en.Wedoany.com Reported - July 2 news, China's Kingsoft Cloud will accelerate the construction of GPU computing clusters in the second half of the year to meet the rapidly growing AI computing needs of its top-tier clients. The related demand primarily comes from China's Xiaomi Group and Alibaba's large model teams. Specifically, Xiaomi's GPU computing demand from Kingsoft Cloud has been upgraded from a ten-thousand-card cluster to a super-large-scale computing cluster, with the corresponding investment budget increasing from an initial nearly 4 billion yuan to over 10 billion yuan.
The core of this expansion lies in the capability to deliver large-scale GPU clusters. AI large model training and inference impose high requirements on server quantity, GPU interconnection, storage throughput, network bandwidth, power supply, cooling, and cluster scheduling. Simply purchasing GPU servers does not directly translate into usable computing power. Cloud vendors must complete the overall deployment of data center resources, eight-GPU servers, switching networks, distributed storage, container scheduling, training platforms, fault monitoring, and operation and maintenance systems to transform hardware resources into AI computing services that clients can continuously utilize.
Alibaba's large model team has signed a five-year computing power lease contract with Kingsoft Cloud, involving over 3,000 eight-GPU servers. Based on the monthly rental price at the time of signing, the monthly revenue after full delivery is approximately 300 million yuan, with an annualized revenue exceeding 4 billion yuan. The significance of such long-term lease contracts for cloud vendors extends beyond increasing revenue orders; they also enhance the certainty of computing cluster construction. GPU clusters require heavy upfront investment and long construction cycles. If client demand is unstable, equipment utilization rates may be low. Long-term contracts allow for clearer planning of construction pace, server procurement, cabinet deployment, and operation and maintenance resource allocation.
The upgrade in Xiaomi's demand reflects the expanding computing power consumption across large models, smartphones, automobiles, and AIoT scenarios. Xiaomi's AI needs are not limited to single model training but may also involve on-device AI for smartphones, automotive smart cockpits, autonomous driving data processing, voice interaction, imaging algorithms, IoT device collaboration, and internal enterprise R&D platforms. The upgrade from a ten-thousand-card cluster to a super-large-scale computing cluster indicates that training, fine-tuning, inference, and data processing tasks are transitioning from project-based needs to long-term infrastructure requirements.
Kingsoft Cloud has previously undergone multiple upgrades in intelligent computing cloud and AI platform layers. Its intelligent computing platform, "Kingsoft Cloud Xingliu," has been upgraded from a resource management platform to a one-stop full-process AI training and inference platform, covering heterogeneous resource scheduling, training task management, inference services, and model APIs. For large model clients, the underlying GPU is just the foundation; what truly impacts usage efficiency is whether resources can be quickly allocated, tasks can run stably, faults can be automatically handled, and training and inference processes can be seamlessly integrated.
The construction of GPU computing clusters will also drive demand for a range of communication and data center equipment. Over 3,000 eight-GPU servers correspond to a large number of high-speed network connections, switches, optical modules, network interface cards, storage devices, cabinets, power distribution, and liquid or air cooling systems. The larger the computing scale, the more critical the network architecture becomes. Large model training requires multi-machine, multi-card collaboration. If network latency and bandwidth are insufficient, GPU utilization will be dragged down, ultimately affecting the client's actual training efficiency.
Kingsoft Cloud's expansion task in the second half of the year will focus on delivery pace. Xiaomi's budget of over 10 billion yuan corresponds to a larger-scale, long-term computing power pool, while Alibaba's five-year contract corresponds to a server cluster with a clear production schedule. For cloud vendors, the next step is to integrate GPU server arrival, data center racking, network debugging, platform access, and client acceptance. Only when the computing cluster achieves stable delivery can it be truly converted into revenue from training, inference, and cloud services.
