JPMorgan Analysis: AI CapEx Approaches $870 Billion, Power and Returns Become the Next Challenge

TL;DR
· TrendForce has raised the global CapEx for the top nine CSPs to approximately $830 billion for 2026, with an annual growth rate increasing to 79%.
· Microsoft, Google, Meta, and Amazon continue to ramp up spending, with rising chip prices, data centers, and power facilities driving up budgets.
· Semiconductors, data centers, and the power supply chain directly benefit, but the speed of AI revenue realization will still affect market patience.

The AI infrastructure budgets of major global cloud providers are still being revised upwards.

TrendForce's May press release indicates that the total capital expenditure of the top nine global CSPs for 2026 has been raised to approximately $830 billion, with the annual growth rate adjusted from 61% to 79%. This figure includes Google, AWS, Meta, Microsoft, Oracle, as well as ByteDance, Tencent, Alibaba, and Baidu.

JPMorgan's related strategy materials also point out that the capital expenditure of large hyperscalers for 2026 has entered the range of over $600 billion, but different statistical methods vary in how they treat leasing, power, land, Chinese cloud providers, and emerging AI cloud platforms. A more cautious interpretation is that regardless of whether a narrow or broad definition is used, the AI infrastructure budget for 2026 shows no signs of cooling.

This round of spending is not just about buying more servers. GPU, self-developed ASICs, networking equipment, data center land, power access, and cooling systems are all contributing to the increased budget. For the market, AI infrastructure has expanded from internal investments by tech companies to semiconductor orders, data center leasing, utility investments, and the free cash flow of large tech companies.

$830 Billion Budget Indicates Continued Growth for Cloud Providers

The $830 billion figure provided by TrendForce represents the total for the top nine CSPs globally. It is broader than statistics that only consider large U.S. cloud providers and includes Chinese cloud providers as well.

This is also why there are discrepancies in current AI capital expenditure figures. A narrow definition is closer to the capital expenditures of several U.S. hyperscalers, while a broad definition includes financing leases, data center construction, power-related investments, and more cloud platforms. The numbers cannot be directly added together, but the direction is consistent: AI infrastructure investment continues to be revised upwards.

The public figures from several leading companies are already sufficient to illustrate the strength of this trend.

Microsoft's investor relations materials indicate that the capital expenditure for the calendar year 2026 is approximately $190 billion. Alphabet's first-quarter earnings call and SEC filings show that the expected CapEx for 2026 is between $180 billion and $190 billion. Meta's first-quarter announcement and SEC documents provide a capital expenditure range for 2026 of $125 billion to $145 billion, which includes the principal of financing leases.

Amazon previously stated that the company's total capital expenditure for 2026 is approximately $200 billion. There are also third-party estimates on AWS-related spending that are higher, but this does not equate to Amazon officially providing guidance of over $230 billion for AWS alone.

In other words, cloud giants are not hitting the brakes due to disputes over the AI return cycle. They are still building data centers, purchasing chips, and securing power in advance for training, inference, and enterprise AI demands.

$190 Billion from Microsoft Includes $25 Billion from Component Price Increases

Microsoft's figures best reflect the complexity of this upward revision.

The company mentioned in its FY2026 third-quarter earnings call that the CapEx for the calendar year 2026 is approximately $190 billion, of which about $25 billion is due to the impact of higher component prices. This means that the increase in capital expenditure does not necessarily equate to a simultaneous increase in the number of servers, GPUs, or AI computing power.

This is important for investors. The larger budget is driven partly by demand and partly by costs.

The demand side comes from the continued expansion of training and inference. Cloud providers need to procure more GPUs, build larger clusters, and are also advancing self-developed ASICs to reduce unit computing costs. The cost side is driven by rising prices for high-end chips, storage, networking equipment, power equipment, construction, and key components.

Thus, semiconductors, networking equipment, data centers, and the power supply chain are seeing orders and construction demand; large tech company shareholders are also facing cash flow pressures, depreciation pressures, and profit margin pressures. If AI revenue growth can keep pace, the market will continue to give these expenditures time. If revenue realization lags behind investment, the debate will return to a more direct question: can the money spent convert to revenue quickly enough?

Data Center Vacancy Rate Drops to 1.6%, Power Becomes a Hard Constraint

What supports cloud giants in continuing to ramp up spending is not just management statements, but also supply-side tightness.

CBRE IM data shows that the vacancy rate in major North American data center markets is approximately 1.6%, with a pre-leasing rate of about 74.3% for under-construction capacity. This indicates that many new data centers have already been locked in by cloud providers and AI customers even before they are delivered.

For CSPs, delaying construction by a year may mean missing customer demand. For data center operators, power equipment, and utility companies, order visibility is extended.

High-end chips are also in tight supply. The demand for next-generation AI chips like NVIDIA's Blackwell remains strong, and supply schedules will still affect server delivery pace. Even if cloud providers advance self-developed ASICs, GPUs remain a key resource for training and high-end inference clusters.

An unavoidable constraint is power. AI data centers cannot start operations just by purchasing servers; they also need substations, power access, backup power, cooling systems, and long-term power contracts. The high proportion of natural gas power generation in the U.S. means that the additional load from data centers will also impact natural gas, grid equipment, and utility investments.

This is also why the beneficiary chain of AI capital expenditure continues to expand. GPUs and ASICs are the first to receive orders, followed by data center developers and lessors, while power equipment, natural gas, grid expansion, and cooling systems will also be stimulated. However, the further along the project, the more it relies on local approvals, grid connection progress, labor, and supply chain delivery.

Money is on the way, but revenue realization speed still needs to catch up

The most easily overstated aspect of this round of AI capital expenditure is equating the upward revision of budgets directly with the completion of AI commercialization.

The reality is more fragmented. Cloud giants are indeed ramping up, data centers are indeed tight, and high-end chip orders remain strong. However, part of the growth in capital expenditure comes from component price increases, and power and delivery may slow the actual deployment of computing power. A larger nominal expenditure in 2026 does not mean that usable computing power will increase at the same rate.

Investors are more concerned not about whether cloud providers will continue to spend money, but how quickly this money will turn into revenue. AI training demand supported the first round of investment, while inference, enterprise applications, and cloud service pricing capabilities will determine subsequent returns.

If the cloud revenue, software revenue, and productivity gains brought by AI cannot cover depreciation, energy, and financing costs, the market's patience for massive CapEx will decline. Money is already on the way, chips and data centers are also in line, but what may truly hold back this round of expansion is not the budget, but power access, project delivery, and the speed of revenue growth.