Research on automotive cloud service platform: with architecture upgrade and computing power improvement, cloud services enter a new stage

In 2026, the Internet of Vehicles industry generates petabytes of data in a single day, and the vehicle backend system communicates automatically with the cloud server ten to hundreds of times a day. As the iteration cycle of VLA models and cockpit agents is further shortened, higher requirements are placed on the stability, low latency, and storage efficiency of cloud computing power, promoting the transformation of cloud infrastructure from "scale-driven" to "value-driven".  

For cloud providers, the focus of competition has shifted from "complementing hardware" to "improving service quality". Algorithm optimization, cloud-native AI, collaborative scheduling, and security compliance have become competitive edges;

For OEMs, through a multi-cloud strategy and rational use of the ecosystem and technical advantages of different cloud providers, they can achieve "cost reduction and efficiency improvement", ensure the stability of real-time cloud services, and accelerate the implementation of core businesses such as autonomous driving, intelligent cockpits, and mobility services, building differentiated competitive edges.  

The focus of cloud providers’ infrastructure shifts to “improving quality and efficiency”.

In 2024, automotive cloud providers found themselves trapped in a dilemma of "chip shortages and insufficient computing power." Cloud providers ramped up their hardware investments to stack servers and GPUs to meet the surging demand for computing power driven by the integration of AI large models and NOA (Navigate on Autopilot) into vehicles. Some providers also began to develop chips in-house.    
In 2026, as the tight production capacity of general-purpose chips gradually eases and algorithms continue to optimize utilization efficiency of cloud computing power (virtualization, segmentation, and pooling technologies become more mature), automotive cloud infrastructure will no longer blindly pursue the expansion of hardware, but will center on improving utilization efficiency, stability, and adaptability of computing power as the focus of developing next-generation automotive cloud service solutions.  

Taking cloud providers such as Google Cloud and Alibaba Cloud as examples, their cloud infrastructure solutions in 2026 focus on improving the efficiency of existing cloud infrastructure with new algorithms and applying new server architectures to optimize the stability of cloud clusters.

1.Google's new algorithm improves cloud computing cluster efficiency

Google introduced the algorithm TurboQuant in early 2026. With quantitative compression and intelligent caching technology, it effectively lowers storage requirements and speeds up inference. It can adapt to the lightweight computing power requirements of automotive scenarios and solve the problem of "insufficient storage hardware restricting the utilization of computing power". It offers the following benefits:
For KV Cache quantization, 3.5 bits per channel achieves near-lossless precision with equivalent accuracy, reducing the storage required by more than 5x compared to the native 16-bit format.
Reduced memory access enables faster inference, with zero additional overhead in the inference pipeline.
The quantization speed is 100,000 to 1 million times faster than PQ/RabitQ.

According to the results released by Google, the TurboQuant curve achieves nearly lossless performance in long context compression (score reaches 0.997).

2.Chinese cloud providers such as Alibaba Cloud apply super-node architectures to improve the operating efficiency of computing clusters.

Among Chinese cloud providers, Alibaba Cloud, Baidu Cloud, and Huawei Cloud launched super-node server architectures that optimize cluster stability in 2025, optimizing inference efficiency and cluster stability, and improving the cost-effectiveness of the entire solutions:  

Alibaba Cloud

Alibaba Cloud released Panjiu AI Infra 2.0 AL128 super node servers at the 2025 APSARA Conference. Through ScaleUp interconnection within the super node, they shorten the completion time of E2E inference tasks and improve foundation model inference experience for users. One of the features of such servers lies in ScaleUp interconnection, a technology that caters to modern GPU design, including:
Native memory semantics: Direct access to the computing core of the GPU is allowed, and it is easy to mount to the SoC bus via the interface. There is no conversion overhead and intrusive design for the computing core.  
Ultimate performance: Extremely high bandwidth (the entire chip can reach TB/s) and extremely low latency can be achieved. In addition to the high message efficiency of the protocol, excellent performance under high load is also required.
Minimalist implementation: Chip area and cost are minimized, allowing valuable resources and power consumption to be reserved for the computing power and on-chip memory of GPU.
Highly reliable link: In a very high-density SerDes environment, high availability is ensured through a high-performance physical layer and link-level retransmission and fault isolation mechanisms.

