This content is reprinted and compiled to present different market perspectives and research viewpoints, and does not imply that this official account endorses the views and conclusions in the text. Source | US Stock Research Society
AI cloud computing is entering a phase where it is harder to fool the market: capital expenditures (Capex) are still surging, but investors have started to demand revenue, profit margins, and cash flow. In the first quarter of 2026, overseas cloud vendors delivered straightforward results. Google Cloud's revenue grew 63% year-over-year to $20 billion, with cloud business operating profit reaching $6.598 billion; AWS revenue increased 28% year-over-year to $37.6 billion, with operating profit hitting $14.2 billion; Oracle's third-quarter RPO (Remaining Performance Obligations) reached $553 billion, an increase of 325% year over year, a significant portion of which came from large-scale AI contracts. The AI Capex poured in by cloud vendors over the past few years is now becoming visible on both the revenue and order fronts.
This will also pull the domestic cloud computing industry into the same evaluation framework. In 2024, the market was buying the story of "who has GPUs and who can get chips"; in 2026, capital cares more about who can turn GPUs into recurring bills, customer renewals, and EBITDA. Kingsoft Cloud's transformation hits right at this node. In the fourth quarter of 2025, the company's revenue hit a new single-quarter high, with public cloud revenue growing 34.9% year-over-year, AI business billing revenue surging 95% year-over-year, and Non-GAAP EBITDA increasing 118.3% year-over-year. This is not just a simple "AI concept stock" story; rather, a company long suppressed by the "second-tier cloud" label is now being re-examined by the market within the framework of "AI cloud monetization capability."
The Exam for Cloud Vendors Has Changed: From "Dare to Invest" to "Does the Math Work"
Over the past two years, the core narrative of AI cloud has been brutal: compute power is insufficient, so just invest first. Microsoft, Google, Amazon, and Meta have continuously raised capital expenditures, leading to successive revaluations of servers, GPUs, optical modules, IDCs, and the power supply chain. During that phase, the market defaulted to a premise: as long as AI demand is large enough, Capex will eventually be absorbed. Investors were willing to buy the future first, letting the financial statements catch up later.
In 2026, the atmosphere has changed. AI demand has not cooled down; instead, it is entering cloud vendors' revenue statements more densely. Alphabet's total revenue in the first quarter grew 22% year-over-year, and Google Cloud's revenue increased 63% year-over-year. Pichai also mentioned that first-party models like Gemini are directly used via customer APIs, processing over 16 billion tokens per minute, a 60% increase from the previous quarter. The meaning of this set of data is clear: AI is no longer just a centralized expense for training large models; inference, API calls, enterprise applications, and Agent workflows are beginning to generate more sustained cloud revenue.
AWS is on the same track. Amazon's AWS revenue in the first quarter grew 28% year-over-year, marking the fastest growth in 15 quarters; AWS operating profit increased from $11.5 billion in the same period last year to $14.2 billion. More interestingly, Amazon disclosed that its chip business annualized revenue has exceeded $20 billion, securing about 2GW of Trainium capacity commitments from OpenAI, and up to 5GW of such commitments from Anthropic. Cloud vendors are no longer just leasing general-purpose servers; they are integrating chips, models, compute scheduling, and cloud services into a comprehensive delivery capability. Oracle's case better illustrates the shift in valuation anchors. Its third-quarter RPO reached $553 billion, an increase of 325% year over year, with cloud infrastructure revenue growing 84% year-over-year. The company also noted that cloud computing demand for AI training and inference is growing faster than supply, and some large AI contracts have reduced Oracle's own financing pressure through customer prepayments or customers bringing their own GPUs.
When looking at Oracle, the capital market no longer just sees the legacy business of the old database giant, but whether it can turn AI infrastructure orders into recognizable revenue over the next few years. This is the gear shift in the AI cloud trading logic. In 2024, the market rewarded "speed of investment"; in 2026, the market rewards "revenue conversion efficiency." Whoever has higher capacity utilization rate, longer customer contracts, and can spread depreciation pressure into recurring bills will more easily get a valuation premium. This change is critical for domestic cloud vendors.
