameter volume of the vehicle-side model.
The Safety Officer Agent, with approximately 5B parameters. This is the innovation highlighted by Zelos at the press conference, which Zhuang Li believes "no other company in the industry has done."
The Safety Officer Agent handles complex last-mile scenarios at speeds between 0 and 5 km/h, including parking spot selection, escape plan generation, and customer preference adaptation. These capabilities are built upon the last-mile operational data accumulated by Zelos over five years, representing experience "honed" order by order in real-world operations.
Before the Safety Officer Agent went online, all abnormal handling in the last mile relied on manual safety officers taking over remotely. The industry assessment standard is to complete the response within 1 minute: from receiving the signal and understanding the on-site situation to starting to control the vehicle, a large amount of time within that minute is spent on "figuring out what happened."
The Safety Officer Agent compresses this process to under 10 seconds, improving efficiency by more than 6 times.
However, more important than the response speed is that the way safety officers work has also changed.
Previously, manual safety officers needed to remotely control the vehicle to complete the entire escape process, taking over from start to finish. Now, the Agent autonomously generates a complete plan—where to stop, how to bypass, and which route to take. The safety officer only needs to review and confirm whether the plan is feasible.
Zhuang Li described it as: "After the plan is generated, the safety officer performs a verification. Once confirmed, all actions are completed relying on the vehicle-side capabilities and the Agent's guidance."
Shifting from "remote control" to "review and release," the role of the safety officer has also changed accordingly.
These three models are not a cascading relationship invoked step by step, but are distilled in parallel from APEX and operate independently. The technical lead emphasized: "The collaboration of multiple agents is a division of labor, ensuring the absolute safety and efficiency of the system. It is not the case that the entire system collapses if one link is missing."
Zelos calls this architecture "Swarm Intelligence," a system comprising a central model plus multiple specialized agents. In Zelos's view, the most important feature of this architecture is scalability. "We call each different downstream task an independent agent, which means we can have other agents performing different tasks."
Following this logic, Zhuang Li mentioned a further possibility in the latter half of the press conference: "When we reach this stage, we will not only be doing autonomous driving; we can do everything that unstructured AI does."
Beyond the model architecture, Zelos has a counterintuitive judgment regarding the data value of L4 and L2.
The installed base of L2 advanced driver-assistance systems far exceeds that of L4, and the total amount of data is naturally much larger. However, Zelos believes that true value cannot be judged solely by mileage. No matter how much data there is of "a car just driving forward," its significance is limited.
L4 naturally lacks a human driver as the last line of defense; all edge cases must be handled by the system itself. The density of edge working conditions generated by L4 unmanned delivery vehicles maneuvering at express delivery points, interacting at extremely close range, and on unstructured roads is something L2 can never collect while driving on main roads. "The complex data under L4 working conditions is incomparable to the L2 industry."
The data advantage Zelos emphasizes lies not in the total volume, but in the density—more edge scenarios per kilometer. This is also the crucial data foundation for APEX to continue advancing from tens of billions to hundreds of billions of parameters.
The Data Flywheel of 30,000 Vehicles
The continuous investment in 10,000-GPU computing power requires commercial revenue to support it. During the Q&A session of the press conference, a reporter directly asked Zhuang Li: Between computing power investment and commercial revenue, is it a positive cycle or continuous bleeding?
Zhuang Li's answer was: "All the vehicles in our company are equivalent to an AI driver each month. After being dispatched, we can receive the salary of the AI driver. Once this salary comes back, it can supplement the current computing power."
Behind the statement "AI drivers earning salaries" lies the three evolutions of Zelos's business model.
The earliest was complete vehicle sales. Around 2023, the selling price of an unmanned delivery vehicle was over CNY 200,000. With product iteration and economies of scale, the average price of Zelos's current product matrix has dropped to about CNY 100,000. Prices are going down, but for many small and medium-sized logistics enterprises, the initial one-time investment still poses pressure.
The second step was financial leasing. Zelos priced hardware and software separately, allowing customers to use an unmanned delivery vehicle by paying a monthly rental fee of CNY 3,000 to 4,000. This monthly rent is about half of a courier's monthly salary.