Huawei

Huawei has released the next-generation AI data center architecture - CloudMatrix and the mass production product - CloudMatrix384, which breaks through the traditional CPU-centric hierarchical design and supports direct high-performance communication between all heterogeneous system components (including NPU, CPU, DRAM, SSD, NIC and domain-specific accelerators), realizing the transformation of the resource supply model from the server level to the matrix level.

In August 2025, Changan Tops AD adopted Huawei Cloud’s CloudMatrix384 super node solution". Based on the CloudMatrix384 super node and Huawei Cloud's high-bandwidth and large-capacity storage cluster, Changan Automobile has achieved efficient training of its autonomous driving model, and adaptation to various autonomous driving models such as VLA and end-to-end models.

Baidu

Relaying on Kunlunxin, a super node server architecture was released. This solution achieves super single-node performance. Its 32-GPU/64-GPU configuration uses faster in-machine communication to increase inter-GPU interconnection bandwidth by 8 times, single-machine training performance by 10 times, and single-GPU inference performance by 13 times, which can support large-scale VLA training and promotion.

Device-cloud collaboration technology optimizes cockpit and vehicle-road-cloud scenario experience.

From 2025 to 2026, device-cloud collaboration technology serves as one of the technical bases to accelerate the penetration into cockpit and vehicle-road-cloud scenarios. With the complementary model of "cloud computing power empowerment + automotive real-time response", it will solve problems such as unsmooth cockpit interaction and vehicle-road-cloud system effects that are not as good as expected, and optimize user experience.
 
1.Cockpit scenario

In 2026, the cockpit device-cloud collaborative architecture upgrades capabilities through the combined approach of "cloud foundation model optimization + vehicle lightweight model execution". The cloud undertakes high-load computing and inference tasks, including complex semantic understanding, multi-turn dialogue tracking, massive knowledge base data invocation, and other tasks requiring high computing power. The vehicle is in charge of real-time response, low-latency interaction, and privacy protection. With technologies such as edge node sinking, the end-to-end latency is controlled within 500 milliseconds to meet user needs. Cloud IVI is a typical application of device-cloud collaboration in cockpit scenarios. 

For example, the Aion Cloud IVI released by GAC and Huawei in September 2025 uses vehicle-cloud intelligent collaboration to reconstruct the cockpit computing power allocation logic: all computing and rendering tasks are handed over to the cloud, and the local IVI is only responsible for interaction and display. The IVI local computing only consumes 0.02-0.03TFLOPS, which greatly reduces the consumption of automotive computing power. This not only ensures a smooth experience of the new IVI system, but also solves the problem of the old vehicle upgrade: there is no need to replace hardware, and smooth intelligent interaction can be achieved even with mid- to low-end chips.

In addition to saving computing resources, this cloud IVI also takes advantage of cloud resources to:
Complete cloud ecosystem aggregation, open up 20,000+ cloud applications, and support the flow of mobile applications to IVI.
Speed up the OTA frequency; all application and system updates are completed in the cloud, and the latest version can be updated in half a day, allowing cockpit functions to always remain "cutting-edge".

2.Vehicle-road-cloud scenario

In the vehicle-road-cloud scenario, the core value of device-cloud collaboration lies in opening up the data links between vehicles, roadside equipment and cloud platforms, and building a complete collaborative closed loop of "vehicle perception, roadside blind spot coverage, and cloud scheduling".  
The cloud is responsible for core tasks such as data fusion, macro traffic flow prediction, and global scheduling optimization. Through multi-dimensional data fusion, intelligent allocation of mobility resources is realized. The cloud control platform adopts a two-level architecture of "edge cloud + zonal cloud" to achieve hierarchical processing and global optimization.
Edge computing nodes serve as vehicle-road connection hubs, ensuring end-to-end latency of ≤10 milliseconds and focusing on real-time data processing and local scheduling.

In August 2025, Dongfeng eπ007 realized the technology of optimizing the smart parking function with vehicle-road-cloud collaboration technology. The technical path is "cloud scheduling + parking lot allocation + vehicle execution". This technology can increase the parking space utilization rate by 45% and increase the number of vehicles parked per unit area by 1.8 times. Thanks to parking lot sensors and cloud technology, Dongfeng eπ007 does not require manual operation after running into the parking lot. The parking lot equipment can instantly recognize license plates, compressing the entry time to within 15 seconds.