The Chinese cloud computing industry has been weighed down by price wars in recent years. Public cloud, government and enterprise cloud, CDN, and storage have all experienced round after round of price cuts. The market's judgment on cloud vendors has been pessimistic: the larger the scale, the heavier the capital expenditures, and the uglier the profits. AI inference demand has given the industry a chance to reprice. Training is like project-based orders, with obvious peaks and valleys; inference is more like water, electricity, and gas, with high call frequency, scattered scenarios, and continuous budgets. Enterprises do not necessarily all train large models, but they will continuously call code assistants, customer service robots, office agents, marketing generation tools, and data analysis Agents.
The broader the token consumption, the easier it is for cloud vendors to turn compute assets into recurring billable services. Kingsoft Cloud is right in this transition. It does not have the ecological breadth of Alibaba Cloud or Tencent Cloud, nor the government, enterprise, and hardware depth of Huawei Cloud, but the second phase of AI cloud looks at more than just scale. The market is shifting to more specific questions: Does this company have stable AI customers? Does it have available compute power? Does it have platformized tools? Are there signs of profit margin recovery?
Kingsoft Cloud's fourth-quarter 2025 financial report provides a preliminary answer.
Kingsoft Cloud's Opportunity Lies in the "Inference Cost-Effectiveness" Hidden in the Cracks of Giants
Kingsoft Cloud's biggest label in the past has been a second-tier cloud vendor. This label is not unearned. In the traditional cloud era, cloud computing competition fought over ecosystems, customers, capital expenditures, and prices. Alibaba Cloud has the foundation of e-commerce and enterprise services, Tencent Cloud has the content and social ecosystem, and Huawei Cloud has advantages in hardware, government and enterprise, and industry digitalization. Kingsoft Cloud has long sought space among video, gaming, office, and government and enterprise projects, lacking scale advantages and facing considerable profit pressure. AI cloud has dismantled the rules of the game.
What enterprise customers want now is not necessarily the most comprehensive cloud product shelf, but faster access to AI compute power, more stable inference task execution, lower-cost access to model APIs, and more flexible management of multi-model and heterogeneous resources. Such demands have given "specialized cloud vendors" a window. The rise of Neoclouds in North America is essentially the same: companies like CoreWeave and Nebius do not have the complete ecosystem of traditional cloud giants, but they can cut into the market with AI-dedicated clusters, Nvidia resources, and customer customization capabilities. CoreWeave's market attention is driven not just by high-growth revenue, but also by long-term contracts and order visibility at the tens of billions of dollars level.
Kingsoft Cloud cannot be simply compared to CoreWeave. The domestic chip supply, customer structure, and cloud market pricing system are all different. But it is indeed starting to move closer to the direction of an "AI infrastructure service provider." A very important change is the upgrade of the Xingliu platform. Kingsoft Cloud disclosed that the Xingliu platform has been upgraded from a resource management platform to a one-stop AI training and inference full-process platform, covering heterogeneous resource scheduling, fault self-healing for training tasks, commercialization of model API services, and more, supporting nearly 40 models including DeepSeek, Xiaomi MiMo, Qwen3, and Kimi. This cannot be seen merely as product packaging.
The profit margin of AI cloud cannot ultimately rely solely on "selling card hours." Pure compute leasing will be caught up in price wars, and hardware depreciation will continue to eat into profits. To improve revenue quality, cloud vendors need to encapsulate compute power upwards into platforms: model management, inference services, scheduling optimization, API calls, enterprise security, and industry application adaptation. What customers buy is not a piece of GPU, but AI capabilities that can run businesses. Another layer of advantage for Kingsoft Cloud comes from the Kingsoft and Xiaomi ecosystems. Lei Jun resigned as a non-executive director and chairman of Kingsoft Cloud in March 2026, succeeded by Zou Tao. In the short term, this is a change in the governance layer, but in the long term, it actually makes Kingsoft Cloud's strategic responsibilities clearer: it can no longer rely solely on the "Lei Jun circle" story to gain market attention, but must deliver on its own AI cloud business. However, ecological synergy remains important. Xiaomi's AI devices, smart cars, IoT (Internet of Things), smartphones, and model applications will bring real inference scenarios; Kingsoft Office's WPS AI, office agents, and enterprise collaboration will also generate stable token calls.
