Foreword:

The most noteworthy signals in the AI industry over the past year lie in the cash flow statements, capital expenditure plans, and data center orders of tech giants.

As AI competition evolves, cloud computing is no longer just a backend resource for hosting applications but is transforming into a new type of infrastructure for the AI era.

Author | Fang Wensan

Image Source | Internet

Capital Expenditures Surge as AI Moves into the Infrastructure Realm

The capital market has long been accustomed to pricing internet companies based on asset-light logic, but AI cloud is pulling tech giants back into an asset-heavy cycle.

Microsoft disclosed in its FY26 Q3 results that capital expenditures reached $31.9 billion for the quarter, with about two-thirds allocated to short-cycle assets like GPUs and CPUs.

During the same period, Microsoft's cloud revenue hit $54.5 billion, up 29% year-on-year; Azure and other cloud services revenue grew by 40%; AI business annualized revenue run rate exceeded $37 billion, up 123% year-on-year.

Microsoft's AI revenue has begun to form quantifiable contributions, but behind this revenue lies even heavier investment in computing power assets.

Alphabet, Google's parent company, saw capital expenditures reach $35.7 billion in Q1 2026, with the vast majority supporting technical infrastructure for AI opportunities. About 60% went to servers, and 40% to data centers and network equipment.

Google Cloud's revenue for the quarter reached $20 billion, up 63% year-on-year; cloud backlog orders approached $462 billion.

Alphabet also raised its full-year 2026 capital expenditure guidance to $180-190 billion.

The simultaneous expansion of cloud revenue, orders, and capital expenditures indicates that AI cloud has shifted from 'strategic investment' to 'order-driven investment.'

Amazon's signals are equally direct. In Q1 2026, AWS revenue hit $37.6 billion, up 28% year-on-year—the fastest growth in 15 quarters—with AWS operating profit reaching $14.2 billion.

Amazon also disclosed that its free cash flow over the past 12 months dropped to $1.2 billion, primarily due to a $59.3 billion year-on-year increase in net property and equipment purchases, reflecting AI investments.

It even mentioned that the annualized revenue run rate for chip businesses like Graviton, Trainium, and Nitro has exceeded $20 billion.

Meta's path is more unique. Without external cloud revenue pillars like AWS, Azure, or Google Cloud, it still raised its 2026 capital expenditure guidance to $125-145 billion, citing rising component prices and increased data center costs for future capacity.

Domestic players are also doubling down in the same direction. Alibaba disclosed that it has invested approximately RMB 120 billion in capital expenditures over the past four quarters to advance AI and cloud infrastructure.

Alibaba Cloud's revenue maintains high growth, with AI-related product revenue achieving triple-digit growth for multiple consecutive quarters.

Tencent's capital expenditures in Q1 2026 reached RMB 31.9 billion, up 16% year-on-year. Tencent stated that its core businesses provide cash flow to support AI investments and linked cloud service growth to AI-related demand.

The More AI Industrializes, the More Cloud Providers Resemble 'Smart Power Plants'

In the mobile internet era, cloud computing solved storage, bandwidth, elastic computing, and application deployment issues.

In the era of large models, the delivery method of AI capabilities has changed, beginning to handle more complex tasks.

Training models, running inference, managing vector databases, orchestrating multi-model services, connecting enterprise data, deploying agents, monitoring security compliance, calculating token costs, and packaging these capabilities into APIs, platform services, and industry solutions.

What enterprises truly purchase is a complete set of 'callable, measurable, scalable, and supervision -able' (regulatory-compliant) intelligent production capabilities. The value of AI cloud lies in folding complex hardware, models, data, and engineering capabilities into services that enterprises can directly use—this is also the deep reason why major players must invest in AI cloud.

Model capabilities will rapidly diffuse, open-source ecosystems will continuously catch up, and barriers to single applications are easily replicated.

Once a cloud platform binds a client's data, workflows, development tools, permission systems, industry knowledge bases, and cost models, migration costs rise significantly.

Models are the entry point, cloud is the foundation, but workflows are where clients are locked in long-term.

AI cloud also has an easily underestimated value: it transforms computing power from a one-time purchase into continuous consumption.

Enterprises no longer need every company to build its own 10,000-card cluster or independently bear the risks of chip procurement, data center construction, and operations.

Cloud providers invest heavily in construction and recover investments through pay-as-you-go, annual/monthly subscriptions, computing power leasing, model services, and industry applications.

In the future, what will truly continuously consume computing power are massive enterprise applications, intelligent agent tasks, and real-time user interactions—all consuming tokens, electricity, memory, network, and storage.

