AI capital expenditure is decentralizing: the second restructuring of technology infrastructure is happening.

In 2026, global AI infrastructure investment is standing at a critical structural inflection point.

Over the past three years, the core narrative of the AI computing race has been highly singular: hyperscalers have expanded data centers and purchased GPUs at near-unprecedented costs, pushing capital expenditures to historic extremes. In 2026, the combined capital expenditures of the four major cloud providers—Amazon, Microsoft, Google (Alphabet), and Meta—are expected to reach $725 billion, a 77% year-over-year increase from $410 billion in 2025. If Nvidia, Apple, Tesla, etc., are included in the Magnificent Seven, this figure approaches $754.2 billion. Gartner predicts that global AI spending will reach $2.59 trillion in 2026, a 47% year-over-year increase.

However, scale itself is losing its status as the sole focus. A deeper change is occurring: AI capital expenditure is shifting from highly centralized to distributed. DIGITIMES defines the tech keyword for 2026 as "dispersal," symbolizing the dual transformation of the AI market and supply chain toward decentralization. This is not just geographic dispersion but a comprehensive restructuring of investment entities, technical architectures, and industrial structures.

The End of Centralization: The $725 Billion "Bill" and Return Anxiety

To understand the beginning of distribution, we must first see the peak of centralization.

In 2026, the total capital expenditures of the four hyperscalers are expected to range between $650 billion and $700 billion, accounting for about 40% of the total capital spending of the Russell 1000 Index, doubling from 2024 levels. The specific numbers for each are as follows: Amazon is locked at approximately $200 billion, Microsoft maintains an expectation of $190 billion, Alphabet has raised its forecast to $175-185 billion, and Meta is at $125-145 billion.

The speed of this adjustment itself is an important signal. In just the past six months, market expectations for cloud provider capital spending in 2026 have increased by nearly 80%. Barclays expects major cloud providers' capital spending to reach $919 billion in 2027 and further rise to approximately $1.16 trillion in 2028. CreditSights estimates that in 2026, approximately 75% of the total capital spending of hyperscalers will go to AI-related infrastructure, i.e., about $450 billion in AI-specific spending.

But centralized scale expansion is facing scrutiny over returns. In June 2026 (Beijing time), Microsoft's stock fell nearly 20% within a month, and its market capitalization evaporated by nearly $1.3 trillion over the past eight months. The focus of investor scrutiny is Microsoft's approximately $190 billion capital spending in 2026—about two-thirds of which goes to short-cycle assets like GPUs and CPUs, which depreciate faster and are directly tied to short-term revenue. Microsoft Cloud's gross margin has been guided to 64%, down 4 percentage points year-over-year. A Goldman Sachs report in June pointed out that U.S. tech investment as a share of GDP has risen to approximately 4.9%, surpassing the peak during the internet bubble around 2000.

The marginal returns on centralized investment are diminishing, providing the most direct driving force for distribution.

The Inference Inflection Point: Why Computing Power Must Go Distributed

The underlying logic of AI capital expenditure distribution is first and foremost the change in the structure of computing demand itself.

Nvidia CEO Jensen Huang explicitly stated at GTC 2026 that AI inference workloads will reach a billion times the scale of training, and the era of inference has fully arrived. IDC predicts that by 2027, inference tasks will account for over 70% of total intelligent computing demand. TrendForce's data is more specific: in 2026, AI inference computing power growth rate is as high as 122%, far exceeding the 56% growth rate of AI training computing power.

Training and inference have vastly different requirements for infrastructure. Training is a centralized, high-density, long-duration computing task, naturally suited for deployment in hyperscale data centers. Inference is a distributed, low-latency, high-concurrency real-time response task—when an AI agent needs to complete an inference and return results within tens of milliseconds, the physical latency of transmitting data from the edge to a centralized data center and back becomes an insurmountable bottleneck.

Akamai's architects point out that gaming scenarios require first-token latency to be controlled within 15 milliseconds, e-commerce recommendations within about 20 milliseconds, while the network latency of tens of milliseconds between traditional centralized data centers and end users has become an insurmountable bottleneck for real-time interactive scenarios. Under centralized deployment, 1GW of computing power requires 75 Tbit/s of egress bandwidth (Blackwell), and the next-generation Vera Rubin will require 135 Tbit/s; after distributing to 20 nodes, each node only needs 3.75 Tbit/s. This is an arithmetic problem determined by physical laws, not a choice of business strategy.

At the same time, multimodal interaction brings massive outbound traffic, and the high bandwidth costs of public clouds are becoming a "silent killer" of AI business profitability. Coupled with tightening data localization regulations in regions like the EU's GDPR, Southeast Asia, and the Middle East, centralized deployment is increasingly trapped in a dilemma where experience, cost, and compliance are difficult to balance simultaneously. AI computing power is no longer concentrated only in the core cloud but is evolving towards a three-tier distributed architecture of "core—regional—edge."

From the Big Four to the Entire Industry Chain: Expanding Participants in Capital Spending

The second dimension of distribution is the diffusion of investment entities.