What AI cloud fears most is idling. Internal ecosystem demand can at least help Kingsoft Cloud complete the early-stage utilization rate climb. This is also the reason the market is refocusing on Kingsoft Cloud. It has not suddenly become the Chinese version of CoreWeave; rather, the previously underestimated combination of "customer scenarios + compute resources + model platform" has caught up with the trend of inference commercialization. The financial reports have already shown some marginal signals. In the fourth quarter of 2025, Kingsoft Cloud's total revenue was 2.761 billion CNY, a 23.7% year-over-year increase; public cloud revenue was 1.902 billion CNY, a 34.9% year-over-year increase; AI business billing revenue was 926 million CNY, a 95% year-over-year increase.
For the full year, Kingsoft Cloud's revenue reached 9.559 billion CNY, a 22.8% year-over-year increase; Non-GAAP EBITDA reached 2.336 billion CNY, with the EBITDA margin rising to 24.4%, significantly higher than 8.2% in 2024. The most worth dissecting in this set of data is the "quality of growth." The acceleration in revenue comes from AI demand, and public cloud has once again become the main engine of growth; the substantial improvement in EBITDA indicates that economies of scale and expense control are starting to work; Non-GAAP operating profit has been positive for two consecutive quarters, indicating that the company is one step closer to stable profitability at the operational level. However, the expansion of the AI business has also pushed up costs.
In the fourth quarter, Kingsoft Cloud's IDC costs increased 12.5% year-over-year, and depreciation and amortization costs rose from 343 million CNY in the same period last year to 741 million CNY, mainly from the depreciation of servers and network equipment related to the AI business. The full-year gross margin dropped from 17.2% in 2024 to 15.7% in 2025, with pressure also coming from server and equipment depreciation. This is the typical financial profile of an AI cloud company: revenue goes up first, and depreciation goes up first; before the utilization rate is fully loaded, the income statement will be suppressed by hard costs. The market is now repricing Kingsoft Cloud, betting not that it has no cost pressure, but that AI inference demand is sustained enough to dilute these hard costs.
The Monetization Window Has Opened, but Kingsoft Cloud Still Has Three Hurdles to Clear
Kingsoft Cloud's story has reached a point where it cannot be written as "AI arrives, second-tier clouds turn the tide." What it really needs to prove are three things.
First, can the AI business turn high-growth bills into high-quality revenue? The AI business billing revenue in the fourth quarter grew 95% year-over-year, which is a striking speed. But the capital market will continue to look ahead: Do these orders come from inside or outside the ecosystem? How long is the customer renewal cycle? What is the proportion of inference tasks? What is the per-GPU utilization rate? Can model API calls bring higher gross margins? If AI revenue is just a short-term project-based explosion, valuation recovery will be very limited; if it can settle into continuous calls and long-cycle contracts, Kingsoft Cloud's valuation anchor will shift from "distress recovery" to "AI cloud growth."
Second, can compute expansion keep depreciation and financing costs in check? AI cloud is not an asset-light business. GPU procurement, server leasing, IDC cabinets, network bandwidth, storage, and O&M teams all require real money. At the end of 2025, Kingsoft Cloud's cash and cash equivalents reached 6.018 billion CNY, a significant increase from the end of the third quarter, mainly from net equity financing, while being offset by capital expenditures. The company's net property, plant and equipment rose from 4.63 billion CNY at the end of 2024 to 10.095 billion CNY at the end of 2025, indicating rapid asset expansion. This is both an opportunity and a pressure. The opportunity lies in the fact that the company has deployed resources in advance for AI demand; the pressure is that once customer utilization falls short of expectations, depreciation, interest, and equipment updates will in turn press down on profits.