Chip procurement capabilities, cluster interconnection efficiency, scheduling systems, model compression, inference frameworks, cache hit rates, energy management, liquid cooling solutions, and data center location all affect final costs.

The reason cloud providers are developing their own chips is precisely to reduce dependence on single hardware supply chains and retain control over cost curves.

Amazon's Trainium, Google's TPU, Microsoft's AI computing power system built around Azure, Alibaba's Tongyi model combined with Alibaba Cloud, and Tencent's infrastructure reconstruction around Hunyuan and industrial scenarios all point to the same issue: AI-era cloud providers must not only be resource lessors but also system integrators of computing power, models, frameworks, toolchains, and application scenarios.

The Core of New Infrastructure Revolves Around Ecosystems and Rule-Setting Power

At first, AI was just a new type of workload on the cloud, with clients renting GPU servers to run their own model training and inference tasks.

But as AI applications deepen, this 'computing power leasing' model no longer meets enterprise needs.

What clients buy is no longer just computing power but a complete AI production system that can be directly implemented.

As of March 2026, China's AI large models had surpassed 140 trillion daily token calls, growing over a thousandfold from approximately 100 billion in early 2024.

When token consumption expands at a thousandfold rate, demand for underlying computing power becomes almost insatiable.

The hundreds of billions in capital invested by major players are also preparing 'ammunition' for this exponential growth in token consumption.

The old model of cloud providers 'building data centers and waiting for clients to rent servers' is failing. Future competition will hinge on who can provide more efficient, cheaper, and more stable computing power services at the token level.

Alibaba Cloud has built a complete vertical integration chain: from T-Head's self-developed GPU chips at the bottom, to Lingjun intelligent computing clusters and cloud operating systems in the middle, to Tongyi Qianwen large models and Bailian MaaS platforms at the top, and finally landing in DingTalk's enterprise-grade AI agents and Qianwen App's consumer-facing applications.

As of February 2026, T-Head's self-developed GPUs had cumulatively shipped 470,000 units, with over 60% serving external commercial clients and supporting AI tasks for more than 400 enterprise customers.

Baidu Intelligent Cloud relies on a four-layer full-stack architecture of 'Kunlunxin chips + PaddlePaddle framework + Wenxin large models + Qianfan platform,' ranking first in the self-developed GPU cloud market with a 40.4% share.

Kunlunxin P800 has completed large-scale validation, delivering multiple 10,000-card clusters, with fully domestic clusters completing training for Wenxin large model version 5.1.

Volcano Engine adopts a more internet-thinking approach, taking MaaS to the extreme. Through Doubao large model's extreme cost-effectiveness and ByteDance's massive user base, model capabilities are 'exercised' and fully validated before external output.

In IDC's 2025 enterprise MaaS call volume statistics, Volcano Engine expanded its lead for three consecutive years with a 49.5% share.

These three paths converge at the same endpoint: cloud providers that control token production and distribution will hold the 'measurement rights' and 'settlement rights' for AI applications.

Gartner predicts that by 2027, cloud-based workloads will account for 80% of all AI inference workloads in China, up from 20% previously.

This means most enterprises' AI applications will run on the cloud, and cloud providers will become de facto AI infrastructure operators.

This shift in status transforms cloud providers from optional suppliers to infrastructure providers—like today's power and water companies—becoming indispensable foundational elements of economic operation.

Unlike traditional infrastructure, however, AI cloud's infrastructure attributes are compounded with data, models, and ecosystem barriers, making its moat far deeper than traditional infrastructure.

Conclusion:

The influx of capital from major players into AI cloud represents their race to secure infrastructure entry points for the next-generation digital economy, marking AI's entry into a more pragmatic phase.

For the Chinese market, AI cloud holds additional significance as a critical foundation for computing power security and industrial upgrading.

With uncertainties surrounding the supply of external high-end chips, relying solely on overseas GPU routes cannot support a long-term stable AI application ecosystem.

Domestic AI cloud providers need to form closed loops in domestic chips, heterogeneous computing power scheduling, model adaptation, inference optimization, and industrial data governance.

This path won't be easy in the short term but will force the co-evolution of software stacks, compilers, communication libraries, model architectures, and application frameworks.

Partial references: TrendForce: '2026 Global Cloud Provider Capital Expenditure Forecast Report,' Gartner: '2026 Global AI Spending Forecast,' Morgan Stanley: 'Global AI Infrastructure Construction Cycle Analysis,' Alter Tech Talk: 'From Alibaba to Baidu, Major Players Compete for AI Cloud 'New Infrastructure' Dividends'