NUPIAO Reporter |
NUPIAO Editor | Wen Shuqi
On July 2, NUPIAO learned exclusively that Kingsoft Cloud will accelerate the construction of GPU computing clusters in the second half of this year to meet the explosive growth in computing demand from its top-tier clients.
Among them, Xiaomi’s GPU computing demand on Kingsoft Cloud has already jumped from a 10,000-card cluster to a hyperscale computing cluster, with the related investment budget soaring from an initial nearly 4 billion yuan to over 10 billion yuan. Meanwhile, Alibaba’s large model team has signed a five-year computing lease deal with Kingsoft Cloud, covering more than 3,000 eight-GPU servers. Based on the monthly rental rate at the time of signing, the monthly revenue after full delivery would be about 300 million yuan, with annualized revenue topping 4 billion yuan.
NUPIAO reached out to Xiaomi and Alibaba for comment, but neither had responded as of press time.
As NUPIAO previously reported, at the China Development Forum 2026 Annual Conference on March 22, Xiaomi founder Lei Jun stated that over the next five years, Xiaomi’s total R&D investment in AI, robotics, and intelligent driving will exceed 200 billion yuan. Over the next three years, the company plans to invest over 60 billion yuan in AI alone, and in 2026, AI-related R&D and capital expenditure will surpass 16 billion yuan.
NUPIAO has learned that Kingsoft Cloud’s management told several institutions recently that the lion’s share of Xiaomi’s computing procurement will land on Kingsoft Cloud, and the company’s 2026 capital expenditure plan has been raised to 15 billion yuan. The specific funding arrangement: 9 billion yuan will be used for computing purchases through its own cash, and 6 billion yuan will be in the form of four-year capitalized lease right-of-use assets, with annual cash lease payments of 1.5 billion yuan. For the full year, the company’s actual cash outlay will total 3.5 billion yuan, consisting of 2 billion yuan in cash purchases plus 1.5 billion yuan in lease payments.
The cooperation model between Xiaomi and Kingsoft Cloud works like this: Xiaomi prepays the computing lease fees, which Kingsoft Cloud then uses for computing procurement. Xiaomi’s contribution covers 20%–30% of the total spending, with the rest coming from financial loans or external financing. Based on Xiaomi’s current 10-billion-yuan investment level, Kingsoft Cloud’s theoretical procurement scale could exceed 40 billion yuan, and the total procurement scale over the next two years could break through the 100-billion-yuan mark.
A source close to Xiaomi told NUPIAO that there are no regulatory or technical hurdles in raising the related-party transaction ceiling between Xiaomi and Kingsoft Cloud—it only requires a routine filing. This means that as Xiaomi’s computing demand dynamically ramps up, its computing spending will directly translate into real revenue for Kingsoft Cloud.
Currently, Kingsoft Cloud’s core computing demand mainly comes from the Xiaomi ecosystem, including full-process training and inference for the MiMo large model, computing adaptation for Xiaomi cars going overseas, large model upgrades for all categories of XiaoAI devices, and WPS Office AI. Xiaomi is also Kingsoft Cloud’s largest customer, with related business demand accounting for about 30% of Kingsoft Cloud’s total revenue, and Xiaomi’s contribution to intelligent computing revenue already exceeding 50%.
Additionally, Kingsoft Cloud is rolling out Xiaomi’s MiMo V3 next-generation training cluster, which has been put into operation in batches. The company has shifted all “old machines” to inference computing, while newly purchased equipment is prioritized for MiMo’s new-generation model training.
NUPIAO understands that Kingsoft Cloud’s overall revenue is divided into two major segments: public cloud and industry cloud. AI-related revenue in the public cloud segment already accounts for more than half of the total revenue, and the top twenty to thirty clients contribute over 80% of the company’s revenue.
Besides big customers like Xiaomi and Alibaba, large model vendors such as Zhipu, Baichuan Intelligence, and MiniMax can currently bring in about 200 million yuan of monthly revenue for Kingsoft Cloud. It’s also worth noting that Kingsoft Cloud has secured the robotics computing demand of Unitree Robotics.
The above-mentioned source disclosed that Kingsoft Cloud’s 2026 revenue target is 12.5 billion to 13.5 billion yuan. Considering that computing power prices across the cloud industry have risen by an average of 50%–60%, Kingsoft Cloud’s capex cash payback period has shortened from four years to two to three years after the price increase, and actual revenue growth could exceed the currently set target range.
Benefiting from Xiaomi’s recent proactive reduction in MiMo token pricing, market call volume has surged, driving a continuous increase in Kingsoft Cloud’s computing lease revenue, with especially remarkable growth in inference business.
On the flip side, due to limited computing power, users can already clearly feel that MiMo’s access speed has dropped and compute supply is insufficient. At present, Xiaomi has only opened MiMo calls on a limited basis in high-end phones and Xiaomi car product lines. Once XiaoAI is fully open across all devices and the MiClaw phone assistant is fully rolled out, inference demand will experience exponential growth.
In response, Kingsoft Cloud’s management told institutions and investors: “The overall market demand for large model tokens is extremely strong right now. Top AI labs and embodied intelligence vendors are generally facing a shortage of computing supply. The only thing holding back our MaaS growth is the hardware procurement cost and availability. It’s just that at this stage, computing power prices are a bit high, so we have chosen to hold off on expansion for now.”
This statement also reveals the current “dilemma” for Kingsoft Cloud: even though 3 billion yuan in funds were deployed for computing purchases in the first quarter of 2026, the company still cannot meet the demand of several top clients.
In the eyes of Kingsoft Cloud’s management, stockpiling cards at current high prices carries significant risk. They judge that computing hardware prices will hit a turning point and start to decline in the third quarter of 2026. In contrast, some aggressively expanding computing leasing companies in the A-share market that rely on large credit lines and long-cycle high-price card hoarding face huge asset impairment risks due to hardware iteration and price drops. Kingsoft Cloud will not emulate such an aggressive model.
NUPIAO learned from the supply chain that, affected by the tight supply of upstream computing cards, Kingsoft Cloud’s traditional CPU server resources have been completely exhausted. GPU computing now only accepts clients on three- to five-year long-term contracts, and one-year contracts have been suspended. The 1,500 eight-GPU servers originally scheduled for delivery to Alibaba in the first half of the year have seen a delivery completion rate of less than one-third. Meanwhile, orders from Xiaomi and others are also facing varying degrees of delivery delays.
From the perspective of core technical barriers, only four vendors in China can independently deliver 10,000-card-level intelligent computing clusters: Alibaba Cloud, Baidu Intelligent Cloud, Kingsoft Cloud, and Range Technology. A 10,000-card cluster has three hard thresholds: large-scale capital reserves, large-scale automated monitoring and deployment capabilities, and high-stability SLA (Service Level Agreement) operation and maintenance capabilities.
According to Kingsoft Cloud, a new type of high-margin managed service has emerged in the market: clients purchase their own hardware, and Kingsoft Cloud can achieve a computing SLA availability rate of 99.5%. Many clients voluntarily co-locate their entire equipment in Kingsoft Cloud’s northwest data centers, where Kingsoft Cloud provides a full suite of operation and maintenance services including IDC facilities, intelligent computing networking, container scheduling, and supporting storage. Several projects have already been signed and put into operation.