Over the past three years, AI infrastructure investment has been almost entirely dominated by the four major cloud providers and Nvidia. But in 2026, this landscape is changing. Calculations by Zhongtai Securities show that in 2026, the combined AI capital spending of the MAG7 is approximately $754.2 billion, while China's domestic AI capital spending totals approximately 805.8 billion RMB (about $110 billion). Combined, the AI capital spending of China and the U.S. contributes approximately 1,007.6 billion RMB to China's GDP, accounting for 0.68% of GDP, with a marginal contribution of about 0.33 percentage points to GDP growth. AI upstream and downstream have surpassed the urban investment chain to become the marginal increment of GDP growth.

Enterprise participation is accelerating. A recent RBC survey shows that enterprises are accelerating AI adoption, with most having shifted from experimentation to formal production. A survey on AI utilization by Japanese domestic enterprises shows that 47.8% of enterprises have reached the stage of full-scale operation (formal production), with 62.7% of large enterprises achieving full-scale operation. Although the adoption rate among SMEs is still limited (about 12% for Japanese SMEs), the 64.7% adoption rate among large enterprises indicates that enterprise AI deployment has moved from proof of concept to the scaling stage.

The participation of sovereign states is also notable. In June 2026 (Beijing time), Jensen Huang revealed at the shareholder meeting that nearly 40 countries and regions, collectively representing $50 trillion in GDP, are building AI factories driven by Nvidia infrastructure. AI infrastructure investment is evolving from an "internal affair of tech companies" to a "strategic competition at the national level."

Distribution of capital spending is also reflected in financing structures. Zhongtai Securities points out that the capital spending of U.S. AI giants has entered a debt-financing-driven phase. Hyperscalers' capital spending is no longer fully dependent on free cash flow but leverages debt financing. This shift in financing mode means that the sustainability of capital spending no longer depends solely on individual companies' cash flow conditions but is linked to broader credit market conditions.

Edge as the Frontline: The Implementation of Distributed AI Infrastructure

The most concrete manifestation of the distribution trend is in the field of edge computing.

In 2026, edge AI is moving from concept to large-scale deployment. Akamai and Nvidia have jointly launched the "AI Mesh," transforming a network of over 4,400 edge nodes worldwide into a distributed AI inference platform. Akamai is transitioning from a leading global cloud delivery service provider to the world's largest distributed AI inference platform, having deployed NVIDIA Blackwell RTX 6000 PRO GPUs at scale globally.

This transformation is not an isolated case. In June 2026 (Beijing time), edge computing company Yuntian Changxiang completed a Series E financing of over 1 billion RMB, led by the China Internet Investment Fund. The company simultaneously announced an upgrade from "edge computing service provider" to a comprehensive strategy for the AGI era: "real-time computing networking." Antimatter secured 300 million euros for the deployment of the first 100 Policloud distributed micro data centers in 2026. NXP strengthened its edge AI product portfolio by acquiring Kinara, adding standalone NPUs.

IDC predicts that by 2027, over 80% of enterprises will deploy distributed edge infrastructure. The growth rate of edge infrastructure construction will surpass that of core data centers. This means the edge is no longer a supplement to cloud computing but is becoming a core component of AI infrastructure.

The business logic of edge AI is clear: inference tasks are much more sensitive to latency than training tasks, and edge nodes are naturally close to data sources and users. For enterprises, edge deployment also solves multiple issues such as data compliance (data does not leave the country), bandwidth costs (reducing cloud transmission), and reliability (local disaster recovery). These issues are difficult to solve simultaneously under a centralized architecture but find operational solutions in a distributed architecture.

The Multi-Layered Infrastructure Era: A Structural Shift in Investment Logic

AI infrastructure is moving from a "single centralized" to a "multi-layered distributed" structure. The implications of this shift for investment logic are profound.

First, the structure of chip demand is changing. On the training side, Nvidia GPUs remain dominant—Nvidia's data center revenue reached $193.7 billion in fiscal year 2026, up 68% year-over-year. But the diverse demands on the inference side are creating incremental markets for ASICs and edge chips. Institutions estimate that in 2026, ASIC chip shipments will be approximately 7.7 million units, with a 45% share, and will surpass GPU share to reach 58% in 2027. Broadcom is expected to capture about 60% of the AI server computing ASIC market by 2027.

Second, the geographic distribution of infrastructure investment is changing. Hyperscale data centers are still expanding—global cumulative investment in data centers is expected to reach $1.6 trillion by 2030—but edge node construction is growing at a faster rate. AI computing power is no longer concentrated only in the core cloud but is spreading to a three-tier architecture of "core—regional—edge."

Third, the evaluation cycle for investment returns is changing. Centralized data center investment has a long payback period and high capital intensity, requiring years to recoup costs. Edge AI deployments are typically smaller, have shorter cycles, and are closer to specific business scenarios, allowing for more granular return evaluation. This difference is changing the valuation logic of capital markets for AI investment—from "who spends the most" to "who spends most efficiently."