The third step is transportation capacity services and ecosystem empowerment. Zelos establishes its own fleets and recruits franchise fleets in various cities. Customers do not need to buy or rent vehicles; they can use unmanned delivery capacity on demand. Meanwhile, Zelos has launched the Zelos Inside model, embedding L4 solutions into the vehicle platforms of partners such as Dongfeng and Yutong.
Due to APEX's distillation capabilities, the cost of adapting to different chassis is not high. Zhuang Li's judgment is: "Accelerators, brakes, and steering—the Chinese automotive industry is highly standardized. Making fine-tuning adjustments based on the characteristics of each chassis means the adaptation cost is not as high as imagined."
The three evolutions are doing the same thing: lowering the threshold to get more vehicles on the road. More vehicles mean more real-world operational data; data is fed into the 10,000-GPU cluster to train better models; better models make the products safer, covering a larger market.
This is the data flywheel Zelos has been talking about: the output of every link is the input for the next link.
The integration with Cainiao in January 2026 made this flywheel spin faster. Cainiao injected its unmanned vehicle assets into Zelos through a capital increase, and no longer produces and sells unmanned vehicles itself. The two parties formed a dual-brand operation: Zelos focuses on the mass market and lower-tier cities, while the Cainiao brand focuses on core logistics scenarios. After the integration, Zelos's fleet size rapidly jumped from about 16,000 vehicles to over 20,000, and its market share in China's unmanned logistics vehicle market also increased to over 70%.
Zelos has a customer in Northeast China. Previously, because remote villages were far from the county town, e-commerce parcels flowing from the Yangtze River Delta to the Northeast took three to five days. The volume distributed to the county was very small, and couriers were unwilling to make a special trip just for a few items.
After a network point deployed Zelos's unmanned delivery vehicles, it achieved automatic dispatching with an unlimited number of parcels per day. Within a week, the parcel volume at this network point increased more than 20 times. An even more interesting change occurred a month later, with the online shopping volume of local residents increasing 200 times.
Unmanned delivery vehicles bring not only an improvement in logistics efficiency but also activate consumer demand that was previously suppressed by distribution bottlenecks.
China Post, on the other hand, is one of Zelos's largest single customers.
Zelos won all 8 bid sections for China Post's nationwide unmanned vehicle leasing procurement, delivering 10,000 unmanned vehicles in 2026 for postal route automation, covering the last mile from cities to towns and townships. "Express delivery to the countryside" is a crucial component of the logistics network among the six networks in the national 15th Five-Year Plan, and Zelos has been explicitly written into Jiangsu Province's logistics network construction plan.
When answering about the return on investment for the 10,000-GPU computing power, Zhuang Li made a comparison using market size: "Currently, we support 30,000 vehicles, but we are facing a replacement market of 30 million traditional transportation vehicles. The proportion of 10,000 GPUs in the middle is relatively small. We even want to step towards 100,000 GPUs."
Regarding how long it will take to go from 30,000 to 300,000 vehicles, Zhuang Li's judgment is about three years. He believes the biggest variable currently lies not in technology, but in policy. There is a lack of unified operational qualification standards for unmanned delivery vehicles nationwide, and regulations vary across regions.
"Multiple ministries and commissions, including the Ministry of Transport, the Ministry of Public Security, and the Ministry of Industry and Information Technology, are already paying attention, and new regulations will be introduced soon. Once the rights, responsibilities, and interests are clearly defined, the subsequent process will be very smooth."
Conclusion
The competition in unmanned delivery is entering the next stage.
Over the past five years, the competition focused on whose vehicles can run, whose costs are lower, and whose city coverage is broader. Zelos won the first round with 30,000 vehicles and a market share of over 70%.
The stakes for the next round are changing.
As Zelos begins to use a 10,000-GPU cluster to train its own world model and distill the vehicle-side VLA, cloud-side VLA, and Safety Officer Agent from it, the industry's entry threshold is no longer just vehicle manufacturing and operational capabilities, but a comprehensive barrier of data scale, model parameters, and computing power infrastructure.
From 30,000 to 300,000 vehicles, Zhuang Li said it will take about three years. But what he truly spent time explaining at the press conference was not the number of vehicles, but how the system's understanding of the physical world will change after APEX's parameters advance from tens of billions to hundreds of billions.
For Zelos, the 10,000-GPU cluster is not the endpoint, but a starting point just crossed.