While the capital market still habitually uses the single metric of "year-over-year growth rate" to measure tech giants, the true logic of competition has already shifted silently.
Over the past decade, the core of the cloud narrative was "scale"—whoever built more data centers and achieved broader coverage; but looking back from the vantage point of Q4 2025, the essence of the story has shifted to "conversion"—that is, who can truly transform massive AI computing power investments into a sustainable and healthy cash flow machine.
In Q4 2025, the world's top four cloud providers submitted their reports simultaneously. These financial reports may seem like routine scorecards, but they are actually watershed declarations. The results show drastic divergence: some are "printing money" by leveraging technological barriers, some are "burning cash" to maintain their ecosystems, and others are quietly completing a strategic repositioning under a seemingly low-profile facade.
This is not merely a fluctuation in financial data, but rather the pain and rebirth of the global cloud computing industry as it transitions from the "infrastructure era" to the "intelligence era".
The "Surface Prosperity" of Cloud Computing and "Structural Divergence"
"Both Are Growing, So Why Is AWS Getting Stronger While Azure Starts to Lose Speed?"
At first glance, the data from the past quarter presents a prosperous picture of "three clouds flying together." Amazon AWS revenue grew by 24% and showed an accelerating trend, Google Cloud took a significant lead with a 48% growth rate, and Microsoft Azure also maintained a high-level growth of 39%.
However, stopping merely at percentage figures would cause one to miss the most critical signal of this cycle. If the data is broken down to the level of profit structure and capacity allocation, a profound structural change emerges: cloud computing is evolving from "general infrastructure leasing" into a "precision business of AI capacity".
In this transformation, the changes in AWS are the most typical and representative. Although its cloud business revenue accounts for only 17% of the group's total revenue, it contributes over 50% of the operating profit. Behind this data is AWS's ultimate control over underlying computing costs. The scaled monetization of self-developed chips Trainium and Graviton not only reduces reliance on external suppliers like NVIDIA, but more importantly, it gives AWS the initiative in pricing power. Currently, the capacity of its next-generation chip Trainium3 has been locked in advance by core customers. This model of "booking future capacity" marks AWS's upgrade from passively "selling computing power" to proactively "selling deterministic AI capacity." Customers are no longer buying elastic servers, but stable intelligent productivity for the next three years.
The explosive logic of Google Cloud is similar, but its path leans more towards being technology-native. 70% of its customers are already using AI-related products, and actual usage exceeds committed quotas by 30%, indicating that the demand is real and urgent. More importantly, the Gemini model has evolved from a mere tool into a traffic entry point, boasting 750 million monthly active users. This means Google Cloud possesses the unique ability to feed back from the C-end to the B-end. AI is transforming cloud business from a corporate "cost center" into a "demand amplifier".
However, the problem lies with the seemingly steady Azure. Behind the superficial 39% high growth hides a reality not much discussed in analyst meetings: its growth rate is beginning to be constrained by internal capacity allocation conflicts. With limited GPU resources, Microsoft has to prioritize supply for internal products, such as Copilot and OpenAI's training needs. This leads to a dilemma where external enterprise customers face their resources being "crowded out" when they need to expand computing capacity. This brings about a core contradiction: in the AI era, Microsoft is experiencing a significant "fight between its left and right hands" for the first time. When the cloud business becomes a "blood supplier" for the internal AI strategy rather than a "profit center," the sustainability and independence of its growth are called into question.
AI Capital Expenditure Spirals Out of Control, Cloud Providers Enter an "Arms Race"
"Behind the $200 Billion Capital Expenditure: Cloud Providers Are Collectively Betting on a Heavier Future"
If the divergence on the revenue side still seems mild, the capital expenditure (CapEx) data completely reveals the industry's anxiety and frenzy. In 2025, the investments of cloud providers are no longer calculated in "hundreds of millions," but in "hundreds of billions" as the unit. AWS expects its 2026 capital expenditure to reach the scale of $200 billion, with Google close behind at between $175 billion and $185 billion, and Microsoft throwing out a staggering figure of $37.5 billion for a single quarter alone.
This is not just a game of numbers; it marks the official entry of cloud computing into the heavy-asset era of "power + chips." In the past, cloud providers competed on the number of servers; now, the competition is about who can get more electricity and who can build more data centers.
A key data point disclosed by AWS is highly symbolic: it added 3.99GW of new power capacity in 2025 and plans to double it by 2027. The subtext behind this is: the future cloud competition is essentially about "who controls more combinations of energy and computing power." To secure electricity, Amazon has even started directly signing nuclear power supply agreements with energy companies, a cross-boundary integration that was unimaginable a few years ago.
However, facing the same heavy-asset pressure, the three giants have chosen vastly different paths, which also determines their future risk exposures.
AWS follows the "ultimate external supply model." It sells most of its AI computing power directly to customers, essentially acting as a "computing power lessor." The advantage of this model is the strong certainty of cash flow; as long as someone rents it, the machines run and the money flows in. But the risk lies in cyclical fluctuations. Once the explosion of downstream applications falls short of expectations, massive computing power assets will face the risk of idle depreciation.
Microsoft chooses the "internal priority model." GPU resources are prioritized for Copilot, OpenAI, and self-developed products. The advantage of this model is the ability to build the tightest ecological closed loop, ensuring a leading experience for its own AI products. But the risks are equally obvious: the growth of the cloud business is suppressed by internal demand, and if internal AI products cannot cover hardware costs through subscriptions, the cloud department's profit margin will be under long-term pressure.