According to Research and Markets, the global AI infrastructure market size will grow from $71.88 billion in 2025 to $90.91 billion in 2026. But this figure only covers the narrow infrastructure hardware market. When enterprise AI deployment, edge computing, and industry solutions are included, the distributed scale of AI capital spending far exceeds this number.

Risks and Constraints: Distribution Is Not a Smooth Path

The trend of AI capital expenditure distribution is clear, but it is not without constraints.

Supply-side bottlenecks remain prominent. Nvidia's Blackwell series products are in a tight supply situation, with demand outstripping supply for several more quarters. Production capacity for key components like HBM has been pre-booked by large customers until 2026 or even 2027. Bernstein Research points out that just the price increase of HBM could increase hyperscalers' overall AI capital spending by about 30%.

Power infrastructure is another constraint. The electricity demand of AI data centers is approaching the carrying capacity of existing power grids. The power connection for a centralized 1GW computing cluster itself is a project that takes years. Distributed architecture reduces the power demand at a single point but imposes new requirements on the distributed access capability of the grid.

Geopolitical risks are also significant. U.S. export restrictions on advanced AI chips continue to affect the global supply chain. Nvidia explicitly excluded the impact of China data center business revenue in its Q1 fiscal 2027 earnings report. Although the mutual mapping of AI capital spending between China and the U.S. is tight, policy uncertainty is increasing friction costs in the supply chain.

Finally, capital market patience for AI investment returns is narrowing. Goldman Sachs has clearly stated that the core contradiction of the AI market is intensifying—fundamentals remain strong, but the market has already priced in too much future earnings. Since November 2022, the market value of AI-related companies has surged by $27 trillion, far exceeding the $9 trillion calculated by macro benchmarks. If distributed investment cannot be converted into revenue and profit more quickly, capital market attitudes may shift from "questioning scale" to "questioning logic."

Conclusion

The distribution of AI capital expenditure is not a denial of centralization but a complement and extension of it.

Training still requires hyperscale data centers; inference is moving to the edge. Giants are still increasing their bets; enterprises and sovereign states are entering the game. GPUs remain the mainstay for training; ASICs and edge chips are opening new battlefronts. This is an era of multi-layered infrastructure—different layers bear different functions, and different participants occupy different niches.

2026 is a critical juncture for this structural transformation. DIGITIMES predicts that global AI market capital expenditure growth will slow from 66% in 2025 to 31% in 2026, but slowdown does not mean stagnation. On the contrary, a slowdown in growth often indicates that the industry is moving from "extensive expansion" to "refined construction." AI infrastructure is evolving from a "winner-takes-all" centralized market to a "layered collaborative" ecosystem.

For investors, understanding the significance of this structural change may be more important than tracking the next quarter's capital spending figures. The distribution of AI capital expenditure is reshaping the long-term investment logic of cloud computing, chip design, enterprise IT architecture, and even national industrial policy. The endpoint of this change is not yet known, but its direction is clear enough.

FAQ

Q1: What is the core driving force behind the distribution of AI capital expenditure?

The explosive growth of inference demand is the core driving force. In 2026, AI inference computing power growth rate reached 122%, far exceeding the 56% for training. The low latency and high concurrency requirements of inference tasks create physical bottlenecks for centralized data centers, making distributed edge nodes an inevitable choice. At the same time, factors such as data compliance and bandwidth costs are also pushing computing power to the edge.

Q2: What are the specific capital expenditures of the four major cloud providers in 2026?

Amazon: approximately $200 billion; Microsoft: approximately $190 billion; Alphabet: approximately $175-185 billion; Meta: approximately $125-145 billion. Total approximately $725 billion, a 77% increase from 2025. About 75% of this goes to AI-related infrastructure.

Q3: What is the relationship between edge AI and cloud computing?

They are complementary, not substitutive. The core cloud handles large model training and complex inference, while edge nodes handle low-latency real-time responses, data preprocessing, and compliant local processing. AI computing power is evolving toward a "core—regional—edge" three-tier distributed architecture, forming a layered collaborative ecosystem.

Q4: What impact does the distribution of AI capital expenditure have on the chip industry?

On the training side, Nvidia GPUs still dominate—data center revenue reached $193.7 billion in fiscal year 2026. However, inference-side demand is creating incremental markets for ASICs and edge chips. In 2026, ASIC shipments are expected to be approximately 7.7 million units, with a 45% share, and in 2027, ASIC share is expected to surpass GPU share. Chip demand is moving from a "single leader" to "multiple coexistence."

Q5: How long can the high growth of AI infrastructure investment last?

Barclays expects major cloud providers' capital spending to reach $919 billion in 2027 and approximately $1.16 trillion in 2028. Nvidia management has raised the upper limit of annual AI industry spending by 2030 to $4 trillion. But growth itself is slowing—from 66% in 2025 to 31% in 2026—indicating the industry is moving from "extensive expansion" to "refined construction."