Third, Kingsoft Cloud needs to find its boundaries under the siege of giants. Alibaba Cloud, Tencent Cloud, and Huawei Cloud will not yield the AI cloud market. Alibaba has the Tongyi and Qwen ecosystem, Tencent has the Hunyuan and WeChat scenarios, and Huawei has Ascend, Kunpeng, and government and enterprise customers. Kingsoft Cloud cannot use the giants' playbook to go head-to-head with them. Its more realistic route is to build a more vertical, flexible, and customer-specific AI infrastructure. This route looks narrow, but it is not bad. AI cloud will enter a layered market.
The top-tier large model training clusters still belong to giants and top AI labs; the middle layer of inference services, model APIs, industry application deployment, office and device ecosystems will generate massive cost-effectiveness demands. What Kingsoft Cloud is truly suited to cut into is this middle layer: customers do not want to build their own infrastructure, nor do they want to be completely bound by giant clouds; they need an AI cloud platform that can run multiple models, control costs, and deliver quickly. This is also where the value of Kingsoft Cloud's Xingliu platform lies.
The more models there are, the more customers need unified scheduling; the more fragmented the applications, the more customers need stable APIs; the more frequent the inference calls, the more customers care about costs and latency. Kingsoft Cloud does not necessarily have to be the protagonist in the model war, but it can be a more pragmatic infrastructure provider in the model deployment link.
From a trading perspective, the most important variable for Kingsoft Cloud has changed. In the past, the market looked at how much it lost, how long its cash could last, and whether second-tier clouds had a way out; now the market is starting to look at the proportion of AI revenue, EBITDA margin, whether the turnaround in operating profit can be sustained, and whether ecosystem customer demand can spill over. This change in the valuation anchor will bring elasticity, but also stricter acceptance. Kingsoft Cloud's true "monetization moment" will most likely not be in the stock price fluctuations of a single quarter, but in the consecutive financial reports of the subsequent quarters of 2026: Can AI billing revenue continue to grow at a high rate? Can public cloud maintain acceleration? Can the gross margin withstand depreciation pressure? Can operating profit get rid of one-off factors? Can cash flow support expansion? The market can trade expectations first, but ultimately it will return to the financial statements.
The Second Phase of AI Cloud Does Not Reward "Having Cards," but Rewards "Knowing How to Do the Math"
In the first phase of AI cloud, the market believed in scarcity. Whoever had GPUs had a story; whoever dared to ramp up Capex could get a valuation. In the second phase, the market begins to believe in the ledger. GPUs are just the starting point; what follows is looking at customers, utilization rates, unit prices, renewals, depreciation, and cash flow. The more expensive the compute assets, the more important the operational capability.
AI cloud sounds like a technology business, but when it comes to finance, it is still a contest of turnover rates and profit margins. Kingsoft Cloud's current highlight lies in this entry point. It is not the largest cloud vendor, nor the most powerful large model company. Its opportunity comes from the demand spillover in the inference era: more enterprises want to use AI, but do not necessarily have to build their own large models; more applications need to call models, but do not necessarily only recognize giant clouds; more scenarios need to control costs, but cannot sacrifice stability. This middle layer of the market is being gradually amplified by AI inference.
This is also the underlying logic for the revaluation of Kingsoft Cloud: in the past, the market discounted it with the "second-tier cloud" label; now, the market is starting to re-examine it with "AI cloud monetization capability." But this revaluation will not be easy. AI cloud is an asset-heavy, high-depreciation, and highly competitive business. If revenue growth cannot outpace depreciation, the story will cool down quickly; if customer stickiness is not strong enough, compute leasing will return to price wars; if platform capabilities cannot be solidified, Kingsoft Cloud will still be squeezed by giants. Therefore, the real test Kingsoft Cloud needs to hand in next is not to prove that it has caught the AI wave, but to prove that it can live more efficiently in the second phase of AI cloud.
The winning hand in AI cloud has shifted from "who buys cards first" to "who can calculate higher revenue from every card." Kingsoft Cloud stands at the window, and the market is willing to take another look at it; whether this look can be turned into long-term valuation depends on whether it can turn AI bills into profits in the coming quarters.