Google, on the other hand, adheres to the "technology-driven model." Relying on its self-developed TPU system, it emphasizes the integration of data and models, and TPUs are mainly not sold externally. This gives it a moat in technological leadership, but the pace of commercialization is often constrained by the speed of technological iteration, easily leading to a situation of "strong technology, slow monetization."
A more hidden but critical change is taking place: cloud providers are shifting from "selling services" to "selling computing power financial products." Customers are required to sign long-term contracts in advance and lock in capacity, which is essentially similar to futures contracts. This financialization trend, while locking in revenue, also means that cloud providers must bear heavier asset delivery responsibilities. Once a disruptive change occurs in the technological route (such as a significant drop in inference costs), these locked long-term contracts could turn from assets into burdens.
The Underestimated Variable—Alibaba Cloud's "Turning Point"
"While Global Giants Are Frantically Burning Cash, Alibaba Cloud Is Instead Walking a Lighter Path"
While the global top three giants are frantically increasing leverage and competing to see who has heavier heavy assets, Alibaba Cloud presents a "different rhythm." Q4 2025 data shows that Alibaba Cloud's revenue growth rate reached 36% and continues to accelerate, with AI-related revenue maintaining triple-digit growth for 10 consecutive quarters, yet its capital expenditure is far lower than that of overseas giants.
This seemingly "restrained" investment may seem like conservative defense at first glance, but it actually hides a larger variable. This is not due to insufficient demand, but a strategic choice based on the characteristics of the Chinese market.
First, the "AI penetration dividend" in the Chinese market is just beginning. Compared to the US market, there is still huge room for the cloud adoption rate of Chinese enterprises, and the commercial application of AI is on the eve of an initial explosion. This means that Alibaba Cloud does not need to make excessive forward-looking infrastructure investments to compete for the stock market like US giants. It faces an incremental market, allowing for a more relaxed pace and a greater focus on return on investment (ROI).
Second, MaaS (Model as a Service) is restructuring Alibaba Cloud's revenue model. Alibaba has explicitly proposed that future MaaS revenue will surpass traditional IaaS (Infrastructure as a Service). This is a qualitative leap. It means Alibaba Cloud is no longer satisfied with just being the underlying "landlord" collecting fixed "utility bills," but is directly participating in the "distribution of model value." Through self-developed models like Tongyi Qianwen, Alibaba Cloud can deeply integrate into customer business scenarios, sharing profits based on performance and call volume. Compared to AWS earning standardized "resource leasing fees," Alibaba Cloud wants to earn "application revenue sharing" in the future. This model may not be as large in scale as IaaS in the early stages, but its profit margin and customer stickiness will far exceed the former.
Third, an advantage ignored by the outside world lies in the cost structure. When overseas cloud providers face the triple pressure of surging GPU prices, rising power costs, and massive depreciation swallowing profits, Alibaba Cloud's path leans more towards "software-hardware synergy." By optimizing inference costs through self-developed models and utilizing architecture designs closer to the application layer, Alibaba Cloud can provide lower-cost inference services under the same computing power. For price-sensitive small and medium-sized enterprises in China, this has a fatal attraction. This may bring higher ROI rather than simply pursuing larger scale. Today, when AI inference costs determine the speed of application popularization, whoever can bring down inference costs will take the initiative in the application ecosystem.
Conclusion: Efficiency is King, the Victory of the Asset-Light Model
"The Winner in the Next Phase Is Not the One with the Most Computing Power, but the One with the Highest Monetization Efficiency"
Looking back at this round of cloud computing financial reports, the truly noteworthy indicators are no longer the simple revenue growth curves, but who is closer to the ultimate form of "turning AI into profit".
The current landscape is already clear: relying on self-developed chips and the external supply model, AWS was the first to run through the cash flow model, making it the current "cash cow"; Google Cloud is using its technological advantages to catch up in scale and ecosystem, acting as the "technology chaser"; Azure, due to internal resource gaming, has fallen into a growth bottleneck, acting as the "ecosystem gamer"; and Alibaba Cloud, standing at a starting point underestimated by the market, is the "efficiency transformer".
If the time frame is extended to 3 to 5 years, the global cloud computing landscape will likely no longer be the traditional "three-way battle." The industry will split into a duel between two models: one is the "heavy-asset computing power empire," which relies on massive capital expenditures to build barriers, betting on the infinite explosion of AI demand; the other is the "light model + application ecosystem," which relies on model optimization and the landing of application scenarios, betting on the precise monetization of AI value.
The heavy-asset model has a scale advantage in the early stages, but with the arrival of power bottlenecks and depreciation pressures, its marginal benefits will diminish. The asset-light model, although seemingly slow to start, is closer to business value, has stronger risk resistance, and is more likely to form network effects in the long-tail market.
In this long marathon, computing power scale is just the ticket to enter, while monetization efficiency is the finish line. The path represented by Alibaba Cloud may well foreshadow the true winner logic in the second half of cloud computing: it is not about how many watts of power you have, but how much commercial value you can generate from every watt of power. When the tide goes out, those providers who can leverage greater application value with lighter assets will eventually cross the cycle and become the new definers.